gbp_2008_03_annrpt_2007

FOOTHILLS RESEARCH INSTITUTE GRIZZLY BEAR PROGRAM

2007 ANNUAL REPORT

Prepared and edited by Gordon B. Stenhouse and Karen Graham March 2008

(Formerly the Foothills Model Forest)

Disclaimer This report presents preliminary findings from the 2007 research program within the Foothills Research Institute (FRI) grizzly bear program. It must be stressed that these data are preliminary in nature and all findings must be interpreted with caution. Opinions presented are those of the authors and collaborating scientists and are subject to revision based on the ongoing findings over the course of this study. Suggested citation for information within this report: Dave Hobson, Gordon Stenhouse, Jerome Cranston, Terry Larsen, Karine Pigeon.. 2008. Program Field Activities 2007. In: Stenhouse, G. and K. Graham (eds). Foothills Research Institute Grizzly Bear Program 2007 Annual Report. Hinton, Alberta. 204 pp. This is an interim report not to be cited without the express written consent of the senior author.

ACKNOWLEDGEMENTS A program of this scope and magnitude would not be possible without the dedication, hard work and support of a large number of people. We would like to thank the capture crew members: Bernie Goski, Dave Hobson, Terry Larsen, Jay Honeyman, and Saundi Norris. Thanks also to the local Fish and Wildlife officers and biologists for their help during the capture season: Exemplary flying (and field!) skills were provided by John Saunders of Peregrine Helicopters of Hinton and fixed wing pilot Mike Dupuis of Wildlife Observation Air Services. Thanks also to the veterinarians who assisted with the captures: Marc Cattet, Nigel Caulkett, Erin Geymonat, Johan Lindsjo and Jessica Paterson Appreciation is also extended to the vegetation plot crewmembers Terry Larsen, Karine Pigeon, Benita Kaytor and Krystal Dixon whose hard work and enthusiasm ensured a successful field season. A huge thank you to Cliff Henderson for his assistance with aircraft and logistical support and Alberta Sustainable Resource Development for their help with logistical support and camp accommodations. Thanks also to Marc Symbaluk and Elk Valley Coal for their participation and ongoing support. The Grizzly Bear Research Partner Group and the Board of Directors of the Foothills Research Institute provided valuable support and assistance that allowed the research to proceed in order to address management needs and we thank Don Podlubny and Jim LeLacheur for this. Thanks to Julie Duval and Debbie Mucha for their help in all areas relating to GIS. A word of praise goes out to Joan Simonton for keeping up with media needs and the special communication requirements associated with our program. Thank you to Judy Astalos, Fran Hanington, Denise Lebel, and especially Angie Larocque for an excellent job in managing the administrative details of this program and keeping all the crews in line. Definitely not an easy task! The staff at the Hinton Training Centre provided a great deal of assistance in many ways this year including food and lodging for field crews during their short stays. Thank you to Jerome Cranston for his ongoing GIS support (and much more!) and to John Boulanger who provided superb statistical advice throughout the year. Dr. David Paetkau at Wildlife Genetics International completed the lab work on all DNA hair samples and Matson’s Lab conducted our tooth aging. This program would not have been possible without the many program sponsors (See Appendix 1).

i

TABLE OF CONTENTS

ACKNOWLEDGEMENTS ............................................................................... i TABLE OF CONTENTS ................................................................................. II

LIST OF TABLES ..........................................................................................VII

LIST OF FIGURES......................................................................................... IX

INTRODUCTION ..............................................................................................1

CHAPTER 1: PROGRAM FIELD ACTIVITIES 2007 ................................2

CAPTURE SUMMARY ..................................................................................................... 2 Introduction..................................................................................................................... 2 Study Area ....................................................................................................................... 3 Methodology .................................................................................................................... 4 Results .............................................................................................................................. 6

Capture Locations...................................................................................................................... 6 Sex and Age Characteristics...................................................................................................... 6 Capture Type ............................................................................................................................. 7 Collars ....................................................................................................................................... 8 Status of Captured Grizzly Bears .............................................................................................. 8 Capture Related Mortalities....................................................................................................... 9 Black Bears ............................................................................................................................... 9 Grizzly Bear Vs. Black Bear Captures ...................................................................................... 9

POST CAPTURE FIELD WORK ................................................................................... 10 GPS Location Data and Denning ............................................................................................ 10 Clear Hills – additional data .................................................................................................... 10 Vegetation and Diet Data Collection....................................................................................... 10 Berry Availability.................................................................................................................... 11 Landscape Change and Road Use Intensity ............................................................................ 11

GRIZZLY BEAR HABITAT ENHANCEMENT TRIAL............................................. 12 Introduction................................................................................................................... 12 Methods.......................................................................................................................... 13 Results ............................................................................................................................ 18 Recommendations ......................................................................................................... 20

CHAPTER 2: REMOTE SENSING MAPPING AND RESEARCH UPDATE ............................................................................................................21

INTRODUCTION AND OVERVIEW............................................................................ 21 PHASE 7 MAP PRODUCTS............................................................................................ 22

Introduction................................................................................................................... 22 Methods.......................................................................................................................... 24 Results and Discussion.................................................................................................. 26

Land Cover .............................................................................................................................. 26 Crown Closure......................................................................................................................... 28

ii

Species Composition ............................................................................................................... 29 Vegetation Phenology ............................................................................................................. 30

Discussion....................................................................................................................... 31 Summary and Conclusions........................................................................................... 33

ANNUAL CHANGES IN LANDCOVER, VEGETATION, AND LANDSCAPE STRUCTURE IN BMAS 3 AND 4, 1998-2005 ................................................................ 34

Methods.......................................................................................................................... 35 Results and Discussion.................................................................................................. 37

Landcover 1998-2005.............................................................................................................. 37 Human Disturbance Features 1998-2005 ................................................................................ 39 Roads and Road Density ......................................................................................................... 41 Cutblocks and Proportion of Study Area Covered by Cutblocks............................................ 41 Wellsites and Wellsite Density................................................................................................ 42 Pipelines and Pipeline Density ................................................................................................ 42 Mine Area................................................................................................................................ 43 Changes in Landscape Structure 1998-2005: Local Scale ...................................................... 43 Changes in Landscape Structure 1998-2005: Regional Scale ................................................. 45

Summary........................................................................................................................ 47 LIDAR PROCESSING FOR CANOPY CLOSURE AND VEGETATION STRUCTURE.................................................................................................................... 48

Introduction................................................................................................................... 48 Methods.......................................................................................................................... 48 Results ............................................................................................................................ 52 Discussion....................................................................................................................... 52 Conclusions .................................................................................................................... 53

NOISE REDUCTION OF NDVI TIME SERIES FOR CHARACTERIZING VEGETATION PHENOLOGY....................................................................................... 55

Introduction................................................................................................................... 55 Remote Sensing of Vegetation Phenology .............................................................................. 55

Methods.......................................................................................................................... 58 Results ............................................................................................................................ 60 Discussion....................................................................................................................... 63 Conclusions .................................................................................................................... 65

CLASSIFICATION OF AGRICULTURAL AREAS.................................................... 66 Methods.......................................................................................................................... 66 Results and Discussion.................................................................................................. 68 Conclusions .................................................................................................................... 71

RELATIONSHIPS BETWEEN LANDSCAPE SPATIAL PROPERTIES AND GRIZZLY BEAR PRESENCE IN AGRICULTURAL AREAS................................... 72

Introduction................................................................................................................... 72 Methods.......................................................................................................................... 72 Results and Discussion.................................................................................................. 74 Conclusions .................................................................................................................... 75

REMOTE SENSING OF MOUNTAIN PINE BEETLE SUSCEPTIBILITY: A REVIEW OF REMOTE SENSING OPPORTUNITIES............................................... 76

Introduction................................................................................................................... 76 Background on Mountain Pine Beetle ........................................................................ 77

iii

Predicting Susceptibility............................................................................................... 78 Remote Sensing of Susceptibility Variables ............................................................... 80

Basal Area ............................................................................................................................... 80 Age .......................................................................................................................................... 81 Stand Density .......................................................................................................................... 82

Future Directions .......................................................................................................... 83

CHAPTER 3: GRIZZLY BEAR/MOUNTAIN PINE BEETLE INTERACTIONS: REMOTE SENSING MONITORING AND MODELING YEAR END REPORT ..............................................................88

EXECUTIVE SUMMARY............................................................................................... 88 Analysis Preparation and Support .............................................................................. 88

DATA BLENDING: BLENDING OF LANDSAT AND MODIS DATA FOR THE GENERATION OF HIGH SPATIAL AND TEMPORAL RESOLUTION IMAGERY FOR CHANGE DETECTION IN VEGETATION ................................... 96

Introduction................................................................................................................... 96 Study area and data description .................................................................................. 96 Methods.......................................................................................................................... 97

Improving MODIS spatial resolution...................................................................................... 97 Generating Landsat ETM data................................................................................................. 97 Data blending .......................................................................................................................... 97 Validation of STARFM predictions ........................................................................................ 98

Results and Discussion.................................................................................................. 99 CHANGE AND BEETLE MAPPING: MAPPING FOREST DISTURBANCE USING A LANDSAT-5 TM TIME-SERIES: INITIAL RESULTS............................ 104

Background ................................................................................................................. 104 Data description .......................................................................................................... 104

Remotely sensed data ............................................................................................................ 105 Mountain pine beetle attack and mitigation data................................................................... 105

Methods........................................................................................................................ 105 Generating an enhanced wetness difference index................................................................ 105 Processing mountain pine beetle mitigation data .................................................................. 105 Additional disturbance data................................................................................................... 106 Thresholding EWDI values ................................................................................................... 106

Results .......................................................................................................................... 106 Discussion..................................................................................................................... 109

CHAPTER 4: POTENTIAL AND REALIZED GRIZZLY BEAR HABITAT BASED ON FOOD RESOURCES ........................................... 111

Introduction................................................................................................................. 111 Methods........................................................................................................................ 111

Distribution of individual food resources.............................................................................. 111 Diet-based weighting of food presence-absence maps.......................................................... 112 Evaluating selection for food resource models ..................................................................... 112 Evaluating selection for food-habitat models........................................................................ 112 Realized food-habitat values and a regional habitat loss index............................................. 113 Correlations between food-habitat models and RSFs............................................................ 113 Relationship between food-habitat models and bear occupancy from DNA studies ............ 113

iv

Results .......................................................................................................................... 114 Food resource distribution models ........................................................................................ 114 Food-habitat models .............................................................................................................. 114 Evaluating selection for food resource models ..................................................................... 114 Evaluating selection for food-habitat models........................................................................ 114 Realized food-habitat values and a regional habitat loss index............................................. 115 Correlations between food-habitat models and RSFs............................................................ 115 Relationship between food-habitat models and bear occupancy from DNA studies ............ 115

Discussion..................................................................................................................... 115

CHAPTER 5: GRAPH THEORY AND RSF GENERATED MOVEMENT CORRIDORS FOR GRIZZLY BEARS IN ALBERTA, CANADA (PHASE 3 AND PHASE 4 RESULTS COMBINED) .............. 131


Study Area and Graph Creation ............................................................................................ 131 Detailed Graph Analyses....................................................................................................... 132

Results .......................................................................................................................... 132 Discussion..................................................................................................................... 133

CHAPTER 6: WILDLIFE HEALTH.......................................................... 143

Introduction................................................................................................................. 143 PROGRESS IN STRESS RESEARCH ......................................................................... 146

Development and Validation of Serum-Based Indicators of Long-Term Stress ................... 146 Development and Validation of Tissue-Based Indicators of Long-Term Stress................... 147

PROGRESS IN WILDLIFE HEALTH RESEARCH.................................................. 149 Development of Grizzly Bear Health Profiles....................................................................... 149 Associations between Age, Stress, and Other Health Function Scores................................. 153

CHAPTER 7: GEOGRAPHIC INFORMATION SYSTEM (GIS) PROGRESS REPORT .................................................................................. 157

Introduction................................................................................................................. 157 Methods and Results................................................................................................... 158

Analysis 1: BMA 3 and 4 ...................................................................................................... 158 Analysis 2: Phase 6 Mapping Extent..................................................................................... 162 Analysis 3: All captures ........................................................................................................ 164 Analysis 4: All captures, annual habitat conditions .............................................................. 165

CHAPTER 8: PROGRESS REPORT ON ANALYSIS OF GIS, LANDSCAPE MATRICES AND HEALTH VARIABLES...................... 167


Variables................................................................................................................................ 167 Analysis strategies................................................................................................................. 169

Results .......................................................................................................................... 170 Sample sizes .......................................................................................................................... 170

v

Exploration of the relationship between bear demography and GIS/Landscape conditions .............................................................................................................................. 171 Relationship between environmental variables and health scores......................................... 176

Discussion..................................................................................................................... 181 Future analyses for final program report ............................................................................... 183

CHAPTER 9: DELIVERY OF NEW PRODUCTS AND DEVELOPMENT OF TRAINING PROGRAMS ..................................... 185

APPENDIX 1: LIST OF PROGRAM PARTNERS................................... 197

APPENDIX 2: LIST OF PUBLISHED PAPERS....................................... 198

vi

LIST OF TABLES

CHAPTER 1: PROGRAM FIELD ACTIVITIES 2007 Table 1: Grizzly bears captured in each capture area 2007. ...........................................................................6 Table 2: Sex and age of captured grizzly bears.................................................................................................7 Table 3: Grizzly bear capture types...................................................................................................................7 Table 4: Status of 2006 research grizzly bears as of January 2008. ................................................................8 Table 5: Black bear captures by sex and age classifications. ...........................................................................9 Table 6: Number of grizzly bears versus black bears captured by year.........................................................9 CHAPTER 2: REMOTE SENSING MAPPING AND RESEARCH UPDATE Table 1: Land cover class descriptions. ...........................................................................................................25 Table 2: Phenological product descriptions. ...................................................................................................26 Table 3: Landsat scenes acquired and processed for quantifying human disturbance, land cover, and

landscape structure from 1998 to 2005. ..................................................................................................36 Table 4: A summary of road density, cutblock proportion, wellsite density, pipeline density, and mine

area across BMAs 3 and 4 from 1998 to 2005.........................................................................................39 Table 5: Road length, change in road length, road density, and change in road density from

1998 to 2005. ..............................................................................................................................................41 Table 6: Cutblock area, change in cutblock area, proportion of study area covered by cutblock, and

change in proportion of study area that is cutblock from 1998 to 2005. ..............................................41 Table 7: Number of wellsites, change in number of wellsites, wellsite density, and change in wellsite

density from 1998 to 2005.........................................................................................................................42 Table 8: Pipeline length, change in pipeline length, pipeline density, and change in pipeline density

from 1998 to 2005......................................................................................................................................42 Table 9: Mine area and change in mine area from 1998 to 2005...................................................................43 Table 10: Temporal occurrence and mean annual amounts of forest change patches across the three

selected 13 x 13 km case study areas, reflecting low, moderate and high levels of change. ................43 Table 11: Edge density, mean patch size, coefficient of variation of mean patch size, mean nearest

neighbour, and coefficient of variation of mean nearest neighbour from 1998-2005..........................45 Table 12: NDVI time series metrics, derived for the characterization of surface vegetation phenology...57 Table 13: Summary of mountain pine beetle susceptibility models. .............................................................79 CHAPTER 3: GRIZZLY BEAR/MOUNTAIN PINE BEETLE INTERACTIONS: REMOTE SENSING MONITORING AND MODELING YEAR END REPORT Table 1: Landsat TM and MODIS image acquisition dates for path 46, row 22, year 2006 for the

Kakwa region. ...........................................................................................................................................89 Table 2: Landsat TM and MODIS image acquisition dates for path 47, row 24 year 2001 for the

Williams Lake study area.........................................................................................................................92 Table 3: Select moderate spatial resolution image sources. ...........................................................................94 Table 4: Association of Landsat and MODIS bands for STARFM prediction. ...........................................98 Table 5: Coefficients used to generate wetness indices from Landsat 5 TM data (from Liang, 2004).....105 CHAPTER 4: POTENTIAL AND REALIZED GRIZZLY BEAR HABITAT BASED ON FOOD RESOURCES Table 1: Critical food resources used to define habitat in west-central Alberta, Canada. Food resource,

abbreviation (code), feeding activity, and general season of use are described.................................118

vii

Table 2: Description and characteristics of environmental variables used to model the probability of occurrence of individual grizzly bear food resources in west-central Alberta, Canada. ..................119

Table 3: Weights used to score food-habitat values (weight × food item presence- 0 or 1) for bi-monthly periods based on reported diet volumes from Munro et al (2006). Final food-habitat values were standardized between 0 and 100 (maximum possible sum of weights in a bi-monthly period) by multiplying the bi-monthly sum by a scaling factor determined by dividing the sum into 100. ......120

Table 4: Coefficients describing the probability of female grizzly bear occupancy (presence of a home range) for six population units in Alberta. All factors represent proportion within a 10-km radius window, except agriculture in the Livingstone and Waterton (phase 4) units where agriculture was measured as proportion within 1-km.. ..................................................................................................121

Table 5: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of 4 letter codes) in non-harvested forest stands near Hinton, Alberta.....................122

Table 6: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of 4 letter codes) in harvested forest stands near Hinton, Alberta.............................123

Table 7: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of the 4 letter codes) for open habitats near Hinton, Alberta.....................................124

Table 8: Evaluation of individual food probability models based on a comparison between animal locations visited in the field and food use identified (number of positive sample sites out of 1,032 indicated as n) against either random locations in the study area or sites without any bear sign and assumed to be locations where bears were travelling or foraging on other minor foods such graminoids. ..............................................................................................................................................125

Table 9: Estimated odds ratio (Odds) of grizzly bear occupancy at 705 DNA mark-recapture bait sites in the FRI core population unit (Boulanger et al 2005) as described by a 1 standard deviation change in potential or realized food-habitat values within a 900-m radius of bait sites. ...............................125

CHAPTER 6: WILDLIFE HEALTH Table 1: Stress-associated proteins detected by the “bear stress chip”, a long-term stress detection tool

developed for use in the conservation of grizzly bears.........................................................................148 Table 2: Constituent variables used to calculate health function scores for grizzly bears captured for the

Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007). ......................149 Table 3: Descriptive statisticsA for health function scores and ages of grizzly bearsB captured for the

Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007). ......................150 CHAPTER 8: PROGRESS REPORT ON ANALYSIS OF GIS, LANDSCAPE MATRICES AND HEALTH VARIABLES Table 1: Landscape variables considered in analyses. Variables marked with an * were updated for each

year so yearly change in these variables reflected both landscape and bear home range shift for a portion of the study area. .......................................................................................................................168

Table 2: Landscape level metrics considered in the analysis .......................................................................168 Table 3: Health variables considered in the analysis. Italics variables were suggested after the analysis

was conducted and will be considered in future analyses....................................................................169 Table 4: Sample sizes of health variables ......................................................................................................171 CHAPTER 9: DELIVERY OF NEW PRODUCTS AND DEVELOPMENT OF TRAINING PROGRAMS Table 1: Training course outline. ...................................................................................................................186

viii

LIST OF FIGURES

CHAPTER 1: PROGRAM FIELD ACTIVITIES 2007 Figure 1: Clear Hills Capture area with site locations. ....................................................................................3 Figure 2: Kakwa and Nose Hill Tower capture area with site locations.........................................................4 Figure 3: Elk Valley Coal capture area with site locations..............................................................................5 Figure 4: The target condition: an opening occupied by Shepherdia canadensis and Hedysarum

alpinum.......................................................................................................................................................12 Figure 5: Surveyed sites in the Moose Mountain area ...................................................................................13 Figure 6: Site 5...................................................................................................................................................14 Figure 7: Site 3...................................................................................................................................................14 Figure 8: Site 1...................................................................................................................................................15 Figure 9: Site 2...................................................................................................................................................15 Figure 10: Site 7.................................................................................................................................................16 Figure 11: Site 4.................................................................................................................................................16 Figure 12: Site 6.................................................................................................................................................17 Figure 13: Site 8.................................................................................................................................................17 Figure 14: Hedysarum seedlings ......................................................................................................................18 Figure 15: Buffaloberry seedlings ....................................................................................................................19 Figure 16: Tilled vs. non-tilled ground ............................................................................................................19 Figure 17: Planting on Site 1 ............................................................................................................................20 CHAPTER 2: REMOTE SENSING MAPPING AND RESEARCH UPDATE Figure 1: The seven phases of the FRIGBP, and the years in which they were adopted. ...........................23 Figure 2: Flow diagram highlighting the general methods used to create the four base map products:

land cover, crown closure, species composition, and phenology...........................................................25 Figure 3: Phase 7 land cover map. ...................................................................................................................27 Figure 4: Phase 7crown closure map. ..............................................................................................................28 Figure 5: Phase 7 species composition map.....................................................................................................29 Figure 6: 16-day NDVI phenology composites for the 2005 growing, Phase 7.............................................30 Figure 7: Examples of the flexible composite map capability of the current land/vegetation

information base. Maps shown here cover a small portion of the Phase 3 study area and were generated through simple raster calculation and overlay procedures in a GIS environment............32

Figure 8: Backdating and updating a reference map through change analysis. ..........................................35 Figure 9: Location of the three 13 x 13 km case study areas in the North Health study area,

representing three levels of relative land cover change between 1998 and 2005. ................................37 Figure 10: Samples of changing land cover over a small portion of BMAs 3 and 4, 1998 to 2005. ............38 Figure 11: Trends in road density, cutblock proportion, wellsite density, pipeline density, and mine

area across BMAs 3 and 4 from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined......................................................................................40

Figure 12: Cumulative change in A) forest area (i.e. forest loss), B) percent edge density, and C) mean patch size, between the years 1998 and 2005 in three case study areas of low (case study 1), moderate (case study 2), and high ( case study 3) annual mean forest change. ...................................44

Figure 13: Trends in edge density, mean patch size, and mean nearest neighbour in BMAs 3 and 4 from 1998 to 2005. ..............................................................................................................................................46

Figure 14: Location of study area and LiDAR transect in west-central Alberta, Canada. The LiDAR transect was designed to sample the diversity of natural subregions in the area, while maximizing forest coverage......................................................................................................................49

ix

Figure 15: Illustration of how the line segment and point count methods for estimating canopy closure are performed on a forest profile generated from LiDAR. The horizontal projection of the vertical structure is estimated by calculating the proportion of vegetation segments (marked c) to ground segments (marked d). ...........................................................................................51

Figure 16: Plots of r2 values derived from regression analyses performed between ground- and LiDAR-estimated values of canopy closure using the line segmentation method with absolute (A) and relative (B) thresholds, the point cound method with absolute (C) and relative (D) thresholds, and the histogram method with absolute (E) thresholds. ..............................54

Figure 17: Map showing the location, context and extent of the study area within west-central Alberta, Canada. .......................................................................................................................................58

Figure 18: Flowchart of the principal steps involved in the empirical comparison of the selected candidate noise reduction techniques......................................................................................................59

Figure 19: Standardized summary performance scores for the seven candidate noise reduction strategies, ordered from lowest (left) to highest (right). ‘None’ refers to the absence of noise reduction. ...................................................................................................................................................61

Figure 20: Pie chart showing the number and percentage of scenarios in which either a noise-reduced time series (light grey) or the noisy, unfiltered times series (dark grey) produced a better unstandardized performance score. ........................................................................................................62

Figure 21: Map of Alberta showing collared grizzly GPS locations and the North and South study areas. ................................................................................................................................................67

Figure 22: Overall accuracy and KIA for all three classification methods. .................................................69 Figure 23: Completed mosaic of the Phase 6 study area with SSM classification, showing new

agricultural classes (top 5 in legend) with those of the FRI land cover map. ......................................70 Figure 24: Study area map showing the distribution of the 107 sub-landscapes in southern Alberta.......73 CHAPTER 3: GRIZZLY BEAR/MOUNTAIN PINE BEETLE INTERACTIONS: REMOTE SENSING MONITORING AND MODELING YEAR END REPORT Figure 1: Grizzly bear focus study area in the Kakwa region. ......................................................................89 Figure 2: Grizzly bear test site in the Williams Lake area.............................................................................91 Figure 3: Select image footprints over Kakwa study area. ............................................................................95 Figure 4: Simplified schema of the STARFM prediction algorithm.............................................................98 Figure 5: A-J: Comparison between observed Landsat scene and difference image between observed

and predicted Landsat scenes. ...............................................................................................................102 Figure 6: TOA reflectance for Landsat scene observed on July 11 (A) and the same area predicted

for July 21 (B) (both Landsat band 1). ..................................................................................................102 Figure 7: T1 (left) and T2 (right) Landsat-5 images. Note the location of inset maps for Figure 10. ......104 Figure 8: Distribution of the number of stems removed at the MPB mitigation sites. Only five sites

had more then ten stems eliminated. .....................................................................................................106 Figure 9: Median, upper and lower quartiles, and non-outlier range of EWDI values for each class.

The mitigation data, categorized as “possibly disturbed,” shows virtually no separation from the healthy forest sites, whilst areas of more severe disturbance are more clearly defined. ............107

Figure 10: T1 and T2 Landsat-5 TM images, a visualization of the EWDI, and an example of a threshold-based classification with actual MPB mitigation data........................................................108

Figure 11: Mitigation amounts and associated EWDI values (top), and idealized EWDI distributions for the classes (bottom). Note the inclusion of the distribution expected for increased levels of red attack. ................................................................................................................................................109

CHAPTER 4: POTENTIAL AND REALIZED GRIZZLY BEAR HABITAT BASED ON FOOD RESOURCES Figure 1: Study area map depicting elevation, populated places, study area where field plots were

visited, and the area modelled and tested. ............................................................................................126

x

Figure 2: Probability of female grizzly bear occupancy as a function of natural sub-region identity and agriculture. .......................................................................................................................................127

Figure 3: Predicted multi-seasonal (1 May to 31 September) distribution of food resources (index from 0 to 1,000) in west-central Alberta, Canada. ...............................................................................128

Figure 4: Mean (±SE) habitat selection (β) of radio-collared female grizzly bears by natural region (mountain in black and foothills in gray) for food-based scores of habitat quality. .........................129

Figure 5: Predicted patterns of realized food-habitat values (a.) and a regional habitat loss index (b.) (0- no loss to 1000- complete loss) based on the difference between realized and potential habitats.....................................................................................................................................................130

CHAPTER 5: GRAPH THEORY AND RSF GENERATED MOVEMENT CORRIDORS FOR GRIZZLY BEARS IN ALBERTA, CANADA (PHASE 3 AND PHASE 4 RESULTS COMBINED) Figure 1: Phase 3 and Phase 4 study region demonstrating RSF-based habitat patches used in

graph theory analyses. ............................................................................................................................134 Figure 2: Phase 3 and Phase 4 study region showing RSF patch centroids or nodes which provided

the basis for least-cost path creation. ....................................................................................................135 Figure 3: Phase 3 and Phase 4 least-cost movement paths further defined by the average daily

movement distance threshold (4942 m).................................................................................................136 Figure 4: Phase 3 and Phase 4 least-cost movement paths further defined by the 95th percentile

distance threshold (6247 m). ..................................................................................................................137 Figure 5: Phase 3 and Phase 4 identified RSF-based habitat patches important to maintaining

overall connectivity based on minimum spanning tree node removal iterations...............................138 Figure 6: Phase 3 and Phase 4 identified RSF-based habitat patches important to maintaining

overall connectivity according to area-weighted node removal iterations. ........................................139 Figure 7: Phase 3 and Phase 4 fine-scale graph theory generated corridors important to

maintaining local connectivity (raster format). ....................................................................................140 Figure 8: Phase 3 and Phase 4 large-scale graph theory generated corridors important to

maintaining regional connectivity (raster format). ..............................................................................141 CHAPTER 6: WILDLIFE HEALTH Figure 1: Although associations are sometimes drawn between occurrences of large-scale,

human caused environmental change and poor performance of resident wildlife populations, evidence of cause and effect is usually lacking. An added complication is considerable time may lapse between occurrences of environmental change and detection of poor population performance. The critical issue here is our lack of understanding of causal mechanisms underlying conservation problems, such as population decline, prevents us from predicting with any accuracy the effects of environmental change on population performance before its occurrence. ……………………………………………………... 143

Figure 2: Tracking and long-term stress and understanding wildlife health is crucial to establishing if large-scale, human-caused environmental change is the cause of, or contributes to, poor performance of resident wildlife populations. ……………………………………………...………...145

Figure 3: Change in health function scores of four adult grizzly bears captured multiple times in the core study area of the Foothills Research Institute Grizzly Bear Program from 1999 to 2003........151

Figure 4: Change in mean population health function scores for grizzly bears captured in the core study area of the Foothills Research Institute Grizzly Bear Program from 1999 to 2003................152

Figure 5: Mean health function scores for grizzly bears that inhabit different regions of Alberta and were captured for the Foothills Research Institute Grizzly Bear Program (1999-2007).. ................154

Figure 6: Associations between age, stress score, and other health function scores in grizzly bears captured for the Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007). .............................................................................................................................................155

xi

xii

CHAPTER 7: GEOGRAPHIC INFORMATION SYSTEM (GIS) PROGRESS REPORT Figure 1: Area and time frame for the annual landscape conditions maps created for the health

analysis.....................................................................................................................................................158 Figure 2: Bear Management Areas (BMA) 3 and 4 and the bear location data available for this area...159 Figure 3: The 95% kernel home range for G020 in 2002 using a single smoothing factor. ......................160 Figure 4: Kernel home ranges created for the second analysis. ..................................................................162 Figure 5: An example of the change in anthropogenic features for a small area of BMA 3 from

2000 to 2004. ............................................................................................................................................164 CHAPTER 8: PROGRESS REPORT ON ANALYSIS OF GIS, LANDSCAPE MATRICES AND HEALTH VARIABLES Figure 1: Conceptual diagram of health-environmental analysis. ..............................................................170 Figure 2: Results of constrained ordination of bear age-sex class with GIS variables as predictor

variables. The first 2 canonical axes are shown. All GIS variables that were underlined were significant predictors of bear age-sex class ...........................................................................................171

Figure 3: Standardized mean scores for significant GIS predictor variables from ordination analysis (Figure 2) ...................................................................................................................................173

Figure 4: The distribution of canopy closure values across the study area. It can be seen that high canopy closure areas are found both in foothills area and mountain valleys whereas low canopy closure is primarily in alpine areas. ..................................................................................174

Figure 5: Constrained ordination biplot showing the relationships between age-sex class and predictor GIS and landscape variables. Underlined GIS/Landscape variables were significant predictors of age-sex class....................................................................................................175

Figure 6: Constrained ordination biplot showing the relationships between landscape and GIS variables. Underlined landscape variables were significant predictors of age-sex class..................175

Figure 7: Constrained ordination biplot showing the relationships GIS/landscape variables and condition indices. Underlined GIS/landscape variables were significant predictors of condition indices......................................................................................................................................176

Figure 8: Plot of straight line length (SLL) versus elevation as a function of age and sex class. A general decreasing trend in SLL with increasing elevation can be seen for each age-sex class. ......177

Figure 9: Plots of straight line length versus LAF (linear access features/roads) and elevation as a function of sex. Blue points represent subadult bears, red points represent adult bears and a star represents females with cubs. ......................................................................................................178

Figure 10: Body condition index versus mean RSF score as a function of age and sex class....................178 Figure 11: Constrained ordination triplot with canonical scores for bears shown as a function of

age and sex class. Significant GIS/landscape variables are underlined.............................................179 Figure 12: Plots of observed GGT and hsp60 levels as a function of contrast edge density (L2) and

canopy closure. Females are depicted by circles and males by squares. Red symbols are adults and blue symbols are subadults. Light blue stars are females with cubs...............................180

Figure 13: Constrained ordination biplots for adult female reproductive hormone analysis. Significant environmental variables are underline. A plot of observed testosterone versus variation in mean patch size is also shown............................................................................................180

Figure 14: Constrained ordination biplots for immunity analysis. Significant environmental variables are underlined.........................................................................................................................181

Introduction

INTRODUCTION

Gordon Stenhouse1

1 Foothills Research Institute Grizzly Bear Program (FRIGBP)[email protected]

The Alberta landscape continues to face pressures to meet societal needs. These needs include resource extraction activities (forestry, mining and oil and gas development) along with the multitude of recreational activities that Albertans pursue on a year round basis. When viewed collectively these activities result in alteration of landscapes along the eastern slopes of Alberta. These areas are also home to grizzly bear populations which are a species widely recognized as an indicator of ecosystem health. Our research team has focused efforts on developing new innovative products and techniques to investigate and understand the relationships between landscape conditions, landscape change (human-caused), and health in grizzly bears. The results of this research effort will enable the Alberta government departments to make well informed and timely land use management decisions to maintain ecosystem health and to support sustainable development. Although this project focuses on Alberta grizzly bear populations, the concepts, techniques and relationships uncovered can be applied to a variety of species at risk in Alberta and Canada. These leading edge innovative products and techniques will make Alberta a recognized world leader in ecosystem management and monitoring. This annual program report for the Foothills Research Institute’s Grizzly Bear Program (formerly known as the Foothills Model Forest Grizzly Bear Research Program) also represents the first year for a number of new initiatives including: work on mountain pine beetle and grizzly bear response, new seasonal grizzly bear food models, and research to investigate the relationship between grizzly bear denning behavior and weather factors. Perhaps the most important accomplishment of our research program to date is contained within the deliverables for 2007 and detailed in the chapter on remote sensing mapping efforts which is the completion of a seamless habitat map (and associated data layers) for all currently identified grizzly bear range in Alberta. To our knowledge Alberta is the only jurisdiction in the world to have such a map product available for the management of this species. This is a major accomplishment and one that has taken our research team nine years to complete. This would not have been possible without the tremendous support provided by our program partners (See Appendix 1). As in previous years this report is divided into separate sections which provide detail on the various program elements within the research effort. These sections have been prepared by the principle investigators of these elements who have or will be publishing these results in scientific peer reviewed journals. A listing of research publications is presented in Appendix 2.

1

Chapter 1: Program Field Activities

CHAPTER 1: PROGRAM FIELD ACTIVITIES 2007

Dave Hobson2, Gordon Stenhouse1, Jerome Cranston3, Terry Larsen1, Karine Pigeon1

1Foothills Research Institute; 2Alberta Sustainable Resource Development, Fish and Wildlife Division;

3Arctos Ecological Services, Hinton, AB

CAPTURE SUMMARY Introduction In order to gather the necessary samples from study animals we conduct an annual spring capture and collaring effort. This activity provides the needed animal health and movement data to relate to landscape conditions. The 2007 grizzly bear spring capture session was the 9th conducted by the Foothills Research Institute Grizzly Bear Program. In 1999, our original study area encompassed 10,000 km2 in an area south of Highway 16, between Edson and Jasper in the north and the Brazeau River in the south. In 2003 the study area expanded to include all of the grizzly bear range between the Berland River and the Montana border (62% of grizzly bear range in Alberta). In 2005, the study area expanded again to include areas between the Berland and Wapiti Rivers plus the Swan Hills. In 2006, the study area expanded northwards to include the Chinchaga River, the Hotchkiss River and the Meikle River area. We also included an area south of the Wapiti River and returned to the Swan Hills. In 2007 our capture and collaring efforts were focused in 3 areas. We returned to the Kakwa and Nose Hill Tower areas of the Weyerhaeuser Forest Management Agreement area and the Clear Hills area north of Worsley which included the southern part of the Chinchaga area. We also did capture work on the coal leases operated by Elk Valley Coal which included the new Cheviot mine southwest of Hinton. There were three primary purposes for capture and collaring grizzly bears in these areas. In the Clear Hills we wanted additional grizzly bear movement data for RSF (Resource selection function) map development for the north west portion of the province. In the Kakwa area our focus was to collect grizzly bear movement data that will relate to ongoing mountain pine beetle activities and ongoing forest management efforts in this area. Our research team has recently completed an RSF map product for the Kakwa area with grizzly bear location data obtained prior to mountain pine beetle activity in this area. Data obtained on Elk Valley’s coal leases were to examine the impact of coal mining activities on grizzly bears.

2


Tooth, blood and tissue samples collected from captured grizzly bears will be used to help understand population dynamics, assist in defining population units through genetic analysis, and track population health. The goal of this years’ capture session was to deploy 15-20 GPS radio-collars on grizzly bears in these study areas. We also embarked on the first major deployment of new camera/sensor systems that were attached to our existing GPS collars. These new camera/sensor systems were being field tested to gather additional information on habitat use and detailed movement paths. Study Area Three capture areas were designated within the 2007 study area (Figure 1). These capture areas were as follows:

1. The Clear Hills area. (Figure 1) 2. The Kakwa and Nose Hill Tower area. (Figure 2) 3. Elk Valley Coal Leases (Figure 3)

Figure 1: Clear Hills Capture area with site locations.

3


Figure 2: Kakwa and Nose Hill Tower capture area with site locations. The 2007 study area comprised parts of provincial Bear Management Areas (BMAs) 1 (Clear Hills), and 2B (Kakwa and Nose Hill Tower). These BMAs make up approximately 66% of grizzly bear range in Alberta. The Elk Valley Coal area occurred in a portion of BMA 3. Methodology Field capture efforts began in May this season. In May, capture operations were conducted by 1 helicopter-based crew (Clear Hills) and 1 ground-based crew (Kakwa Tower). Each crew consisted of biologists with experience in grizzly bear capture and a project Veterinarian. The Clear Hills team utilized an aerial snare line approach with daily checks on all active sites. Due to the forest cover and low elevation aerial darting was not possible. During this same time period the Kakwa based team worked exclusively along existing forest access roads to establish snaring sites.

4


Figure 3: Elk Valley Coal capture area with site locations. On June 4th, after 4 weeks of continuous capture efforts, the helicopter crew moved to the Kakwa Tower area to allow capture efforts to expand beyond roaded areas. Primary capture efforts in the Kakwa were concluded at the end of June. Capture efforts for the Elk Valley Coal area occurred in mid July and then again in early September 2007. Only culvert traps were used. Bears were immobilized using a drug combination of Telazol and xylazine (XZT). The drugs were administered by rifle/pistol once the bear had been restrained in a snare, or culvert trap. Atipamazole was used to reverse the xylazine, after handling procedures were completed. Once immobilized, grizzly bears were weighed, and measured (chest girth, zoological length, and straight-line length). Samples were collected (blood, hair, ear plug tissue and tooth). Radio-collar and ear tag transmitters were attached. Vital functions and blood-oxygen levels were monitored during the processing.

5


Black bears and other non-target species were marked with ear tags and released from the snare after immobilization. Vital conditions were monitored while these species were under anesthesia but measurements and samples were not collected. After administering the reversal, all bears were monitored until they became mobile. As required by the Alberta SRD research permit conditions, all captured bears were checked within 24 hours of capture to ensure that they had recovered from immobilization and handling procedures. Results Capture Locations In total, 16 grizzly bears and 15 black bears were captured this trapping season. Of the grizzly bear captures, only 1 was caught in the Clear Hills capture area, while 13 were caught in the Kakwa capture area and 2 were caught in the Elk Valley Coal area (Table 1). One bear, G267 was a bear that was relocated to the Sheep Creek area in August and was collared to collect additional movement information. Table 1: Grizzly bears captured in each capture area 2007.

Capture Area Captured Grizzly Bear IDs Clear Hills G242

Kakwa Tower G222, G223, G224, G226, G238, G260, G261, G262, G263, G264, G265, G266, G267

Elk Valley Coal G008, G110 Sex and Age Characteristics Of the 16 grizzly bears captured, 11 (69%) were adults, 5 (31%) were sub-adults. Eleven (69%) were males and 5 (31%) were females (Table 2). Adult males made up the largest component of captured grizzly bears (56%) while sub-adult males, adult females, and sub-adult females comprised 12%, 12%, and 19% of captured grizzly bears respectively. No cubs of the year or yearlings were caught.

6


Table 2: Sex and age of captured grizzly bears.

Grizzly Bear IDs Area Age Sex G008 Elk Valley Coal Adult Male G110 Elk Valley Coal Adult Male G222 Kakwa Adult Male G223 Kakwa Adult Female G224 Kakwa Sub-adult Female G226 Kakwa Adult Male G238 Kakwa Sub-adult Female G260 Kakwa Adult Female G261 Kakwa Adult Male G262 Kakwa Adult Male G263 Kakwa Adult Male G264 Kakwa Adult Male G265 Kakwa Sub-adult Female G266 Kakwa Sub-adult Male G267 Kakwa Sub-adult Male G242 Clear Hills Adult Male

Of the 2 captured adult females, only G223 was observed with dependant young (3 cubs). Capture Type Capture types were categorized as ground capture with snare, ground capture with culvert trap or helidarted. Most ground capture sites used snares in 3 different formats, pail sets, cubby sets and/or trail sets. Three new lightweight aluminum culvert traps were deployed. Of the 18 capture events (G223 and G260 were captured twice)(Table 3), helicopter captures accounted for 12% (2) of capture events and ground captures accounted for 88% (16) of capture events. Thirteen ground-capture events involved snares and 3 involved a culvert trap. Table 3: Grizzly bear capture types.

Grizzly Bear IDs Capture Type Grizzly Bear ID Capture Type G008 Helidart G110 Culvert G222 Snare G261 Snare G223 Snare G262 Snare

G223 2nd capture Helidart G263 Snare G224 Snare G264 Snare G226 Snare G265 Snare G238 Culvert G266 Snare G260 Snare G267 Snare

G260 2nd capture Culvert G242 Snare

7


Collars Seventeen radio-collars and ear tag transmitters were deployed. Radio-collars deployed consisted of 3 types; 13 Tellus (including new UHF model), 3 ATS GPS and 1 Telonics Argos GPS. ATS and Telonics collars were programmed to obtain a location every 2 hours during the active period (April 1 to November 31) and to shut off during the denning period (December 1 to March 31) to preserve battery power. Tellus collars were programmed to obtain a location every hour during the active period and every 2 hours during the denning period. ATS collars are expected to last one season, while Telonics and Tellus collars should last until the fall 2009.

All radio-collars were outfitted with a remote release mechanism and all collars were equipped with a rot-off system as a backup in case of electronic failure. G008’s ATS collar was successfully remotely released from the bear and retrieved in the fall of 2007 and the remaining 2 ATS collars will be remotely released and retrieved in the spring of 2008.

G223s Tellus collar was replaced with another Tellus collar when she was caught the second time. Status of Captured Grizzly Bears Table 4 lists status of collared grizzly bears as of January 2008. Table 4: Status of 2006 research grizzly bears as of January 2008.

Grizzly Bear IDs Fate as of January 2008 G008 Collar retrieved fall 2007 G110 Alive, ATS collar working G222 Last heard 03 July 2007, Tellus collar pulled off after

24 hours. G223 Alive, Tellus collar working. G224 Alive, Tellus collar working. G226 Alive, Tellus collar malfunctioned; Dropped and

retrieved; Collar banding was cut G238 Alive, Tellus collar working. G260 Alive, Tellus collar working. G261 Last heard 14 August 2007. Argos collar pulled off

after 4 days. G262 Tellus collar last heard 14 August 2007 G263 Last heard 08 June 2007. Tellus collar pulled off after

3 days. G264 Tellus collar last heard Sept 2007 G265 Alive, Tellus collar working G266 Alive, Tellus collar working G267 Alive, ATS collar dropped while in den; to be

retrieved spring 2008. G242 Tellus collar last heard 16 June 2007.

8


Capture Related Mortalities There were no capture related mortalities of either grizzly or black bears this year. Black Bears A total of 15 black bears were captured (7 in the Kakwa area, 7 in the Clear Hills area and 1 in the Elk Valley Coal area). Adult males constituted the largest number of captured black bears (64%). Adult females, and sub-adult males comprised 29% and 7% of captured black bears respectively. No cubs of the year, yearlings or sub-adult females were captured. Male/female percentages of captured black bears were 71/29 and adult/sub-adult percentages were 93/7. One black bear was caught in a culvert trap and released without handling therefore age and sex of this bear could not be determined. The black bear in the Elk Valley Coal area was caught a total of 4 times. Table 5: Black bear captures by sex and age classifications.

Kakwa Tower area Sex Cub of the Year Yearling Sub-adult Adult Total Male 0 0 1 3 4 Female 0 0 0 3 3 Total 0 0 1 6 7

Clear Hills area Male 0 0 0 5 5 Female 0 0 0 1 1 Total 0 0 0 6 6

Elk Valley Coal area Male 0 0 0 1 1 Female 0 0 0 0 0 Total 0 0 0 1 0 Grizzly Bear Vs. Black Bear Captures The numbers of grizzly bears vs. black bears captured have varied between years and areas (Table 6). Table 6: Number of grizzly bears versus black bears captured by year.

Year Grizzly Bears Black Bears Total 1999 24 5 29 2000 25 13 38 2001 29 10 39 2002 28 5 32 2003 28 14 42 2004 25 25 50 2005 23 22 45 2006 15 38 53 2007 16 15 31

9


10

POST CAPTURE FIELD WORK

GPS Location Data and Denning Our research team continued to collect grizzly bear movement data from collared animals up to the time of denning, which occurred this year during the first week of December. Using the exact locations of den sites we have also established weather monitoring stations at a number of grizzly bear dens to collect microsite weather data to relate to temperature data being collected from collars. The objective of this data collection effort is to further investigate possible triggers for den emergence over the winter months and in the spring. Clear Hills – additional data Since our capture success in the north eastern portion of the province (Clear Hills) was very limited (for a second consecutive year) we undertook an effort to gather additional data on the possible distribution and number of grizzly bears in this area. In June prior to the crew departing this area, we established 6 large bait sites where we positioned remote camera systems to take pictures of all activities that occurred at these sites during a one month period (June). In that time, we obtained photos of black bears, cougars, lynx, ravens, wolves, coyotes and 4 different grizzly bears. Vegetation and Diet Data Collection Research personnel worked in the field between May 1 and September 15, 2007 to collect comprehensive grizzly bear habitat use information and track human industrial activity. This work occurred within Weyerhaeuser’s Grande Prairie Forest Management Agreement Area south west of Grande Prairie, Alberta. Our vegetation sampling strategy was to visit one female grizzly bear GPS location per bear use day paired with an associated random location throughout the field season. Data collected at GPS locations included a detailed inventory of forest floor and under story vegetation, forest composition and structure, canopy cover measurements, distribution and abundance of grizzly bear foods, identification of grizzly bear and other wildlife sign, presence of human activity, habitat classification according to our remote sensing mapping criteria, and grizzly bear scat collection. Field personnel visited additional GPS locations to collect grizzly bear activity (feeding or bedding) information for those individuals using habitats in proximity to major road networks or where other industrial activities were occurring. Scat collection and analysis followed protocols established by the Foothills Research Institute Grizzly Bear Program used in previous field seasons. One scat per bear use day collected throughout the field season was analyzed for content based on our ability to positively identify food matter. The percent volume of discernable food items was quantified to determine their seasonal importance value. This information will be used to link seasonally important grizzly bear foods to specific habitat attributes and model the distribution of these resources across the study area. The habitat use information collected this field season will be used as part of our on-going effort to generate habitat classification maps. It will also be used to create and validate a Resource Selection Function model which is the probability of occurrence of grizzly bears within specific habitat types. We can then assess the current state of the landscape for


grizzly bears and make predictions regarding habitat quality and quantity and if industrial activity or natural factors such as Mountain Pine Beetle and forest fires alter portions of the land base. Berry Availability During the summer of 2007, 50 permanent berry plots were set-up in the Kakwa study area. These plots will be used to assess yearly berry abundance and variations. Five key berry species were recognized as important bear foods (Vaccinium membranaceum, Vaccinium myrtilloides, Vaccinium caespitosum, Vaccinium vitis-idea, and Shepherdia canadensis) and at least 10 berry plots were set-up per species (more than one species can occur in a particular plot). In order to reduce environmental variations, plots were only set-up in the 2 most abundant habitat types within the study area (conifer forests and regenerating stands). Furthermore, to reduce variations caused by aspect, plots were positioned on slopes of less than 10º. Berry plots were visited at least once in August and September to assess berry production and timing. Once most berries within a plot were ripe, they were picked and weighted for biomass calculations. All plots will be revisited consistently within the next few years to assess yearly variations.

In order to assess berry species abundance and variation over the landscape, random locations were visited within conifer and regenerating stands. Locations with and without berry species were marked and a percentage of plant cover was determined if key species occurred. Once a significant number of locations are visited, an accurate measure of berry species availability over the landscape will be derived.

Landscape Change and Road Use Intensity New and existing industrial activity was tracked within grizzly bear home ranges. This included quantifying the intensity of traffic on major road networks, identifying new physical changes to the landscape, and mapping existing infrastructure. Using hand held GPS units field personnel mapped the locations and status of well sites (new, active, and abandoned), roads, pipelines, seismic lines, and cut blocks from existing road networks. This spatial data will be used to verify landscape change layers produced by our remote sensing team from satellite imagery. Infrared digital trail cameras were installed throughout the study area to sample vehicular traffic on major access roads. Traffic data will be used to determine the type, volume, and time frame of road use activity. Monitoring and classifying vehicular traffic will allow us to determine the types of industrial activities occurring on the landscape and when they are taking place. This information could then be used to determine acceptable threshold levels of human industrial activity relative to the needs of grizzly bears and their ability to tolerate development and traffic volumes.

11


12

GRIZZLY BEAR HABITAT ENHANCEMENT TRIAL Introduction This interim report describes the surveys and treatments applied by the Grizzly Bear Habitat Enhancement Trial, conducted by the Foothills Research Institute Grizzly Bear Program (FRIGBP) in the Moose Mountain area in Kananaskis, Alberta, in July 2007. The objective of the Grizzly Bear Habitat Enhancement Trial is to determine:

i) whether grizzly bear foods such as buffaloberry (Shepherdia canadensis) and alpine sweet-vetch (Hedysarum alpinum) can be established on reclaimed oil and gas sites (Figure 4); and

ii) once established, whether grizzlies will be attracted to these sites.

Funding for this project was provided by Shell Canada and Husky Energy through the Moose Mountain Environmental Enhancement (MMEE) fund.

Figure 4: The target condition: an opening occupied by Shepherdia canadensis and Hedysarum alpinum.


Methods Reconnaissance of potential sites (Figure 5) was conducted on July 28, 2006 (sites 5 and 6), September 7, 2006 (sites 1, 3, and 4), and on July 14, 2007 (sites 2, 7, and 8). Sections of the Interconnect pipeline Right-of-Way (RoW) were also surveyed for potential treatment; however, the RoW is currently precluded from consideration for treatment due to safety concerns.

Figure 5: Surveyed sites in the Moose Mountain area.

13


Site 5: (NW-12-23-6-W5) This site is a non-productive well drilled by Petro-Canada in 2005 (Figure 6). There is a 250m access road. Site is flat, with scattered herbaceous cover. The soil was compacted but plantable. Lodgepole pine seedlings had been planted in spring 2006, on the site and access road. This site was considered suitable for treatment.

Figure 6: Site 5. Site 3: (SE-23-23-6-W5) This site occupies a ridge crest position (Figure 7). The soil is extremely compacted and not plantable. This site was not considered for treatment due to soil compaction.

Figure 7: Site 3.

14


Site 1: (SE-3-23-6-W5) This well site was reclaimed approximately 15 years ago (Figure 8). There are clusters of advanced spruce regeneration but most of the site is open and grassy. Slope is gentle but the soil is hard and compacted. There is a recreational trail 100m to the east of the opening. The 4 km access road has been reclaimed. This site was considered suitable for treatment.

Figure 8: Site 1. Site 2: (SE-20-23-7-W5) This well site was reclaimed approximately 15 years ago (Figure 9). Very steep side slope on

2.2 km reclaimed access road, not accessible by ATV. Soil is extremely compacted, not plantable. There is scattered natural Lodgepole pine regeneration despite abundant cones in adjacent stands. Slope is steep (~ 60%). Primary vegetation cover is grass (Phleum spp) and locoweed (Oxytropis monticola). Minimal bear sign was observed (digging, scat) along stream in regenerating stand adjacent to access road. This site was not considered suitable for treatment due to soil compaction.

Figure 9: Site 2.

15


Site 7: (SE-23-22-6-W5) This well site was reclaimed approximately 15 years ago (Figure 10). There is a 1.9 km reclaimed access road. The site is extremely compacted and is covered in grass and locoweed. There is a recreational trail 50 m to south. This site was not considered suitable for treatment due to soil compaction.

Figure 10: Site 7. Site 4: (NE-17-23-6-W5) This reclaimed well site lies 3.1 km past the Interconnect pipeline RoW on a reclaimed road (Figure 11). This site is extremely compacted with minimal vegetation cover (grass) and was not considered suitable for treatment due to soil compaction.

Figure 11: Site 4.

16


17

Site 6: (SE-3-23-6-W5) This reclaimed well site lies along the Interconnect pipeline route (Figure 12). The site is flat, and the soil is extremely compacted. Vegetation cover is clover (Trifolium spp). This site was not considered suitable for treatment due to soil compaction.

Figure 12: Site 6. Site 8: (NW-15-24-6-W5) This reclaimed well site lies at the end of a 1.2 km reclaimed access road (Figure 13). The site is flat, and the soil is extremely compacted. Vegetation cover is clover (Trifolium spp).

The site was not considered suitable for treatment due to soil compaction.

Figure 13: Site 8.


Results Site 5 was planted on July 9th and 10th, 2007. The Hedysarum alpinum seedlings were in a fragile condition in pots, and could not be transported the 4 km to the site by ATV (Figure 14). Therefore the decision was made to plant all the Hedysarum at this site. Although there were approx. 900 pots delivered, there was poor germination and only 580 seedlings were planted. Almost all were planted along the east, north, and west edges of the opening. All were staked with 18” pigtails. A rainstorm in the morning of the 9th wet the soil, but the weather for the remainder of the plant was hot and dry.

Figure 14: Hedysarum seedlings. The Shepherdia seedlings were more robust and came in 412 plugs (Figure 15). Root collar caliper averaged 2 mm and heights ranged from 15 to 25 cm. A total of 1080 Shepherdia seedlings were planted on Site 5 on July 10th.

18


Figure 15: Buffaloberry seedlings. Growth-limiting factors on this site are soil compaction that may inhibit root development, and seedling desiccation due to hot, dry weather at the time of planting. The Lodgepole pine seedlings planted the previous year were suffering severe mortality (~80%) and damage, likely due to frost or drought. A post-treatment survey of this site will be conducted in July to assess seedling condition. The first 30m of the access road had been tilled (below, right), creating much better growing conditions than on the well site itself (Figure 16).

Figure 16: Tilled vs. non-tilled ground.

19


Site 1 was planted with 1080 Shepherdia seedlings on July 11th (Figure 17). Of these, 900 were staked with 18” pigtails and the remainder were flagged. The primary growth-limiting factors on this site are competition from grass, soil compaction that may inhibit root development, and seedling desiccation due to hot, dry weather at the time of planting. Figure 17: Planting on Site 1. Recommendations The primary factor that precludes reclaimed well sites from treatment aimed at restoring grizzly bear habitat is severe soil compaction, noted on six of the eight sites surveyed. Mechanical tillage may reverse this condition but such treatments are very expensive, and may be of limited effectiveness given that grasses have become well-established on all sites. It is recommended that future reclamation efforts attempt to preserve soil productivity by minimizing compaction, and that a vegetation management regime that includes the establishment of grizzly bear forage species be implemented at the time of reclamation.

20

Chapter 2: Remote Sensing Mapping and Research

CHAPTER 2: REMOTE SENSING MAPPING AND RESEARCH

UPDATE

Greg McDermid4, Adam McLane4, Adam Collingwood5, Jennifer Hird4, Adrian Faraguna4, David Laskin4, Julia Linke4, Jerome Cranston3, Xulin Guo5,

and Steven Franklin5

3Arctos Ecological Services, Hinton, AB; 4Foothills Facility for Remote Sensing and GIScience, Department of Geography, University of Calgary; 5Environmental Remote Sensing Laboratory, Department of Geography, University of Saskatchewan;

INTRODUCTION AND OVERVIEW 2007/08 marks the eighth year in which researchers and technicians from the University of Calgary and the University of Saskatchewan have contributed remote sensing research and mapping initiatives in the Foothills Research Institute Grizzly Bear Program (FRIGBP). Activities in the past fiscal year can be divided broadly into two main categories: (i) those related to deliverable products, generated primarily by technical staff, and (ii) those related to research pursuits, performed primarily by graduate students. In this report, we summarize the activities taking place within the Foothills Facility for Remote Sensing and GIScience at the University of Calgary and the Environmental Remote Sensing Laboratory at the University of Saskatchewan. On the deliverables front, the remote sensing/mapping team completed two significant milestones (major contributors are listed in brackets):

1. Phase 7 map products (David Laskin and Adam McLane) 2. Annual Changes in Landcover, vegetation, and landscape structure in BMAs 3 and 4,

1998-2005 (Julia Linke, Adam McLane, David Laskin, and Jerome Cranston) On the research front, efforts this year were concentrated on five additional topics:

3. LiDAR processing for canopy closure and vegetation structure (Adam McLane) 4. Noise reduction of NDVI time series for characterizing vegetation phenology

(Jennifer Hird) 5. Classification of agricultural areas (Adam Collingwood) 6. Relationships Between Landscape Spatial Properties and Grizzly Bear Presence in

Agricultural Areas (Adam Collingwood) 7. Mountain pine beetle susceptibility: a review of remote sensing opportunities (Adrian

Faraguna) Information on the status and achievements of each of the seven topic areas above are contained in the following sections.

21


22

PHASE 7 MAP PRODUCTS Introduction Since its launch in 1999, the FRIGBP has evolved through seven phases of expansion in its objective to map the 228,000 km2 that comprise grizzly bear range in province of Alberta (Figure 1). Initial efforts from 1999 to 2002 were constrained to the 10,000 km2 Phase 1 and 2 study areas centered on the Yellowhead ecosystem, including the eastern portion of Jasper National Park and adjacent public lands. In this era, remote sensing activities were concerned primarily with evaluating decision rules capable of handling the mixed sets used to generate the hybrid land use/land cover maps that were the focus of this initial work. For example, Franklin et al (2001) described the Integrated Decision Tree Approach (IDTA) to classifying mixed land use and land cover categories using combined satellite and GIS data sources. The IDTA selectively merged supervised, unsupervised, and GIS decision rule criteria with a three-level classification scheme. The hierarchical, multi-step nature of the strategy enabled class-specific decision rules (e.g. water separated from shadow on the basis of slope and elevation) and variable sets to be applied in a selective manner best suited to the particular level of classification. Used in conjunction with tasseled cap greenness maps – rough indicators of biomass or vegetation amount – these products contributed to ground-breaking efforts to map grizzly bear habitat in the mountains and foothills of Alberta (Nielsen et al, 2003; Nielsen et al, 2006). While these early activities were certainly productive, our current approach to delivering dynamic and effective remote sensing products within the FRIGBP has evolved considerably. As McDermid et al (2005) described, the multi-disciplinary collaboration required for large-area ecological applications such as the FRIGBP is challenging, and often constrained by a lack of common understanding between resource managers with extensive knowledge of ecology and remote sensing scientists with backgrounds in geography. Our experiences have revealed a series of valuable lessons that have helped to improve the effectiveness of our remote sensing efforts, even as we pursue them under increasingly challenging conditions.

Lesson #1: Minimize your reliance on categorical maps that are inflexible and incapable of change. Classification techniques, while far and away the dominant information-extraction paradigm in remote sensing, have two limitations that should be borne in mind at all times in collaborative wildlife projects: (i) the information products they produce are categorical in nature, with pixel values assigned at a nominal or ordinal level; and (ii) once defined, class boundaries are very difficult to adjust. In the FRIGBP, we have chosen to avoid categorical maps where possible in favour of ratio-level end products that maintain their flexibility.

Lesson #2: Eliminate spatial inconsistencies that undermine the confidence of end users. Spatial inconsistencies across political and jurisdictional boundaries are one of the major limitations of traditional inventory-based information sources, and a commonly-cited strength of satellite-derived products. However, remote sensing-based maps are also subject to spatial issues, including shadows, topographic effects, and seam lines between image tiles in a multi-scene mosaic. Particular care must be taken to reduce or eliminate these issues as much as possible, in order to maximize the utility of end products and maintain the confidence of end users.


Lesson #3: Try not to over-simplify a complex phenomenon. Unfortunately, many remote sensing products can be criticized for presenting an overly-simplified depiction of vegetation and other aspects of the natural landscape. The natural world is complex, and cannot be captured by a single map product – particularly a categorical one produced exclusively by classification! In our current approach, we produce a series of map products, each focused on individual habitat attributes.

Figure 1: The seven phases of the FRIGBP, and the years in which they were adopted.

23


24

As a result of these and other hard-won experiences, we have formalized (McDermid et al, 2005) and then applied (McDermid et al, 2008) an application framework designed to link the needs of ecologists with the tools and information-extraction strategies of remote sensing scientists. Guided by hierarchy theory and the remote sensing scene model, the framework calls for the development of multi-attribute information databases in which individual land and vegetation attributes are identified and mapped separately using a variety of scale-sensitive data sources and extraction strategies. In its current implementation, the data products produced by the remote sensing team are comprised of four groups: (i) land cover/physiognomy, (ii) crown closure, (iii) species composition, and (iv) vegetation phenology. The approach differs from that used in most other large-area remote sensing mapping programs in at least three important respects. First is the recognition that land and vegetation information exists at a variety of spatial and temporal scales, and that no single map is capable of capturing the full range of variability observed in nature. While it would be foolish to claim that even a complex multi-layer information base is capable of simulating reality, we do contend that an attribute-based approach that attempts to identify and account for the major spatial, structural, and temporal patterns observed on the landscape in a series of spatial databases is more appropriate than a single catch-all map product. The second major difference involves the selective application of multiple information-extraction techniques that are sensitive to the scale and data levels at which these attributes appear on the landscape. Finally, a concerted effort has been made to map selected attributes as continuous variables wherever possible, in order to achieve maximum flexibility with the completed product.

Methods An overview of the general procedure followed to generate the standard base map products – land cover, crown closure, species composition, and phenology – is shown in Figure 2. For a detailed description of the entire methodological process, please see McDermid et al, 2008. To summarize, we begin with raw data sets comprised of digital data from three main sources: MODIS imagery, Landsat imagery, and a digital elevation model. Each raw data set undergoes a battery of pre-processing routines designed to reduce geometric and radiometric errors and produce digital variables suitable for further analysis. We then employ a variety of empirical modeling techniques designed to explain the variance patterns recorded across a sample of training sites containing ground biophysical observations acquired through field work. A ten-class land cover model (Table 1) is obtained through object-oriented classification, while generalized linear modeling approach is used to model continuous-variable estimates of crown closure (0-100%) and tree species composition (% coniferous tree cover). Once obtained, the various models are inverted to generate wall-to-wall estimates of the final map products across the entire study area. A custom software program written in IDL is used to extract a series of ten phenological attributes (Table 2) from the MODIS time series data.


Figure 2: Flow diagram highlighting the general methods used to create the four base map products: land cover, crown closure, species composition, and phenology.

Table 1: Land cover class descriptions.

Land Cover Class Description

Upland Trees Vegetated sites with tree crown closure >5%; dry or mesic

Wetland Trees Vegetated sites with tree crown closure >5%; wet

Upland Herbs Vegetated sites with herbaceous cover >5%; dry or mesic

Wetland Herbs Vegetated sites with herbaceous cover >5%; wet

Shrubs Vegetated sites with shrub cover >5%;

Shadow Ground features obscured by shadow

Water Non-vegetated sites with water cover >50%

Barren Land Non-vegetated sites with barren cover >50%

Snow/Ice Non-vegetated sites with snow/ice cover >50%

Cloud Ground features obscured by cloud cover

25


Table 2: Phenological product descriptions.

Metric Meaning/Description

Timing of SOS Composite period in which growing season starts

Timing of EOS Composite period in which growing season ends

LGS Number of composite periods comprising the growing season

Maximum NDVI Maximum level of photosynthesis reached during the growing season

Timing of Maximum NDVI Composite period in which maximum level of photosynthesis is reached

NDVI Amplitude Range between maximum level of photosynthesis, and average of pre- and post-growing season minimum levels of photosynthesis

Maximum Green-Up Maximum rate of increase in photosynthetic activity during spring green-up

Timing of Maximum Green-Up Composite periods between which maximum rate of increase in photosynthetic activity occurs

Average NDVI Mean level of growing season photosynthesis

Integrated NDVI Cumulative photosynthetic production during the growing season

Results and Discussion Field and laboratory efforts from personnel on the FRIGBP mapping team has enabled the production of high-quality maps of land cover, crown closure, species composition, and vegetation phenology over the Foothills Research Institute Grizzly Bear Program’s Phase 7 study area, covering almost 25 million hectares of rugged, multi-jurisdictional terrain in western Alberta. Land Cover The composite land cover map over the Phase 7 study area is shown in Figure 3. The overall accuracy of the map varies from region to region, ranging from 75% (Kappa=0.46) in the Phase 5 expansion area to 88% (Kappa=0.70) in the Phase 6 expansion area, with the results predictably tied to the number of ground points available in each workzone. Upland treed areas were generally well classified across all regions except phase 5, where many upland treed objects were mis-classified as wetland herb. We attribute this to the thinner broadleaf forests in the area displaying spectral values to other wetland classes. Smaller amounts of confusion between wetland treed and upland treed classes is evident in phase 6, explained in part by the relatively small sample size of the wet-treed class in that area. We recommend further refinement in the delineation of wetland classes in order to increase the accuracy of future map products.

26


Figure 3: Phase 7 land cover map.

27


Crown Closure A seamless mosaic of the Phase 7 crown closure model is shown in Figure 4. Overall accuracy of a four-class configuration of crown closure in the Phase 6 extension was calculated at 78%.

Figure 4: Phase 7 crown closure map.

28


29

Species Composition The phase 7 species composition map is shown in Figure 5. Accuracy of product in a four-class configuration in the Phase 6 extension was calculated at 80%.

Figure 5: Phase 7 species composition map.


Vegetation Phenology The vegetation range product reflects the magnitude of seasonal vegetation dynamics, where NDVI change indicates the rates and/or levels of change occurring in vegetated surfaces between composite periods (Figure 6). Other NDVI composite surfaces include Minimum, Maximum, and the timing of maximum NDVI, reflecting the per-pixel level and time during which maximum/minimum surface vegetative production is reached during the growing season.

Figure 6: 16-day NDVI phenology composites for the 2005 growing, Phase 7.

30


Discussion The end result of the multi-attribute mapping approach used by the FRIGBP to map grizzly bear habitat is a multi-layer database of land and vegetation information products, including layers for land cover, crown closure, species composition, and phenology. This rich, multi-layer information base stands in sharp contrast to the single-layer classification maps generated by most large-area remote sensing projects, and adds significant new dimensions of quality and flexibility to the final product. By separating broad categorical information products that vary at the stand level – land cover – from high-level attributes like LAI, crown closure, and species composition that vary continuously across the landscape, two key objectives have been realized. First, a framework has been established that matches diverse information needs with appropriate and effective image processing techniques. Based on hierarchy theory and the remote sensing scene model, this framework provides a foundation for complex information extraction that is more effective than the indiscriminate application of classification techniques. Second, the generation of multiple – and, where appropriate, continuous – estimates of individual attributes across the study area has resulted in the creation of a land cover/vegetation information database with exceptional flexibility. By maintaining high-order information, the system provides managers and researchers with divergent needs the opportunity to define categories that suit their individual application. By re-coding, re-classing, and combining information from the vegetation/land cover information base with geospatial layers from other GIS sources, the potential exists to produce a large number of information products potentially suitable for a broad range of habitat mapping/environmental management objectives. For example, Figure 7 shows examples of three different map products generated over a small portion of phase 3 using simple procedures in a GIS environment. Figure 7A depicts vegetation structure with three classes of forest species composition: pure conifer (>70% coniferous), pure broadleaf (<30% broadleaf), and mixed forest (30-70% coniferous), but only two crown closure classes: open (<50% crown closure) and closed (>50% crown closure). The prevalence of old-growth conifer forests in this portion of the study area – the Athabasca valley in Jasper National Park – creates a map dominated by the closed coniferous forest class. However, the simple process of re-generating the map with a three-class configuration of crown closure: open (<30% crown closure), medium (30-70% crown closure), and dense (>70% crown closure) produces a product – Figure 7B with much more structural detail in the forests common to in this part of the study area. Contrasting these two structural vegetation maps is the phenological view of vegetation shown in Figure 7C. In this case, a composite map was generated to highlight three categories of LAI change across the summer growing season: vegetation senescence (leaf area decreasing between the early- and late-hyperphagia periods), no change (leaf area remaining the same between two periods), and vegetation growth (leaf area increasing between the two periods). Configured in this manner, the information base presents a markedly different view of the study area, highlighting the substantial increase of foliage observed in the upper valley slopes.

31


A

B

C

Figure 7: Examples of the flexible composite map capability of the current land/vegetation information base. Maps shown here cover a small portion of the Phase 3 study area and were generated through simple raster calculation and overlay procedures in a GIS environment.

32


Summary and Conclusions Field and laboratory efforts from personnel in the remote sensing team at the Foothills Facility for Remote Sensing and GIScience at the University of Calgary and the Environmental Remote Sensing Laboratory at the University of Saskatchewan has enabled the production of high-quality maps of land cover, crown closure, species composition, and NDVI phenology over the FRIGBP’s Phase 7 study area, covering more than 25 million hectares of rugged, multi-jurisdictional terrain in western Alberta. The effort completes the team’s work to provide baseline mapping of habitat conditions over the practical range of grizzly bears in the province of Alberta. This work is the basis for the following forthcoming journal article: McDermid, G.J., D.N. Laskin, A.J. McLane, A.D. Pape, S.E. Franklin, and G.B. Stenhouse: Mapping and update of vegetation and land cover for grizzly bear research and conservation. Canadian Journal of Remote Sensing, in preparation.

33


34

ANNUAL CHANGES IN LANDCOVER, VEGETATION, AND

LANDSCAPE STRUCTURE IN BMAS 3 AND 4, 1998-2005 Increasing levels of human activity in west-central Alberta surrounding the timber, coal, and petroleum industries have lead to widespread growth in anthropogenic disturbance features on the landscape, and corresponding changes in land cover structure and composition. A primary goal of the mapping team in this project has involved the accurate detection, mapping, and quantification of human-induced landscape changes at a temporal resolution matching our grizzly bear health and stress data sets as closely as possible. We employed satellite imagery and change-detection procedures to generate spatially-explicit layers of land cover, vegetation, and various anthropogenic disturbance features across the Health Study Area covering BMAs 3 and 4, then quantified their annual levels and trajectories using a variety of summary statistics and landscape metrics in a classic landscape pattern analysis (LPA).

Several studies have reported large bias in landscape metrics caused by classification error in single-date maps. In particular, Langford et al (2006) warned that there is potential for large metric errors in nearly every LPA study ever published. This caveat also applies to multi-temporal LPA studies where the land-cover maps for each time period are generated and analyzed independently from one another. A significant portion of this work –though not substantially reported here – involved dealing with the challenges and issues involved in performing LPA with multi-temporal satellite imagery.

Since classification errors in one date will be independent of errors in the other dates, this procedure is prone to creating spurious change. One strategy for reducing this problem is to avoid classifying each image independently, and instead update (project forward in time) or backdate (project backward in time) a reference map by only re-labeling raster cells in areas where change has been detected, thus drastically reducing the area over which spurious changes may be introduced. The detection of change between two remote sensing (RS) images from the same area acquired at different periods can take on a variety of forms, but generally involves comparing spatially coincident units and their allocation to change and no change. However, this procedure is not without challenges.

Backdating an existing reference land-cover map (T0) to reflect the land cover conditions at a previous time period (T-n) entails the removal or re-labelling of change patches at T-n that appeared on the landscape prior to the reference year T0 (Figure 8). Updating an existing reference map to a future time period (T+n), on the other hand, requires the insertion or re-labelling of change patches at T+n that have appeared on the landscape after T0 (Figure 8). In both cases, change patches are detected and labelled through bi-temporal change detection. In the backdating case, two situations are possible: (i) land-cover patches spatially coinciding with change patches will be fully removed from the reference map if the change patch originated after T-n, or (ii) land-cover patches spatially coinciding with change patches will remain in the backdated map but can be assigned a new attribute label reflecting an earlier successional stage if the change patch originated before T-n. In the updating case, the situation is analogous, with new change patches being inserted to the reference map and existing patches undergoing attribute changes related to succession.


Figure 8: Backdating and updating a reference map through change analysis. Methods We acquired a series of 22 Landsat Thematic Mapper and Enhanced Thematic Mapper Plus images (Table 3) to track annual change patterns across BMAs 3 and 4, located in the core portion of the Phase 7 study area, from 1998 to 2005. Linear disturbance features (roads, pipelines, rail) were mapped primarily through manual digitizing and update of available GIS layers, while area-based features (cut blocks, mines, and wellsites) were delineated through change detection of co-registered satellite imagery. Annual disturbance layers were visually inspected and manually corrected where necessary. Disturbance objects were transformed to land cover classes using decision rules (e.g. road features = barren; pipelines = herbaceous), and spatially mosaicked to create annual update layers. Finally, we overlaid the annual update layers on the co-registered 2003 base land cover map to backdate and update land cover products for each year of interest.

35


Table 3: Landsat scenes acquired and processed for quantifying human disturbance, land cover, and landscape structure from 1998 to 2005.

Landsat Path/Row Image Acquisition Date Sensor September 5, 1998 Landsat 5 TM September 8, 1999 Landsat 5 TM August 17, 2000 Landsat 7 ETM+ September 14, 2001 Landsat 7 ETM+ August 23, 2002 Landsat 7 ETM+ September 3, 2003 Landsat 5 TM August 12, 2004 Landsat 7 SLC-Off

45/23

July 22, 2005 Landsat 5 TM August 29, 1998 Landsat 7 ETM+ August 24, 1999 Landsat 7 ETM+ September 27, 2000* Landsat 7 ETM+ September 14, 2001 Landsat 7 ETM+ June 13, 2002 Landsat 7 ETM+ July 10, 2003 Landsat 5 TM August 13, 2004 Landsat 5 TM

44/23

September 13, 2005 Landsat 5 TM June 22, 2002 Landsat 7 ETM+ June 17, 2003 Landsat 5 TM June 19, 2004 Landsat 5 TM 43/24

August 25, 2005 Landsat 5 TM 2002 N/A - July 10, 2003 Landsat 5 TM 2004 N/A - 44/24

September 17, 2005 La80dsat 5 TM The thematic accuracy of the change features were assessed at two levels: (i) change identification, and (ii) change labeling. Change identification assessed our ability to accurately separate change areas from no change, using 178 test points evaluated through manual interpretation of temporally-coincident, high-spatial-resolution orthophotos. Change labeling accuracy was determined with an additional 256 test points distributed in a stratified random sample, allocated proportionally. A spatio-temporal analysis of landscape structure from the period 1998 to 2005 was conducted using Fragstats 3.3. We derived the landscape metrics Mean Patch Size, Mean Nearest Neighbor Distance, and Edge Density from the ten-class landcover map for each year of interest. The same metrics were also derived from a class perspective to track the patterns of forest/non-forest through time. In order to investigate the range of trajectories in landscape structure observed across time at the local scale, we selected three 13x13 km2 case study areas from the Health study area

36


selected to represent low (Area 1), moderate (Area 2), and high (Area 3) levels of change (Figure 9).

Figure 9: Location of the three 13 x 13 km case study areas in the North Health study area, representing three levels of relative land cover change between 1998 and 2005. Results and Discussion Landcover 1998-2005 Samples of the resulting land cover maps generated over BMAs 3 and 4 from 1998 to 2005 are shown in Figure 10. Results of the accuracy assessment suggested efficient thematic performance of the procedure, both in the identification (100% overall accuracy; Kappa=1.0) and labeling (93% overall accuracy; Kappa=0.889) of change features, and a visual inspection revealed very few slivers or other spatial anomalies, lending good confidence to the quality of the map products and resulting LPA.

37


Figure 10: Samples of changing land cover over a small portion of BMAs 3 and 4, 1998 to 2005.

38


Human Disturbance Features 1998-2005 A summary of the amounts, densities, and proportions of key human disturbance features across the Health Study Area for the time period 1998 to 2005 can be found in Table 4. These same trends are shown graphically in Figure 11. It is important to note, however, that that the estimates from 1998 to 2001 were derived exclusively from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined. The changing base areas over which estimates were summarized reflects the varying spatial extents over which grizzly bear health and stress data were available. Grizzly bear captures in the North Health study area (BMA 3) span the years 1998 to 2005, while those in the South Health study area (BMA 4) range only from 2002 to 2005. This changing base area leads to a number of apparent incongruities when values are summarized across the entire study area. For example, at this scale, pipeline density appears to drop from 0.138 km/km2 in 2001 to 0.127km/km2 in 2002. However, the pattern merely reflects the incorporation of new areas in BMA 4 from 2002 onward, where pipeline densities (and other disturbance features) are generally lower. These apparent anomalies only appear in metrics summarized at very broad scales, and do not affect subsequent statistical analyses taking place at the individual home-range scale.

For each year, we observed substantial increases in human disturbance features, clearly reflecting the growing impact of fast-developing resource extraction industries in west-central Alberta grizzly bear habitat. A brief summary of each disturbance element is presented in the following subsections.

Table 4: A summary of road density, cutblock proportion, wellsite density, pipeline density, and mine area across BMAs 3 and 4 from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Road Density (km/km2)

Cutblock Proportion (%)

Wellsite Density (#/km2)

Pipeline Density

(km/km2)

Mine Area (km2)

1998 0.35 2.04 0.09 0.126 1.10 1999 0.36 2.33 0.10 0.130 1.10 2000 0.37 2.64 0.11 0.136 3.50 2001 0.39 2.97 0.12 0.138 6.58 2002 0.42 3.03 0.11 0.127 10.64 2003 0.44 3.36 0.12 0.129 12.20 2004 0.44 3.68 0.13 0.131 14.26 2005 0.45 4.04 0.14 0.132 15.73

39


Road Density 1998-2005

0.00

0.10

0.20

0.30

0.40

0.50

1998

1999

2000

2001

2002

2003

2004

2005

Year

Roa

d D

ensi

ty (k

m/k

m2)

Cutblock Proportion 1998-2005

0.000.501.001.502.002.503.003.504.004.50

1998

1999

2000

2001

2002

2003

2004

2005

Year

Cut

bloc

k Pr

opor

tion

(%)

Wellsite Density 1998-2005

0.000.020.040.060.080.100.120.140.16

1998

1999

2000

2001

2002

2003

2004

2005

Year

Wel

lsite

Den

sity

(#/k

m2)

Pipeline Density 1998-2005

0.115

0.120

0.125

0.130

0.135

0.140

1998

1999

2000

2001

2002

2003

2004

2005

Year

Pipe

line

Den

sity

(km

/km

2)

Mine Area 1998-2005

0.002.004.006.008.00

10.0012.0014.0016.0018.00

1998

1999

2000

2001

2002

2003

2004

2005

YearM

ine

Are

a (k

m2)

Figure 11: Trends in road density, cutblock proportion, wellsite density, pipeline density, and mine area across BMAs 3 and 4 from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

40


Roads and Road Density Road density changed from 0.35km/km2 in 1998 to 0.45 km/km2 in 2005 (Table 5): an increase of more than 28%. We documented an average annual increase of approximately 100m/km2 over the time period examined. The total length of roads mapped over the Health Study Area (BMAs 3 and 4) was 26,357 km in 2005. Table 5: Road length, change in road length, road density, and change in road density from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Length (km) ∆ Length (km) Density (km/km2)

∆ Density (km/km2)

1998 13377.4 - 0.35 - 1999 13794.1 416.9 0.36 0.01 2000 14351.3 557.2 0.37 0.01 2001 14959.3 608.0 0.39 0.02 2002 24560.6 9601.3 0.42 0.03 2003 25361.9 801.3 0.44 0.01 2004 25640.2 278.3 0.44 0.00 2005 26357.0 716.8 0.45 0.01

Cutblocks and Proportion of Study Area Covered by Cutblocks The proportion of the study area that is covered by cut blocks changed from 2.04% in 1998 to 4.04% in 2005 (Table 6): an increase of about 98%. The rate of average proportional increase was calculated at 0.29%, with total cut block coverage in the study area exceeding 2300 km2 in 2005.

Table 6: Cutblock area, change in cutblock area, proportion of study area covered by cutblock, and change in proportion of study area that is cutblock from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Area (km2) ∆ Area (km2) Proportion (%)

∆ Proportion (%)

1998 780.9 - 2.04 - 1999 891.5 110.6 2.33 0.29 2000 1012.0 120.5 2.64 0.32 2001 1136.6 124.6 2.97 0.32 2002 1760.4 632.8 3.03 0.06 2003 1951.5 191.1 3.36 0.33 2004 2137.5 186.0 3.68 0.32 2005 2347.0 209.5 4.04 0.36

41


Wellsites and Wellsite Density The density of wellsites occurring in the study area changed from 0.09/km2 in 1998 to 0.14/km2 in 2005 (Table 7): an increase of about 56%. The rate of average change in wellsite density over the entire study area was calculated at 0.007 wellsites/km2, with the total number of wellsites estimated at 8,162 in 2005.

Table 7: Number of wellsites, change in number of wellsites, wellsite density, and change in wellsite density from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Number Change Density (#/km2)

∆ Density (#/km2)

1998 3482 - 0.09 - 1999 3749 267 0.10 0.007 2000 4057 308 0.11 0.008 2001 4431 374 0.12 0.010 2002 6612 2181 0.11 -0.002 2003 7038 426 0.12 0.007 2004 7524 486 0.13 0.008 2005 8162 638 0.14 0.011

Pipelines and Pipeline Density Pipeline density changed from 0.126 km/km2 in 1998 to 0.132 km/km2 in 2005 (Table 8): an increase of about 5%. We documented an average annual increase of approximately 10m/km2 over the time period examined. The total length of pipelines mapped over the Health study area (BMAs 3 and 4) was 7,650 km in 2005.

Table 8: Pipeline length, change in pipeline length, pipeline density, and change in pipeline density from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Length (km) ∆ Length (km) Density (km/km2)

∆ Density (km/km2)

1998 4819.7 - 0.126 - 1999 4977.0 157.3 0.130 0.004 2000 5200.4 223.4 0.136 0.006 2001 5278.0 77.6 0.138 0.002 2002 7400.3 2122.3 0.127 -0.010 2003 7501.1 100.8 0.129 0.002 2004 7598.1 97.0 0.131 0.002 2005 7650.5 52.4 0.132 0.001

42


Mine Area The total amount of mine area changed from 1.10 km2 in 1998 to 15.73 km2 in 2005 (Table 9): a 13-fold increase. We documented an average annual change of approximately 2 km2

over the time period examined. The total mine area mapped over the Health study area (BMAs 3 and 4) was 15.73 km2 in 2005.

Table 9: Mine area and change in mine area from 1998 to 2005. Please note that estimates from 1998 to 2001 were derived from the North Health study area, while those from 2002 to 2005 were derived from the North and South Health study areas combined.

Year Area (km2) ∆ Area (km2) 1998 1.10 - 1999 1.10 0.00 2000 3.50 2.40 2001 6.58 3.08 2002 10.64 4.06 2003 12.20 1.56 2004 14.26 2.06 2005 15.73 1.47

Changes in Landscape Structure 1998-2005: Local Scale As expected, landscape structure and conditions displayed significant spatial variability across the study area, both in terms of their mean annual amounts of change (Table 10) and structural trajectories over time (Figure 12). The mean annual area of forest change varied from 38.5 hectares distributed across 3.2 new patches per year in Area 1 to 322.3 hectares distributed across 21.3 new patches per year in area 3. Annual trends over time revealed consistent increases in edge density and decreases in mean patch size. Despite the wide range of cumulative loss of forested areas, the selected landscape metrics appear fully capable of tracking logically-consistent patterns of landscape structure over time, efficiently revealing the changing patterns of land cover associated with increasing human activities.

Table 10: Temporal occurrence and mean annual amounts of forest change patches across the three selected 13 x 13 km case study areas, reflecting low, moderate and high levels of change.

Occurrence of Annual Forest Change Patches

Annual Amount of Forest Change Patches

Case Study Area Number of Years

(n) Time Frame

(Years)

Mean Annual Area of Change

(ha)

Mean Number of Change Patches

(n)

1 – Low Change 4 1998-2000; 2001-2003 38.5 3.2

2 – Mod. Change 7 1998-2005 252.5 10.6 3 – High Change 7 1998-2005 322.3 21.3

43


Figure 12: Cumulative change in A) forest area (i.e. forest loss), B) percent edge density, and C) mean patch size, between the years 1998 and 2005 in three case study areas of low (case study 1), moderate (case study 2), and high ( case study 3) annual mean forest change.

44


Changes in Landscape Structure 1998-2005: Regional Scale A summary of the selected landscape metrics tabulated across the Health Study Area for the time period 1998 to 2005 can be found in Table 11. These same trends are shown graphically in Figure 13. In all cases, the metrics seem to indicate a pattern of increased fragmentation that one would expect from the transition of large natural forest patches to roads, cutblocks, wellsites, and other human disturbance features. Edge density – linear distance of edge per unit area of landscape – changed from 43.7 m/ha in 1998 to 48.5 m/ha: an increase of 11%. Over the same time period, mean patch size changed from 42.0 ha to 37.1 ha: a decrease of 12%. Mean nearest neighbour – a measure of edge-to-edge distance – changed from 271 m in 1998 to 244m in 2005: a decrease of 10%. Table 11: Edge density, mean patch size, coefficient of variation of mean patch size, mean nearest neighbour, and coefficient of variation of mean nearest neighbour from 1998-2005.

Year Edge

Density (m/ha)

Mean Patch Size (ha)

CV of Mean Patch Size

(ha)

Mean

Nearest Neighbour

(m)

CV of Mean

Nearest Neighbour

(m)

1998 43.7 42.0 3416 271 233 1999 44.3 41.5 3434 269 234 2000 45.0 40.8 2457 265 235 2001 45.6 40.1 2449 261 237 2002 46.2 29.5 3462 257 238 2003 47.0 38.7 3532 253 239 2004 47.6 38.1 3436 250 241 2005 48.5 37.1 3425 244 242

45


41.0

42.0

43.0

44.0

45.0

46.0

47.0

48.0

49.0

1998 1999 2000 2001 2002 2003 2004 2005

Year

Edge

Den

sity

(m/h

a)

34.0

35.0

36.0

37.0

38.0

39.0

40.0

41.0

42.0

43.0

1998 1999 2000 2001 2002 2003 2004 2005

Year

Mea

n Pa

tch

Size

(ha)

230

235

240

245

250

255

260

265

270

275

1998 1999 2000 2001 2002 2003 2004 2005

Year

Mea

n N

eare

st N

eigh

bour

(m)

Figure 13: Trends in edge density, mean patch size, and mean nearest neighbour in BMAs 3 and 4 from 1998 to 2005.

46


Summary We employed satellite imagery and bi-temporal change-analysis procedures to generate spatially-explicit layers of land cover, vegetation, and various anthropogenic disturbance features across BMAs 3 and 4, which have formed the basis for a multi-year landscape pattern analysis documenting the structural changes in grizzly bear habitat across its core range in western Alberta. This work is the basis for the following forthcoming journal articles: McDermid, G.J., A.D. Pape, J. Linke, D.N. Laskin, A.J. McLane, and S.E. Franklin: Object-based approaches to change analysis and thematic map update: challenges and limitations. Canadian Journal of Remote Sensing, in review. Linke, J., G.J. McDermid, A.D. Pape, A.J. McLane, D.N. Laskin, M. Hall-Beyer, and S.E. Franklin: The influence of patch delineation mismatches on multi-temporal landscape pattern analysis. Landscape Ecology, in review. Linke, J., G.J. McDermid, A.J. McLane, D.N. Laskin, A.D. Pape, M. Hall-Beyer, and S.E. Franklin: A framework for temporally- and categorically-dynamic land cover maps using integrative approaches. Photogrammetric Engineering and Remote Sensing, in preparation.

47


LIDAR PROCESSING FOR CANOPY CLOSURE AND VEGETATION STRUCTURE

Introduction Canopy closure can be defined as the proportion of the sky hemisphere obscured by vegetation when viewed from a single point (Jennings et al, 1999). It has been a particular interest of forest ecologists studying the effects of forest succession, natural disturbance, silvicultural prescriptions (stand-level operations plans) and timber harvesting on the survival, pattern, and diversity of understory plants and trees, and represents one of the base products used by FRIGBP ecologists to characterize grizzly bear habitat. Canopy closure has traditionally been measured in the field through the use of ground-based optical tools such as the spherical densiometer and hemispherical photography: labour-intenstive strategies that consume significant project resources. This work is an investigation into how canopy closure can be estimated from discrete-return profiling LiDAR (Light Detection and Ranging) data: an efficient, reliable, and repeatable information source that has the potential to improve the measurement of canopy closure and other aspects of vegetation structure while reducing project costs. Similar to canopy closure is the concept of canopy cover, which can be defined as the proportion of the forest floor covered by the vertical projection of the tree crowns (Jennings et al, 1999). Several studies have introduced methods of estimating canopy cover from LiDAR, including Nelson et al (1984), where it was estimated by dividing the number of canopy hits by the total number of hits altogether; and Ritchie et al (1992), where it was estimated by counting the number of laser measurements in a height category and dividing by the total number of laser measurements for a given segment of the profile. Canopy cover has also been considered synonymous, although wrongly so, by several authors with canopy closure (i.e. Philip, 1994; and Avery and Burkhart, 1994). The most important difference between these two measurements is that canopy closure measurements integrate information over the sky hemisphere above one point on the ground, whereas canopy cover assesses the presence of canopy vertically above a sample of points (Jennings et al, 1999). This difference is important because while field-based methods such as hemispherical photography measure canopy closure, a profiling LiDAR system measures canopy cover, and thus any comparison between the two would be the result of the comparison between two distinct variables. This research was designed to address three specific objectives: (i) determining the best strategy for separating canopy from ground in discrete-return LiDAR profiles, (ii) determining how to select the threshold between canopy and ground; and (iii) determining if stratification is necessary in order to enable the algorithm to operate efficiently across diverse forest types. Methods The study area for this research is located in the west-central portion of Alberta, Canada (Figure 14). We acquired LiDAR data along a 1200km transect over a period of two days: August 19-20, 2006 The flight path was designed to sample a broad range of diversity while maximizing forest coverage. The forests that lie underneath the transect range from pure

48


hardwood to pure softwood, with a variety of mixed stands in between. Apart from a small amount of obvious burns and regenerating forest plots, the forest stands under coverage were generally healthy and displayed a range of canopy closure from approximately 5% to approximately 70%. There were, however, areas in the extreme west of the study area that may have been affected by mountain pine beetle.

Figure 14: Location of study area and LiDAR transect in west-central Alberta, Canada. The LiDAR transect was designed to sample the diversity of natural subregions in the area, while maximizing forest coverage. Distance ranges from LiDAR (in reference to sea level), were acquired along the transect using a Riegl LMS-Q140i-80 sensor on board a fixed-wing aircraft that was flown by Laser Imaging Technologies of Calgary, Alberta. The LMS-Q140i-80 sensor is a discrete-return device that operates through the use of a rotating polygon mirror to scan the target surface in a parallel manner, up to a maximum scan angle range of 80 degrees. The pulse is considered a class 1 laser (human eye safe) that is emitted at 10 kHz at a near infrared wavelength, and has a beam divergence of approximately 3 mrad. Although accuracy changes over the extent

49


of the transect, average accuracies are approximately 0.66m and 0.04m in the horizontal and vertical planes, respectively and the density of points approaches 2.5 returns per m2. Although the system is capable of recording multiple returns per pulse, only the first and last returns were recorded, and only the first returns will be used in this analysis. To obtain a profile from this data, returns corresponding to an incidence angle of less than one degree were selected.

Canopy closure estimates from hemispherical photography were acquired during the same time period by field crews on the ground. The field protocol involved measurements of vegetation composition (species, percent cover) and structure (height, and volume) using standard vegetation sampling and timber cruise methods across a 30-metre plot analogous in size to a Landsat Thematic Mapper (TM) pixel. Upon arrival at a given field site, five hemispherical photographs were taken at fixed locations within the plot. These five digital photographs were then analyzed in WinSCANOPY, where gap fraction was used to calculate canopy closure. The resultant canopy closure estimates were then averaged to determine a single measurement for each 30 x 30 m plot.

All data analysis was performed using software developed specifically for this research using IDL (Research Systems, Boulder CO). Thirty-metre sample profile segments covering the ground plots were extracted from a file containing all points with incidence angle values of less than one degree. Ground surface elevation was defined for each 30 m segment as the minimum elevation along that segment, whereby anything above that minimum value was considered to be vegetation. Elevation values were then scaled based on this minimum value along with the minimum horizontal value, providing a profile representative of the local area.

Canopy cover was determined using (i) a histogram method described by Ritchie et al (1992), (ii) a point count method outlined by Nelson et al (1984), and an innovative line segment method developed for this research. Both the point count method and the line segment method use the same basic principle of defining a height threshold whereby anything below that threshold is considered ground and anything above that threshold is considered vegetation. The difference between the two however, is that the line segment method connects the dots between the points and sums the line segments above and below the threshold, as opposed to simply counting the number of returns above and below the threshold.

Figure 15 provides an illustration of how canopy cover is estimated using the line segment method developed in this project. If the location of two points – (p1, located above a crown-defining threshold, and p2, located below the threshold) are known, then the Euclidean distances marked a and b can be easily calculated through subtraction, while distances c and d can be calculated with basic trigonometry. Once estimated, the algorithm adds distance c is added to the vegetation segment and distance d is added to the ground segment. This same procedure is repeated for every pair of points until the entire profile is processed. The result is a ratio of vegetation segments to ground segments over a given distance, and can be considered a LiDAR-derived horizontal projection of the vertical structure of the canopy. Although this LiDAR-derived variable is better defined as canopy cover, it is believed to be closely associated with the measurement of gap fraction made through the use of hemispherical photography and is therefore considered a close estimate of canopy closure.

50


Figure 15: Illustration of how the line segment and point count methods for estimating canopy closure are performed on a forest profile generated from LiDAR. The horizontal projection of the vertical structure is estimated by calculating the proportion of vegetation segments (marked c) to ground segments (marked d). Each LiDAR-derived canopy cover estimation method was performed at various height thresholds on 80% of the 30 m profile segments that corresponded to a field site location (20% of the data was held back for validation). This empirical relations process was performed to see which method worked the best when compared to field data that measured canopy closure (objective 1), and at what height threshold this occurred (objective 2). Absolute height thresholds ranging from 0.1m to the maximum height of the canopy were compared, as were proportional height thresholds ranging from 1% to 100% of maximum canopy height.

51


Once canopy cover values were determined for each height threshold at each sample location, they were regressed with field measurements from each location using a simple linear regression. Each relationship was analyzed for goodness of fit, and testing of assumptions associated with linear regression was also performed. This was done with the data both stratified and un-stratified according to species and whether or not the area was considered upland or wetland, so as to assess whether or not species composition or moisture had an effect on the relationship between estimates of canopy cover from LiDAR and canopy closure from field measurements (objective 3). The best model from each method was tested against the others through use of the significance of the difference between two correlation coefficients for correlated samples procedure, as outlined by Ferguson (1981). Results Line plots portraying the r2 values from the histogram, point count, and newly-established line segment methods for both absolute and proportional thresholds are shown in Figure 16, with individual lines reflecting results obtained for stratified and unstratified data. The innovative line segment method performed better than both the point count method and the histogram method for this study area. This was determined by locating the highest peak (or set of peaks from a stratification) in the r2 distribution for each method, and directly comparing that with the peaks from the other methods. The peak r2 value for the line segment method using the unstratified iteration was 0.81 at an absolute threshold of 2.7m, and although the softwood strata has a higher peak r2 value of 0.89, the hardwood and mixed strata only have r2 values at that threshold of 0.30 and 0.71, respectively. The peak r2 values for the point count and histogram methods also occurred in the case of the unstratified global model, however their values were both 0.71 at an absolute threshold of 0.7m. In the test of statistical significance between the models, the innovative line segments method was significantly better than best model from both of the other methods. This verifies that at a threshold of 2.7 m, explaining 81% of the variability in canopy closure measurements and having a standard error of 10.32 on 53 degrees of freedom, that the line segments method for estimating canopy cover is the best method overall of the ones compared in this study. A plot of observed versus predicted canopy closure values for the 14 independent test plots produced an r2of 0.87, with an RMSE of just over 5m. Discussion Despite the conceptual difference between canopy closure and canopy cover, it has been established here that the two forest canopy descriptors share a statistically significant linear relationship. This indicates that a profiling LiDAR system can be successfully used to estimate canopy closure, despite the fact that it physically measures canopy cover. This is an important result since canopy closure is directly related to the microclimate and light regime at a given area, thus providing a link to growth and plant survival. Therefore, the efficiency, repeatability and accessibility benefits that LiDAR provides can be utilized to estimate an important descriptor of the forest that otherwise might not be possible over large and diverse areas through traditional field methods. In addition, the cost of obtaining these canopy closure estimates will be greatly reduced.

52


While the cost of acquiring a summer’s worth of field data and the LiDAR data for this study area were roughly comparable, the amount of data acquired at the end of the LiDAR mission far exceeded that of the field mission. This suggests that in a relative comparison between the two methods of data volume per dollar spent, the acquisition of LiDAR out-performs the acquisition of field data for the purposes of determining canopy closure. Also, when one takes into consideration the subjective concept of risk, the cost of obtaining field data is a great deal higher than that of obtaining LiDAR data. This has to do mostly with the fact that a great deal more time is spent acquiring field data, thus increasing the chance of complication, but also with the fact that the acquisition of field data is performed almost completely by people instead of equipment, as is the case with LiDAR. Although damage to expensive equipment and physical resources can pose a serious problem, damage to human resources is a much more difficult situation to deal with. Conclusions In an attempt to isolate the best method possible for separating canopy from ground, it has been discovered through an empirical relations investigation that an innovative line segment method introduced in this study provides the best and most consistent results. It has also been discovered that this method produces the best relationship between LiDAR and field measurements at a threshold of 2.7 m above the lowest elevation point in the sample profile (which can be considered ground level) and that stratification is not necessary in order to enable the line segment method to work over diverse forest types. This occurs despite the conceptual difference between canopy closure (which is measured in situ through the use of hemispherical photography) and canopy cover (which is measured remotely by LiDAR), and is important to researchers because it creates a link between the growth and plant survival characteristics of canopy closure with the efficiency, repeatability and accessibility benefits of LiDAR. This work is fully documented in McLane (2007), and is the basis for the following forthcoming journal article: McLane, A.J., G.J. McDermid, and M.A. Wulder: Processing discrete-return profiling LiDAR data to estimate canopy closure for large-area forestry mapping. International Journal of Remote Sensing, in prep.

53

Chapter 2: Remote Sensing Mapping and Research A

C

B

D

E

Figure 16: Plots of r2 values derived from regression analyses performed between ground- and LiDAR-estimated values of canopy closure using the line segmentation method with absolute (A) and relative (B) thresholds, the point cound method with absolute (C) and relative (D) thresholds, and the histogram method with absolute (E) thresholds.

54


NOISE REDUCTION OF NDVI TIME SERIES FOR CHARACTERIZING VEGETATION PHENOLOGY

Introduction The study of recurrent vegetative, biophysical events such as springtime budburst, leaf-out, flowering and autumnal senescence is referred to as vegetation phenology (Badeck et al 2004). Agricultural interests in phenology date back thousands of years, but more recent interests in the topic have expanded considerably to include those of biologists, ecologists, climatologists and other researchers (Zhang et al 2004). Phenology is a driving force in the lifecycles of innumerable organisms, and plays a critical role in the timing and abundance of food supplies for insects, birds, rodents, ungulates, and many other organisms. Grizzly bear activity and movement are highly dependent upon seasonal vegetation: the spring, summer and autumn shifts in the distribution of high energy food sources influence grizzly presence over a variety of landscapes (Hobson 2005, Munro et al 2005). Time series of satellite-derived spectral vegetation indices (VIs) such as the normalized difference vegetation index (NDVI) are fundamental to the remote sensing of vegetation phenology, but their application is hindered by prevalent noise resulting chiefly from varying atmospheric conditions and sun-sensor-surface viewing geometries. Guided by a series of informal research questions, this work was designed to address the following four objectives:

1. Performing a comprehensive review of the literature concerning the acquisition and processing of NDVI time series, including noise reduction techniques, and the subsequent extraction of phenological variables. It was asked: what techniques are available? Were and when are they used? What is their success in the literature? What selection of techniques would be most appropriate for an empirical comparison in the current research?

2. Determining which of the selected noise reduction techniques is most effectively applied to the present data set and study area. It was asked: is one technique superior to all other selected techniques? If so, which one? If not, why?

3. Exploring factors that influence the effective application of the selected noise reduction techniques. It was asked: do land cover; year; the choice of phenology-based NDVI metric; and level, strength and type of noise influence the performance of the selected techniques? Why or why not?

4. Examining whether it is always best to apply NDVI time series noise reduction. It was asked: is noise reduction always beneficial? When is it, or is it not beneficial? Do land cover, noise level, year and choice of NDVI metric influence the beneficial nature of noise reduction?

Remote Sensing of Vegetation Phenology The remote sensing of vegetation phenology is a diverse field of research, comprising a multitude of disciplines, research objectives and approaches to studying the health and development of surface vegetation as reflected in the occurrence and timing of periodic biophysical events. The importance of vegetation phenology to a variety of biological, ecological and climatological concerns is evident in the literature, as is the need for repeatable regional- to global-level data sets in support of large-area studies of earth system

55


dynamics and processes, as is provided by a variety of satellite-borne optical sensors. Data from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors on board the Terra and Aqua satellites are becoming increasingly popular for phenological research. While MODIS provides data at spatial and temporal resolutions similar to that of previous-generation AVHRR imagery, it also provides improved radiometric correction, as well as a multitude of additional vegetation related products. The MODIS VI data sets are especially interesting to those who study vegetation phenology as these are shown to be closely related to levels of photosynthetic activity occurring at the surface. Of the two VIs, the NDVI is particularly prevalent throughout the literature. The simplicity of its calculation, its normalizing nature, and its availability as pre-processed data sets, such as those provided by MODIS, have rendered it very popular for phenological research. A MODIS NDVI vegetation product provides the remotely-sensed data set for the current research. The free and easy access to MODIS data sets through the Land Processes Distributed Active Archive Center and the superior radiometric quality of these data to the more common AVHRR data sets led to the selection of MODIS VI data. In addition, the continued popularity of the NDVI rendered it the most practical choice for providing the most substantial contribution to present-day research and literature regarding the remote sensing of vegetation phenology. Nevertheless, the inherent noise and error resulting from high-frequency variations in atmospheric conditions and sun-sensor-surface viewing geometries that is present in the time series of NDVI so widely applied in phenological research are an important concern. It is the presence of noise in multi-temporal NDVI data sets, and the importance of its minimization to subsequent efficient use in studying vegetation phenology, which provides the impetus for the current research. Numerous researchers have addressed this problem, providing a variety of noise reduction filtering and function-fitting techniques that work on a per-pixel basis for the minimization of undesirable high-frequency noise. However, no standard methods or approaches prevail in the literature and few comprehensive studies testing the relative success of many of these techniques are found. While it is obvious that a number of such techniques demonstrate considerable potential for noise reduction in NDVI time series, rarely are these techniques applied elsewhere beyond their initial proposals. The present project aims to provide a more comprehensive comparison of several of the noise reduction techniques presented above, and thus relies heavily on this portion of the present literature review. Additional information is provided regarding more peripheral aspects of the remote sensing of vegetation phenology, particularly in relation to the current research, but these portions of the literature review are relevant to a comprehensive understanding of the topic and of the methods used to undertake the present empirical comparison. The final element of the present review concerns the extraction of phenological information in the form of metrics from multi-temporal remotely-sensed NDVI data sets (Table 12). Although such metrics do not form the main component of the present research, knowledge of their importance and application in the remote sensing of vegetation phenology is pertinent to this research. They comprise an important aspect of the empirical comparison described in the following chapter, and therefore require consideration in the above literature review. In particular, the extraction of start, end and length of growing season metrics are

56


Table 12: NDVI time series metrics, derived for the characterization of surface vegetation phenology.

Metric1 Description Methods of Estimation

Start of Growing Season (SOS)

Timing and level of photosynthesis at the start of measurable photosynthesis, representing the start of the growing season

Thresholds, inflection points, curve derivatives, or Fourier-based methods

End of Growing Season (EOS)

Timing and level of photosynthesis at the cessation of measurable photosynthesis, representing the end of the growing season

Thresholds, inflection points, curve derivatives, or Fourier-based methods

Length of Growing Season (LGS)

The duration of measurable photosynthesis, representing the length of the growing season

Number of composite periods or days between SOS and EOS

Mid-Season Timing and level of photosynthesis at the middle of the growing season

Point midway between time at which 90% of the seasonal amplitude is reached and time at which 10% of senescence has occurred

Integreted NDVI (I-NDVI)

Overall productivity and biomass produced during the growing season

Area under the NDVI curve between SOS and EOS

Rate of Green-Up Speed at which spring green-up occurs Slope of line between SOS and maximum NDVI

Green-Up Fraction Portion of growing season spent in green-up Number of days/composite periods between SOS and maximum NDVI over the total LGS

Green-Up I-NDVI Portion of total growing season productivity produced during green-up

Area under the NDVI curve between SOS and maximum NDVI, over I-NDVI

Start of Senescence Timing and level of photosynthesis at the start of rapid decrease in photosynthesis

Inflection point method

Rate of Senescence Speed at which autumnal senescence occurs Slope of line between maximum NDVI and EOS

Senescence I-NDVI Portion of total growing season productivity produced during senescence

Area under the NDVI curve between maximum NDVI and EOS, over I-NDVI

Maximum NDVI Timing and level of photosynthesis at maximum photosynthesis reached during the growing season

Maximum NDVI between SOS and EOS

Minimum NDVI Timing and level of photosynthesis at minimum photosynthesis reached during the growing season

Minimum NDVI between SOS and EOS

Mean NDVI Overall mean level of photosynthesis occurring throughout the year

Mean of NDVI values over a year

NDVI Amplitude Range in levels of photosynthesis occurring during the growing season

Range between maximum and minimum NDVI

Modality Number of growing seasons within a year Periodicity of significant peaks in NDVI time series

Relative NDVI Range A standardized NDVI amplitude, used in inter-annual comparions of biomass productivity

NDVI Amplitude divided by I-NDVI

Harmonic Amplitude Similar to NDVI Amplitude Wave amplitude of harmonics derived from Fourier transforms

Harmonic Phase Timing of maximum photosynthesis Wave phase of harmonics derived from Fourier transforms

1For more details on these, see Reed et al. (1994), DeFries et al. (1995), Azzali and Menenti (2000), Jönsson and Eklundh (2002, 2004), and Pettorelli et al. (2005).

57


important phenological measures, playing integral roles in Earth-atmosphere energy and mass exchanges and climatological studies because of their responsiveness to local and regional environmental conditions. Deriving these metrics from multi-temporal NDVI data sets involves per-pixel time series calculations, but is done using a variety of methods including thresholding and inflection points. Such metrics continue to support numerous biological, ecological and climatological research objectives and provide a method for studying, measuring and monitoring short- to long-term variations in vegetation phenology at regional to global scales.

Methods The study area for this research covers a portion of west-central Alberta, Canada, along the front ranges of the Rocky Mountains, and covers approximately 71,500 square kilometres (Figure 17). The study area represents the portion of the current FRIGBP study area that is covered by one MODIS tile, and represents the greatest diversity in landscapes of the three tiles covering the area. By employing this study area the results, conclusions and contributions made by the present research can be more easily applied to the current remote sensing-based mapping and modeling efforts of the FRIGBP.

Figure 17: Map showing the location, context and extent of the study area within west-central Alberta, Canada.

58


Figure 18 presents a flowchart outlining the methods and the principal steps involved to complete an empirical comparison of selected strategies for the noise reduction of NDVI time series. The analysis was undertaken within a model framework, wherein the application of selected noise reduction techniques was tested on a series of model NDVI time series with introduced noise. The decision to employ a model environment was based on a consideration of the difficulties involved in acquiring and applying the ground truth data that would have been necessary for a real-world analysis of the selected noise reduction techniques.

Overall RMSE and Metric Scenario

Tallies

Stratified RMSE and Metric Scenario

Tallies

Overall Standardized RMSE

and Metric Performance Scores

Stratified Standardized RMSE

and Metric Performance Scores

Time Series Metric Results

RMSE Results

Noisy NDVI Time Series

Model Construction

MODIS 250m 16-Day Composite

NDVI, 2003-2005

Model NDVI Time Series

Candidate Noise Reduction Techniques

Noise-Reduced NDVI Time Series

Introduction of Noise

Application of Noise Reduction

RMSE Calculations Time Series Metric Calculations

Performance Score Calculations

Performance Score Calculations

Raw RMSE Performance Scores

Raw Metric Performance Scores

Summations of RMSE and Metric

Scenarios

Standardization of RMSE and Metric

Performance Scores

= Input

= Processing Step

= Intermediate Output

= Final Output

Figure 18: Flowchart of the principal steps involved in the empirical comparison of the selected candidate noise reduction techniques.

59


A set of model NDVI time series reflecting ideal conditions and constructed from a MODIS NDVI data set covering the study area for 2003 through 2005 were created. Several levels of noise were introduced to these model time series, affording a variety of modeled noisy NDVI time series on which the six candidate noise reduction techniques, selected from the literature, were tested. It was the goal of the present research to test the quality of these noise reduction techniques in terms of both i) their ability to return noisy NDVI measurements to their ‘true’ values and ii) their ability to maintain the integrity of the original NDVI signal for the subsequent extraction of phenological metrics. The empirical comparison of the six techniques therefore comprised two separate evaluations: one involving root mean square error (RMSE) and the other, phenology-based metrics derived from NDVI time series. Each of these evaluations provides a different measure of noise reduction technique performance, and each is equally important to a comprehensive understanding of the abilities of the selected techniques to minimize noise in NDVI time series and to maintain the integrity of the original signal. The results of both the RMSE and metric evaluations were transformed into performance scores, which allowed for the direct comparison between different types and categories of results that would not have been possible otherwise. These performance scores were further used in scenario summations, which comprised an assessment of how often and in what circumstances the application of a noise reduction technique to noisy NDVI time series produced a better result than when such a technique was not applied. Overall and stratified results were produced from this analysis, the latter representing the re-organization of these results into land cover, noise level and other categories in order to evaluate the effects of this stratification on the results and their reflection on the performance of the selected techniques. In addition to this, further analysis of RMSE and metric performance scores involved the generation of standardized performance scores, from which summary performance scores could be calculated. The latter allowed for the subsequent overall comparison and ranking of the six candidate noise reduction techniques and their respective performances, while stratification of these results provided for additional investigation into the effects of land cover, noise level and other factors on technique performance. Results Overall standardized summary performance score results (Figure 19) indicated that the two function-fitting techniques, the Double Logistic and Asymmetric Gaussian, performed the best of the six candidate noise reduction techniques, while the raw, unfiltered data produced the worst overall performance. The RMSE and metric components of these summary scores differed, however; the Savitzky-Golay filter generated the best overall RMSE results, with the two function-fitting methods, and the 4253H-Twice filter produced the best overall metric results.

60


1.78 1.79 1.86 1.932.06 2.10

2.48

0.0

0.5

1.0

1.5

2.0

2.5

3.0

D.L. A.G. 4253H-Twice S.G. ARMD3-ARMA5

MVI None

Noise Reduction Technique

Stan

dard

ized

Per

form

ance

Sco

re

Figure 19: Standardized summary performance scores for the seven candidate noise reduction strategies, ordered from lowest (left) to highest (right). ‘None’ refers to the absence of noise reduction. Stratification of the results revealed that standardized summary, RMSE and metric performance scores varied considerably with land cover. The consistently poor performance by the unfiltered, noisy data across the majority of land covers was the one observable trend. Similarly, no one technique performed the best across all three levels of introduced noise according to the stratified results. Nevertheless the standardized summary, RMSE and metric performance scores showed that not applying noise reduction produced the worst performance scores at the 40% and 70% noise level, while at the 10% noise level the Savitzky-Golay filter performed poorest. An unexpected result was also encountered in these stratified performance scores. It was observed that the unfiltered data with 40% noise generated worse standardized performance scores than the unfiltered data with 70% noise. The opposite would have been expected. Stratification by year and metric was necessarily restricted to standardized metric performance scores; RMSE and summary results were not included in this portion of the analysis. According to metric performance scores stratified by year, more variation in scores occurred between the candidate noise reduction techniques for each year, than between the three years for each technique. In addition, the Double Logistic and Asymmetric Gaussian function-fitting methods, and the 4253H-Twice filter consistently produced better standardized metric performance scores than the MVI, ARMD3-ARMA5 and Savitzky-Golay filters across the years 2003 through 2005. The stratification by metric of standardized metric performance scores revealed more variation in results between metrics

61


than was observed for land cover, noise level or year. Indeed, the general superiority of the Double Logistic and Asymmetric Gaussian function-fitting methods over the MVI and ARMD3-ARMA5 filters across the majority of metrics was the only observable pattern in these results. Overall scenario tallies from the scenario analysis showed that in the majority of all cases, applying a noise reduction technique did not improve on the raw performance scores produced from raw, noisy data (Figure 20). This was also demonstrated by the overall metric scenario tallies, but the overall RMSE results showed a contrary pattern. This dichotomy between RMSE and metric results was generally consistent. In the majority of RMSE scenarios, applying noise reduction produced an improvement in performance scores. RMSE and metric scenario tallies varied little when stratified by noise reduction technique, but RMSE scenario results varied considerably with land cover. Metric scenario results showed more variation with land cover than with noise reduction technique, but not to the same degree as that demonstrated by the RMSE scenarios. Stratification by noise level revealed that noise reduction is consistently less beneficial at the 10% noise level than at the 40% and 70% noise levels, according to both RMSE and metric scenario tallies. Stratification of metric scenario results by year produced little variation in the results; tallies with the majority of scenarios demonstrating no improvement in performance scores with noise reduction were consistently calculated for all three years. The final stratification, that of the metric scenario results by metric, revealed considerable variation. No discernable trends were found in these results, although three of the metrics generated tallies contrary to all other metric scenario results. More than half of the NDVI amplitude, maximum rate of green-up and average NDVI scenarios demonstrated improved performance scores with the application of noise reduction.

1120, 33%

2228, 67%

Noise-ReducedNoisy

Figure 20: Pie chart showing the number and percentage of scenarios in which either a noise-reduced time series (light grey) or the noisy, unfiltered times series (dark grey) produced a better unstandardized performance score.

62


Discussion The current empirical comparison showed that over the broad range of scenarios investigated in this experiment, both of the function-fitting techniques, the Double Logistic and Asymmetric Gaussian functions, demonstrated a superior ability over the four filtering methods for minimizing noise in NDVI time series and maintaining the integrity of the original, desired signal. This superior performance is likely the result of i) a better capability for approximating constant winter NDVI, which reflects dormant vegetative conditions, and ii) a strict preservation of the upper envelope of NDVI values. Not only were they better able to deal with noise that was primarily negatively biased, but subsequent metric calculations generally tended to be more accurate because of better approximations of winter NDVI, and beginning and end of season transitions. These findings support those of Jönsson and Eklundh (2002, 2004) and Beck et al (2006, 2007), who demonstrated the flexibility of these methods when confronted with irregular, asymmetrical NDVI time series, particularly in comparison with more common Fourier-based strategies. They also found these techniques to be less vulnerable to high levels of noise than other techniques, such as the Savitzky-Golay filter and BISE method. Both the MVI and ARMD3-ARMA5 filters showed consistently poor performance in the current empirical comparison despite successful applications asserted in the literature. Both lack the ability to maintain the upper envelope of NDVI values, and rely heavily on averaging techniques. The former factor results in a greater vulnerability to negatively-biased noise; the latter, a tendency to distort the shape and amplitude of NDVI time series signals. Both the 4253H-Twice and Savitzky-Golay filters demonstrated overall better performance than the MVI and ARMD3-ARMA5 filters, but varied in their performances according to RMSE and metric results. The Savitzky-Golay filter provided the best general noise minimization but was less capable of maintaining the integrity of original NDVI signals for subsequent metric extraction. The 4253H-Twice filter, however, was more successful at metric extraction but less so at general noise minimization. The ability of the former filter to maintain the upper envelope of NDVI values and the lack of such a bias in the latter is a likely explanation for the better general noise minimization observed in the former. The greater capability of the 4253H-Twice filter for deriving accurate time series metrics could doubtless be explained by the reliance of the 4253H-Twice filter on running medians and the potential for user error in selecting moving window widths for the Savitzky-Golay filter. The current research demonstrated notable variation in performance of the candidate noise reduction techniques with land cover type. This variation was interpreted not to be caused by differing phenologies across land covers as one might expect, but rather, differences in the strength and type of noise. In particular, the presence of negatively-biased noise led to underestimation of original NDVI signals by techniques which did not aim to preserve the upper envelope of NDVI values, while techniques incorporating this aim over-estimated time series for which noise was not negatively-biased. Observations of the results stratified by noise level revealed that any of the six candidate noise reduction techniques were at least somewhat successful at the higher noise levels (i.e. 40% and 70% introduced noise), while at the 10% noise level only the 4253H-Twice filter produced a better standardized performance

63


score than no noise reduction. The application of noise reduction to time series with minimal levels of noise was found to degrade these relatively clean signals, while at moderate to high levels of noise the application of such strategies appeared advantageous. The performance of the six candidate noise reduction strategies, though varying with land cover and strength, type and level of noise, appeared to be most affected by one’s choice of metric. Indeed, the variation in noise reduction technique performance with metric was so great that few broad trends could be observed. Nonetheless, it can be argued that the application of noise reduction techniques to NDVI time series generally resulted in more accurate derivations of time-based metrics; four of the five metrics directly related to the timing of phenological processes demonstrated the worst standardized summary performance scores where no noise reduction had been applied. It was demonstrated that this likely results from an increased difficulty in deriving such time-based metrics from noisy time series containing a number of spurious spikes or drops, which could be mistakenly identified as SOS, EOS or the timing of maximum green-up. Because the application of any one of the six candidate noise reduction techniques minimized the amount of high-frequency noise found in NDVI time series, it therefore follows that the difficulties in deriving time-based metrics from jagged, noisy time series were also minimized with the application of noise reduction. When compared with these standardized performance score results, however, the results of the scenario analysis appeared to be contradictory. While the former demonstrated a notable support of the candidate noise reduction techniques as a superior choice to the use of uncorrected, noisy data, the latter suggested something quite different. Overall results of the scenario analysis indicated that in nearly 70% of individual RMSE and metric scenarios, the noisy, unfiltered NDVI time series performed equally to or better than a noise-reduced time series. This disparity can be attributed to differences between the two methods of analysis. While the scenario analysis represents the performance of the candidate noise reduction techniques relative to noisy data over individual scenarios, the standardized performance scores are total measures of overall inaccuracy. Thus, while noise reduction may not have produced benefit over the majority of all scenarios, for those cases where it was beneficial the alternative choice of not applying noise reduction caused considerably worse errors than applying noise reduction when it was not beneficial. The same was true of all metric scenarios, when examined separately, but RMSE scenarios showed an opposing trend, demonstrating better performance by a noise reduction technique in more than half of RMSE scenarios. Thus, while the application of noise reduction was frequently successful in terms of the general minimization of noise, it generally led to less accurate metric derivations because the NDVI signal could be considerably altered with the application of such techniques. Stratification revealed little variation in these results with either the selection of noise reduction technique, noise level or year; none of these factors demonstrated a strong effect on the probable benefit of applying noise reduction. Stratification by land cover generated considerable variation in scenario analysis results however, particularly in the case of RMSE scenarios where half to 100% of scenarios demonstrated improved performance with the

64


application of noise reduction across the six land cover types. Nonetheless, the small number of scenarios on which these numbers were based necessitates the acquisition of larger sample sizes before the accuracy of these observations could be confirmed. The factor appearing to most influence the potential benefit of noise reduction for NDVI time series is one’s choice of metric. Scenario results varied considerably with metric, and are not likely to be exaggerated because metric sample sizes are much larger than for RMSE results. One of the few patterns discernable from these stratified metric scenario results is the contradictory findings observed for NDVI amplitude, average NDVI and maximum green-up. All metric scenario results indicated an equal or improved performance by unfiltered, noisy NDVI time series than by noise-reduced time series, with the exception of these three metrics. That is, for each of these metrics applying noise reduction is beneficial in the majority of scenarios. These contrary results can likely be attributed to the positive effect of minimizing high-frequency fluctuations in noisy data sets on the derivation of these metrics; by reducing these fluctuations, these particular metrics can be more accurately derived. Conclusions Satellite-derived NDVI time series are fundamental to the remote sensing of vegetation phenology, but their application is hindered by prevalent noise resulting chiefly from varying atmospheric conditions and sun-sensor-surface viewing geometries. A model-based empirical comparison of six selected NDVI time series noise reduction techniques revealed the general superiority of the Double Logistic and Asymmetric Gaussian function-fitting methods over four alternative filtering techniques. However, further analysis demonstrated a strong influence by noise level, strength and bias, and the extraction of phenological variables on technique performance. Users are therefore strongly cautioned to consider both their ultimate objectives and the nature of the noise present in an NDVI data set when selecting an approach to noise reduction, particularly when the derivation of phenological variables is a final goal. This work is crucial to improving current understandings of NDVI time series noise reduction and its role in the remote sensing of vegetation phenology. This work is fully documented in Hird (2008), and is the basis for the following forthcoming journal articles: Hird, J.N. and G.J. McDermid: Review of methods for remote sensing of vegetation phenology, part 1: sensors and products. Progress in Physical Geography, in prep. Hird, J.N. and G.J. McDermid: Review of methods for remote sensing of vegetation phenology, part 2: metrics and noise reduction. Progress in Physical Geography, in prep. Hird, J.N. and G.J. McDermid: Noise reduction for time series of the normalized difference vegetation index. Remote Sensing of Environment, in prep.

65


CLASSIFICATION OF AGRICULTURAL AREAS Agriculture and its associated activities is a major cause of increased conflict between bears and humans, and a decline in bear populations. Kansas (2002) identified reducing human-grizzly conflict on agricultural lands as a priority for mitigating the long term decline of the species. In a study of grizzly-human conflict on agricultural lands in Montana, Wilson et al (2005; 2006) found that there were many different attractants for bears on private lands that are a part of the natural bear habitat. One of the most important factors was the use of riparian areas by bears as both habitat and transportation corridors (Wilson et al, 2005). The bears use these areas to reach anthropogenic attractants, such as cattle, sheep, beehives, and boneyards. The more attractants that were in an area, and the closer that area was to wetlands or riparian areas, the more likely the bears were to use that area as habitat. Alberta has a large agricultural footprint. Agriculture and related activities exist right up to the edge of the foothills of the Rocky Mountains. However, one problem currently facing grizzly bear habitat mapping in Alberta is the lack of a classification scheme that differentiates between different agricultural and herbaceous areas. By finding an appropriate classification scheme for this purpose, the current land cover maps being used by the FRIGBP for grizzly bear habitat analysis will be updated with greater thematic resolution, which could lead to increased resource modeling accuracy. The purpose of this research is to demonstrate the use of remote sensing for land cover classification in western Alberta, specifically focusing on the classification of herbaceous and agricultural areas in grizzly bear habitat. The specific goals of this project are: (i) to find the best possible classification approach from a limited selection of methods for determining multiple classes of agricultural and herbaceous land cover, and (ii) to create land cover maps of agricultural and herbaceous areas which will be integrated into existing grizzly bear habitat maps for western Alberta. Methods The study area for this project covers sections within the greater FRIGBP study area that contain herbaceous and agricultural areas, and that are within the natural range of the grizzly bear (Figure 21). Two areas were examined in detail: one in the northern part of the province, located west of Grand Prairie (the North study area), and one in the south, located around the Nanton/Chain Lakes area (the South study area). The two study areas were selected from agricultural areas that are within the current range of grizzly bears in the province, and that have bear GPS collar location data present within them. Large portions of both of these study areas were also located within Landsat scene overlaps, which made cloud-free image acquisition more likely.

66


Figure 21: Map of Alberta showing collared grizzly GPS locations and the North and South study areas. One scene was collected for each of the North and South Study areas. The North study area was covered by Landsat scene 47/21 acquired on July 26, 2007; the South study area was covered by Landsat scene 42/25, acquired on July 23, 2007. A stratified random sample scheme was used to collect field level data in late July, 2007, which corresponds to the week in which the images that cover the North and South study areas were taken. A total of 506 ground samples were collected on five agricultural classes : Bare Soil/Fallow, Canola, Grass/Forage, Legumes, and Small Grains (which includes barley, wheat, and oat varieties). 70% of these samples were used for training data; the remaining 30% saved for validation. The Landsat scenes (both TM and ETM+) were orthorectified using 5th order polynomial geometric correction in PCI OrthoEngine. Ground control points (GCPs) were collected from

67


existing geo-referenced scenes of the same areas; a minimum of 30 GCPs were used for each image. Root Mean Square (RMS) error for all images was lower than 0.2 pixels. Radiometric and atmospheric correction was performed using the ATCOR-2 algorithm in PCI Geomatica 10. While most traditional remote sensing land cover classification is pixel-based, many newer studies are turning to object-based classification methods as a way to improve. Object-based classification divides the satellite image into objects or segments that represent a homogenous unit on the ground. The entire object is classified based on the overall statistical properties of the pixels that make up the object, instead of classifying each pixel separately as in pixel-based classifications. Three different object-based classifications were performed and analyzed; one unsupervised classification, and two supervised classifications. Validation of the results was done using both the standard error matrix and the Kappa Index of Agreement (KIA) for both overall and class specific results. The most accurate and useful classification method from among those tested was applied to the agricultural scenes across the remaining portion of the Phase 6 study area. The complete classification of the eight Landsat scenes was then added to the land cover map, with the new classification being overlain on top of the existing classification as one large image mosaic. Results and Discussion The supervised classifications gave higher accuracy results than the unsupervised classification (Figure 22). The supervised NN classification had an overall average accuracy of 85.7%, with an average KIA of 80.1%. The accuracy of the north scene was again higher than that of the south, with an overall accuracy of 86.7% and a KIA of 82.4% compared to the south scene’s 84.8% overall accuracy and 77.8% KIA. The supervised sequential masking (SSM) technique gave the highest classification accuracies of the methods tested, with the highest average overall accuracy (88.0%) and KIA (83.4%) values. The accuracy of the south scene was higher than that of the NN method, but the north scene had slightly lower accuracy results than the NN method (Figure 22). Individual class accuracies were also very good, with classes such as Bare Soil / Fallow, Canola, and Peas having average KIA per class values above 95%. The lowest accuracy was the southern Bare Soil / Fallow class user’s accuracy, at 44%.

68


Figure 22: Overall accuracy and KIA for all three classification methods. Due to the higher average accuracies, as well as other benefits, such as easy adaptation to new areas without training sites, the SSM method was chosen as the best method of classification, and was applied to the other six Landsat scenes (5 TM, 1 ETM+) that make up the agricultural area in the FRIGBP study area. The complete mosaic, with the SSM agricultural classification can be seen in Figure 23. The agricultural classification was added to the existing FRI land cover map, increasing its thematic resolution. There are some class similarities between the SSM and FRI classified parts of the map. For example, the SSM Bare Soil / Fallow class is spectrally similar to the Barren class of the FRI map, though the SSM class represents a different use of the land cover. Another example is the SSM Grass / Forage class, which is similar spectrally to the Upland Herbs class of the FRI map, though again the use of the land cover is different between these two classes, with the SSM class existing within an agricultural framework. The new SSM land cover map could contribute to more accurate resource selection models and would give a better understanding of bear activity in agricultural areas.

69


Figure 23: Completed mosaic of the Phase 6 study area with SSM classification, showing new agricultural classes (top 5 in legend) with those of the FRI land cover map.

70


Conclusions The objectives of this research were to test a small selection of classification methods, and of those methods, find the one most appropriate for determining multiple classes of agricultural and herbaceous land cover for the purpose of land cover mapping in areas of grizzly bear habitat. The most appropriate method was determined to be the Supervised Sequential Masking classification, which gave an overall accuracy of 88% and a Kappa Index of Agreement (KIA) of 83%. It had the highest classification accuracies, was the most operationally useful, and it is flexible and easily expandable to other classification problems. This work is fully documented in Collingwood (2009).

71


RELATIONSHIPS BETWEEN LANDSCAPE SPATIAL PROPERTIES AND GRIZZLY BEAR PRESENCE IN AGRICULTURAL AREAS

Introduction Activities such as oil and gas exploration and extraction, forestry, agriculture, and recreation all contribute to grizzly bear habitat fragmentation and loss (Garshelis et al, 2005). Another important factor is the network of roads and trails that all of the aforementioned activities depend on, as well as the seismic exploration lines that are cut for oil and gas exploration (Mace et al, 1996; Linke et al, 2005). These linear features allow access to otherwise remote areas by people, which leads to conflict and a declining bear population (Kansas, 2002). Fragmentation not only fragments the landscape, but reduces the total area of available habitat, and may limit grizzly bear movement. Management plans to reduce problem bear conflicts in agricultural areas were mentioned by Kansas (2002) as one of the strategies with the greatest potential to mitigate human-induced harmful effects on grizzly bear populations in Alberta. It has also been recommended by Alberta’s Endangered Species Conservation Committee that the species be elevated from “may be at risk” status to “threatened” status (Stenhouse et al, 2003). Any change in status would require appropriate management and conservation planning, including management plans for agricultural areas that are a part of traditional grizzly bear habitat.

The purpose of this research is to investigate the possible relationships between metrics that represent landscape structure and grizzly bear (Ursus arctos) presence in agricultural areas. The characteristics of certain landscape elements and landscape composition and configuration are examined to identify their relationships with grizzly bear location information. Using satellite imagery, existing bear location GPS data, and a statistical landscape analysis program (FRAGSTATS) this research is designed to determine the configurational and compositional differences between areas that the bears use and areas that they avoid in the agricultural landscape. Information about these relationships between landscape and bear presence could be critical in determining land management practices in agricultural areas that border current grizzly bear habitat. Landscape metrics have been shown to be an important element in grizzly habitat selection (Linke et al, 2005). Therefore, the specific goals of this research were to:

1. identify landscape composition and spatial configuration in the agricultural areas of western Alberta;

2. determine if landscape composition and spatial configuration are related to grizzly presence or absence in an area;

3. determine which landscape metrics have the strongest relationships with certain grizzly population and biological measures that are available from collared bear GPS datasets;

4. determine the extent of the difference between landscape metric values when calculated at different spatial and thematic scales.

Methods The study area for this project was the foothills region to the west and south of Calgary, Alberta (Figure 24). The area was chosen based on grizzly GPS location data that suggested that bears were present in agricultural areas in this part of the province. The landscape of this area is dominated by grassland and agricultural crops, with patches of forest, changing to

72


largely forested areas further west in the foothills. Roads are a dominant feature in much of this landscape, with higher densities in the agricultural areas, and lower densities in the foothills. The total study area covers 4494 km2, which was made up of 107 square sub-landscapes of 42 km2 each, 71 of which contained bear occurrence points. The scale of the sub-landscapes in this research is based on the recommendations of Linke et al (2005), who found that grizzly bears move through and select habitat at a landscape scale of around 35 – 50 km2. Each sub-landscape was analyzed separately in the FRAGSTATS program (McGarigal et al, 2002), and had its own landscape metrics generated.

Figure 24: Study area map showing the distribution of the 107 sub-landscapes in southern Alberta.

73


The imagery used for this research was from the Indian Remote Sensing (IRS) satellites (IRS-1C and IRS-1D) panchromatic sensors. The IRS imagery was acquired as 6 bit image data, resampled to 8 bit by the company Space Imaging (maximum number of distinct grey levels = 64). Each image has been orthorectified to Alberta provincial 1:20,000 vector data files. The imagery has a geometric accuracy of +/- 15 meters across each scene. The images are a compilation of scenes from as many as 7 dates, acquired between April and October, and some span more than one year. The panchromatic IRS images were classified using the Definiens Professional object-based image analysis software package. Each image was classified separately, using the same SSM classifier as described in the previous section. The SSM method was chosen for its relatively high accuracy and so that each image that was classified could simply use a modified version of the SSM classifier that was used on the previous image; the SSM classifier is easily adapted to suit each scene. A limited number of classes were used in this study, due to its focus on agricultural settings and limitations of interpretability for the panchromatic imagery. Four classes were used: agriculture (which includes open shrubland, grassland, and pastureland in addition to agricultural crops) forest, water, and other (which includes features such as roads, cities, bare rock, snow, etc.). A variety of configurational and compositional landscape metrics were chosen for this analysis based on their simplicity and accuracy in measuring different elements of the landscape. Metrics were computed at the landscape level in the FRAGSTATS program; landscape level analysis measures the aggregate properties of the entire landscape mosaic for each sub-landscape (McGarigal et al, 2002). Individual grid cells of the same land cover type were merged to form discrete patches using the 8-cell patch neighbor rule (McGarigal et al, 2002), and the sub-landscape borders were not counted as edges, which are the same parameters used by Linke et al (2005). A total of 16 variables were included in the analysis. Multiple regression analysis was conducted, using a stepwise approach, to see which metrics could be used to predict grizzly bear abundance, and how much of the variation can be explained by the given metrics. Finally, logistic regression based on presence/absence of bears was conducted, using a conditional forward stepwise method. Logistic regression was done to test predictions of the presence or absence of bears in a given area. Results and Discussion The multiple regression analysis indicated that a model that included the metrics Patch Density, the area-weighted mean of the Contiguity Index, and the coefficient of variation of the Euclidean Nearest Neighbor Distance was a likely predictor of grizzly bear density. All of these metrics were very significant (p < 0.01) in the model. The R value for the model was 0.61, with an adjusted R2 value of 0.35, which indicates that about 35% of the variance seen in the grizzly density is explained by these metrics. The multiple regression analysis indicated that a model that included the metrics Patch Density, the area-weighted mean of the Contiguity Index, and the coefficient of variation of the Euclidean Nearest Neighbor Distance was a likely predictor of grizzly bear density. All of these metrics were very significant (p < 0.01) in the model. The R value for the model was

74


0.61, with an adjusted R2 value of 0.35, which indicates that about 35% of the variance seen in the grizzly density is explained by these metrics. The coefficients of the model suggest that bear use of an area increases with increasing patch density, increasing amounts of large, contiguous patches, and increasing variation in the distances between similar patches. The landscape metrics included in the logistic regression model by the conditional forward stepwise regression procedure are the coefficient of variation of the Shape Index, the median of the Contiguity Index, and the mean and area-weighted mean of the Euclidean Nearest Neighbor distance measure. From the coefficients for the logistic model, it would appear that grizzly bear presence is associated with an increase in the variation of the patch Shape Index, a higher median Contiguity Index, a decrease in the mean Euclidean Nearest Neighbor distance between patches of the same class, and an increase in the area-weighted mean Euclidean Nearest Neighbor distance between patches of the same class. Grizzly bear presence was predicted with 87% accuracy, and the overall prediction accuracy, including both presence and absence prediction, was 71%. The prediction accuracy is based on the number of correctly predicted presence or absence values (using the regression equation) for each sub-landscape when compared to the observed values (the GPS locations). While this study did not find a direct link between grizzly bear abundance or presence and the amount of agricultural land present, it did find links with spatial attributes that correspond to reduced agricultural activity and human-caused fragmentation. Size, shape, and position of land cover patches in areas of grizzly habitat had a measureable relationship with the presence/absence and abundance of the bears. There was a link between decreased grizzly bear landscape use and agricultural activity. Nielsen and Boyce (2002) suggested that grizzly bears tend to select habitat that is highly variable, which suggests natural, patchy landscapes, like those to which bear presence was correlated with in this study. Natural, patchy landscapes are different from human- fragmented landscapes, which are characterized by patch isolation, geometric patterns, and increased human presence. Relationships between landscape metrics that were representative of human fragmented landscapes and bears were negative, in that bears were less likely to be present in this type of landscape. It may be important to know for future work which landscape metrics are important for analyzing grizzly habitat, as well as what spatial and thematic resolution these metrics should be calculated at; this research is a step towards these goals. Conclusions Knowledge about grizzly bear selection of habitat in agricultural areas is very limited. While it is known that grizzly bears tend to avoid anthropogenic disturbance, this research presents the first evidence that the physical structure and composition of agricultural areas may play a part in this behavior. There were significant differences among landscapes that grizzly bears did use versus those they did not use. Landscape spatial structure seems to have at least some role in determining whether or not bears will use an area in an agricultural landscape. The results of this research, while not definite, could be helpful in informing other grizzly bear resource selection models. This work is fully documented in Collingwood (2009).

75


REMOTE SENSING OF MOUNTAIN PINE BEETLE SUSCEPTIBILITY: A REVIEW OF REMOTE SENSING OPPORTUNITIES

Introduction Over the past decade an epidemic outbreak of the mountain pine beetle (Dendroctonus ponderosae; MPB) has caused devastating loses to pine forests in western Canada and the north-eastern United States. In 2006 the Government of British Columbia (B.C.) estimated that over eight million hectares of lodgepole pine (Pinus contorta) forest were infested with MPB, with 50% of mature pine expected to be dead by 2008 and 80% dead by 2013 (Government of British Columbia 2006). The Government of Alberta has been monitoring the MPB population in forests since the last epidemic in the late 1980s. Beetle populations remain at endemic levels in the south-western portion of the province but 1992 a population was detected in west-central Alberta. In 2006 a sharp increase in the number of infected forest stands was observed. Thus is due to an expanding Alberta MPB population and long-distance dispersal of MPB from B.C. (ASRD 2006). It is believed the continued eastern spread of MPB could result in infestation of jack pine ecosystems found in the boreal forest across North America. This could result in major ecological and economic impacts for much of Canada. This epidemic has had serious ecologic and economic consequences. Forest ecosystems are a principal component in maintaining the biodiversity of a landscape. Mountain Pine Beetle infestation has been found to have positive initial effects on some wildlife species, but over time those decrease and ecosystems are left with a biodiversity deficit which can take decades to recover. While initial MPB infestations can increase avian populations due to the increased food source of beetle larvae and mature beetles, post-infection bird populations in affected stands are significantly reduced over time (Martin et al 2006). Lynx are associated with the occurrence and distribution of snowshoe hare, which prefer dense, mature conifer forests. Loss of mature pine stands could seriously affect snowshoe hare populations and consequently Lynx populations can also be expected to decline (Koehler and Aubry 1994). Defoliation associated with post-MPB landscapes also results in the loss of both thermal and security cover for numerous ungulate species such as elk (Cervus elaphus), moose (Alces alces), deer (Odocoileus sp.) and woodland caribou (Rangifer tarandus caribou) (Bunnell et al 2004), a species listed as “threatened” in Canada (COSEWIC 2002). There may also be negative consequences on other species listed as “at risk” or “threatened” such as grizzly bear (Ursus arctos) and wolverine (Gulo gulo).The proliferation of standing dead trees can initially add to habitat available for cavity nesting bird species, and can serve as feeding areas for insectivorous bird and mammal species (Bunnell et al 2004). However, MPB attacked trees are prone to breakage at the base of the tree, and over half of dead trees can be expected to fall in the initial ten years following infestation (Lewis and Hartley 2005). This can increase the amount of surface fuel and could promote large-scale high-intensity forest fires, further reducing the amount of forests remaining and further impacting ecosystem function, forest users and the forest industry. Determining susceptibility of forest stands to MPB attack is a critical aspect of MPB management. This review is intended to provide a summary of techniques available for

76


estimation of susceptibility variables using remote sensing datasets, and a critical assessment of these techniques. A background and summary of MPB biology and management options is also provided for context. The aim of this paper is to review the success of methods currently available for estimating forest variables, and to propose future directions for research in improving accuracy of forest structural attribute estimation using remote sensing data sets. Background on Mountain Pine Beetle Mountain pine beetle is a native forest insect and an endemic part of natural lodgepole pine forests. At low population levels they are part of a healthy ecosystem, aiding in nutrient cycling and forest thinning by elimination of weak and stressed trees. However, when conditions are favourable, MPB populations can grow rapidly resulting in epidemic outbreaks (Bradley 1989). Attacks of lodgepole pine are initiated during the summer when beetles successfully enter the cambium of a host tree. The initial attacking MPBs release a pheromone which serves to attract large numbers of beetles to the host tree. This is termed “mass attack” (Wulder et al 2004). Mass attack is intended to overcome trees natural defences, such as the release of defensive chemicals and utilization of carbohydrate stores to produce excess sap to force out the beetles (Bradley 1989). Upon successful penetration of the cambium the beetles also introduce spores of blue-stain fungus. This fungus undergoes rapid growth in the carbohydrate rich environment and soon clogs the transport systems of the tree, leading to tree desiccation and mortality (Safranyik 1989). The MPBs then lay eggs under the bark of the dead tree in galleries, and overwinters. Mountain pine beetle reproduction is climate dependent and they are generally able to produce one brood per year (Wulder et al 2004a). Mountain pine beetle can be found as far south as Mexico, north along the Rocky Mountains to southern B.C. and into the eastern slopes of Alberta. They prefer lodgepole pine, but will attack any pine species depending on availability (Hiratsuka et al 1995). There are two main environmental factors that limit the spread of MPB: (i) the amount of mature lodgepole pine habitat on the landscape (Taylor and Carroll 2004) and; (ii) favourable climatic conditions (Carroll et al 2004). The current MPB outbreak is a result of favourable conditions of both factors. Intensive fire suppression campaigns through most of the 20th century have led to an abundance of mature pine stands (Taylor and Carroll 2004), and successive years of mild winters have led to an increase in overwinter survival of massive numbers of beetles (Carroll et al 2004). These favourable conditions are allowing MPB to expand their historical range into new areas, including northern Canada, higher elevations and increasingly eastern areas across North America (ASRD 2006b). Carroll et al (2006) have shown that MPB will continue to expand their range into climatically unfavourable areas due to climate warming. An examination of climatically suitable habitats has shown an increase of suitable habitat from 1921 to 2000 into areas formerly unfavourable for MPB colonization, including the boreal forest ecosystem (Carroll et al 2006). This increases the likelihood that MPB will spread through the boreal forest, and across North America, if effective management and control measures are not implemented. Susceptibility mapping

77


can play a key role in determining the best locations and methods for monitoring and management of an outbreak of MPB across North America. Monitoring of MPB is accomplished through a combination of methods including aerial surveys, pheromone bait stations, beetle probing of trees in susceptible areas and analysis of remote sensing imagery. The methodology is largely dependent on the scale of the survey and the extent of the area of interest. Monitoring programs conducted in Alberta in 2005 and 2006 were able to confirm the presence of MPB at 14 pheromone bait stations and along numerous flight transects along the eastern slopes of the Rocky Mountains (ASRD 2005, 2006a). Active beetle probing programs are also conducted across B.C. and along the eastern slopes of the Rocky Mountains in Alberta in order to determine where management efforts should focus (ASRD 2006b). Management options for MPB fall into two categories; direct and indirect. Direct control methods are focused on destruction and physical removal of MPB from a forest stand. This is accomplished through such methods as harvesting and processing infected timber during the MPB overwintering period; felling and burning infested wood; debarking of infested wood and; application of the insecticidal herbicide monosodium methanearsonate (MSMA) (Shore and Safranyik 2004). Indirect control methods involve manipulation of potential habitat to reduce forest stand susceptibility to MPB attack. This is accomplished through silvicultural methods and typically involve stand thinning to reduce available habitat while still allowing for future harvest of mature timber (Whitehead and Russo 2005). An action plan for MPB management in Alberta, with a goal of mitigating the social, economic and environmental effects of MPB in Albertan forests, outlines the primary method of managing MPB as the removal of the most susceptible prime brood stands before MPB are able to infest the given stand (ASRD 2006b). Predicting Susceptibility The need for predicting forest stand susceptibility in advance of MPB attack is key in planning control and preventative mitigation measures. Indirect control methods require lead time in order to be implemented effectively, and mapping susceptibility over large areas has the potential to fulfill this requirement. There have been a number of susceptibility indices developed over time (Table 13), but the one being implemented by the B.C., Alberta and Canadian governments is the susceptibility index developed by Shore and Safranyik (1992). This model derives susceptibility as:

LDAPS ×××= Eq. 1 Where S is the susceptibility of a site to attack by mountain pine beetle; P is the percent of total stand basal area made up of pine; A is an age factor; D is a stand density factor (stems/hectare); and L is a location factor. Susceptibility values range from 0 to 100 when calculated with this formula (Shore and Safranyik 1992). The advantage of the Shore and Safranyik (1992) model over alternative models is that the Shore and Safranyik technique returns a range of continuous values that can then be divided into classes of the user’s choosing.

78

Chapter 2: Remote Sensing Mapping and Research Table 13: Summary of mountain pine beetle susceptibility models.

Author(s), Year Variables Used for Susceptibility Summary

Amman et al 1977 Age, elevation, DBH This model tends to overestimate susceptibility, and incorporation of stand basal area, as per Shore and Safranyik (1992) results in a more accurate representation of stand susceptibility (Randall and Tensmeyer 2000).

Mahoney 1978 Periodic radial growth Periodic radial growth is a ratio of the most recent five years of growth compared to the previous five years of growth. Lower periodic radial growth values indicate a decrease in tree vigour over time, thus increasing a trees susceptibility to MPB attack due to potentially weakened defensive responses. This model has not been able to accurately predict MPB attack (Shore et al 1989; Bentz et al 1993). The primary problem with this model is that trees typically slow in growth after age 30, thus resulting in lower scores for all mature trees.

Berryman 1978 Stand resistance, phloem thickness Stand resistance was determined from a ratio of periodic growth ratio and stand hazard rating. Stand hazard rating was calculated from a crown competition factor and stand density. However crown competition factor has been found to be inversely related to MPB mortality (Shore et al 1989). Thus this system accumulated problems associated with both periodic growth ratio and crown competition factor, and consequently had low rates of success in predicting susceptibility (Katovich and Lavigne 1986).

Anhold and Jenkins 1987 Stand density index It was found that stand density index was not a good predictor of MPB population changes, but stand density index values were found to give accurate predictions of a stands potential for attack. This implied that stand density was a significant variable for determining MPB susceptibility.

Shore and Safranyik 1992 Age, location (latitude, longitude and elevation), stand density, % pine basal area

This model builds on the efforts of previous susceptibility models and incorporates some of the positive aspects of those models. Testing of the Shore and Safranyik model found that 40 of 41 test sites evaluated fell within the 95% confidence interval, and showed a strong relationship between tree mortality and the susceptibility index. This model is also currently being used by the B.C., Albertan and Canadian governments. As such, it is the model that will be considered for the purposes of this paper.

79


Remote Sensing of Susceptibility Variables Typically, forest measurements are collected through labour intensive field programs, but over the last twenty years, techniques, models and protocols have been developed that allow resource managers to accurately segment forested and non-forested areas, define individual tree crowns, identify species, obtain tree measurements and group stands according to need (Gougeon and Leckie 2003). Originating with the interpretation of aerial photographs, remote sensing has become a valuable tool for resource managers. Aerial photographs can be used to determine stand heights, crown closure/cover and stand density. This is accomplished through a variety of methods established over time that are proven to accurately estimate forest stand attributes (Howard 1991). However, these measurements are subjective and vary depending on an individuals training, experience and interpretation proficiency (Congalton and Green 1998). The launch of the Landsat series of satellites allowed for large-scale evaluation of forest resources through multispectral analysis techniques. The multi-spectral resolution of Landsat allows for determination of variables such as species composition, forest health and change detection. However, the coarse spatial scale of the imagery does not allow for resolving of individual trees making estimation of forest structural attributes difficult (Wulder et al 2005). Other multispectral systems have been developed that have finer spatial resolution, such as SPOT 5, IKONOS and QuickBird, making forest inventory with remote sensing imagery more feasible. A good deal of success has been achieved with these systems (Cohen and Spies 1992; Nelson et al 2004; Kayitakire et al 2006). Recent advances in remote sensing, such as hyperspectral imagery and LiDAR, can allow resource managers to estimate forest structural variables to accuracies comparable with conventional aerial photographs. Basal Area Basal area is the cross-sectional area of a single tree stem measured at breast height, or the sum of all single stems in a given stand, expressed per unit of land area (Avery and Burkhart 2002). Basal area can be extracted from remote sensing imagery due to correlations between basal area and other structural variables that can readily be measured remotely, such as height or crown width. A number of studies have been conducted in order to estimate basal area from remote sensing datasets. Accuracy of interpretation of digital air photos has been shown to be improved by integrating auxiliary data sources, such as field samples and existing forest inventories, with interpretation results (Tuominen et al 2003). Hyyppa et al (2000a) has also shown that aerial photographs can be used to give reasonable estimates of basal area. Using Landsat ETM+ to map basal area should be based on spectral signature of the forest stand, as the spatial resolution of Landsat imagery is too coarse for methods of estimation such as texture analysis or image segmentation (Cohen and Spies 1992). Using the kNN estimation procedure can significantly increase accuracy when using Landsat imagery, with lower k values yielding better accuracies. However care must be taken when applying kNN as the representativeness of the training samples has a strong influence on the kNN results due to the non-parametric nature of the procedure (Maselli et al 2005). The kNN variation used (e.g. distance, weighted function, number of neighbours) also has a strong influence on the accuracy of the estimation.

80


Mean stand texture of SPOT 4 panchromatic imagery shows strong correlation with basal area, indicating that these data capture the relative degree of sunlight and shadow associated with stand complexity. However, there is little relationship between forest structural attributes and original SPOT data (Cohen and Spies 1992). The accuracy of Radio Detection and Ranging (RADAR) for determining basal area requires further study before it can be applied for forest inventory purposes. The addition of backscatter variables in RADAR models improves accuracy of estimates of forest structure (Hyyppa et al 1997, 2000a). Studies have shown strong relationships between RADAR and basal area, but low R2 values imply a need for further refinement of RADAR-based models for estimation of basal area (Hyyppa et al 2000a). LiDAR data return estimates of basal area with very high accuracy (Hyyppa et al 2000a). Integrating LiDAR with multispectral or panchromatic imagery provides little additional accuracy in estimates of basal area. However, LiDAR has been shown to overestimate the number of trees per hectare compared to ground estimates, resulting in an overestimation of basal area. Age Regardless of sensor type, tree age is the most difficult variable to estimate with remote sensing data. When determining age from remote sensing datasets crown diameter, tree height, spectral signature or image texture are the variables typically utilized in age models. Numerous techniques have been developed for estimating forest stand age with remote sensing imagery. The use of neural networks has provided insights into the non-linear nature of tree age, and results suggest that it may be possible to accurately predict age from these models. Coregistration of Landsat TM imagery and topographic models provides the most accurate results when neural networks are employed. Reflectance in the near-infrared (NIR) portion of the spectrum has shown a strong inverse relationship with pine age making aging younger stands the most valid application for those models. Age has also shown to be closely correlated with most other attributes (e.g., tree height, crown diameter, DBH) in a forest stand (Cohen and Spies 1992). Using historical Landsat TM image integrated the Normalized Difference Vegetation Index (NDVI), tasselled-cap greenness and wetness, a Band 5 differencing index and a vegetation condition index to detect forest growth over a two year period. While errors in expected areas of change (i.e. cutblocks that were not immediately replanted) provide a good method for focussing field verification programs, historical differencing of imagery does not provide any information on tree age, unless a series of multitemporal images can be used to accurately age a forest stand.

81


Tasselled cap indices determined from Landsat ETM+ imagery, particularly wetness and brightness, show moderate correlation with stand age (Wulder et al 2004b). The correlation with wetness increases as the forest matures and plateaus at around 200 years (Cohen and Spies 1992). Using a stepwise segmentation process, Cohen et al (1995) demonstrated that determining age classes from Landsat imagery requires further research. Utilizing the tasselled cap (wetness, greenness and brightness) Cohen et al (1995) segmented closed canopy forest stands and then classified those into young and mature age classes. Attempts to further classify those classes into more age categories yielded results of decreasing accuracy with increasing numbers of age classes. It was also shown that most of the spectral variation in an image can be attributed to Band 4 (NIR) and Band 5 (MIR) of TM images (Cohen et al 1995).

Texture and backscatter of airborne Synthetic Aperture RADAR (SAR) was evaluated, and determined to be of little value in estimating forest age in temperate forests, while returning strong correlations in tropical forests. The reason is likely due to differences in forest stand development in temperate versus tropical ecosystems.

It is possible to aggregate homogeneous tree age classes into polygons, using panchromatic IKONOS imagery and Voronoi polygons. The best successes are achieved in mature forest stands, with accuracy of the aggregation decreasing in younger stands. Transition zones from young to mature trees also confound the model. This method allows for easy removal of non-forest polygons as they are aggregated together and may allow for fuzzy boundaries inherent with forest stands (Nelson et al 2004).

Stand Density Stand density is a quantitative measure of the number of trees, typically of a given minimum diameter, in a given unit area (Avery and Burkhart 2002). Generally, younger stands have higher density values than more mature stands. This is primarily due to natural thinning processes associated with forest succession. Determining stand density from remote sensing data is typically derived from individual tree segmentation, or image texture analysis. .

Stand density is typically assessed with 1:5,000 scale aerial photograph stereo pairs. This method has shown to be very strongly correlated with field stem counts.

Landsat imagery can be used to accurately estimate stand density after application of kNN filtering. The use of three Landsat ETM+ spectral indices (i.e. NDVI, IRI and MIRI) has been shown to have no correlation with stem density. Strong correlations between IKONOS texture and image analysis and stand density have been found, yielding stem counts comparable to 1:5,000 scale aerial photograph interpretation (Kayitakire et al 2006).

LiDAR has proven to be of enormous utility in determining estimates of stand density and structure. LiDAR alone has proven to be of equal capability in estimating stand density as when integrated with multispectral or panchromatic Advanced Land Imager (ALI) imagery.

82


Future Directions It is apparent that a comprehensive study of the ability to predict susceptibility variables from remote sensing datasets, and an evaluation of each of the techniques used along with a quantified evaluation of each of the data types (e.g. Landsat, IKONOS, SPOT, and LiDAR) is required. Error analysis of different forest structural attribute estimation techniques is needed to understand where real innovations have been made, and where different estimation and analysis techniques produce results not significantly more accurate than others. Existing susceptibility models were not developed with remote sensing applications in mind. As a result it is difficult to directly estimate the variables necessary for susceptibility calculations with remote sensing datasets. For example, the Shore and Safranyik (1992) model requires basal area, tree age and stand density for determining susceptibility. While those variables can be derived through allometric models of tree structure and growth habits, some of these variables cannot be directly measured via remote sensing. In order to obtain the most accurate estimates of MPB susceptibility possible, a model derived from attributes that can be directly and simply estimated with remote sensing datasets is needed. In this way large-scale maps could be created and easily updated on the same temporal scale as the remote sensing platform of the model. Estimation of forest attributes from remote sensing data is a diverse field, with many techniques, models and methods developed over time. The need for protocols for predicting forest attributes for use in MPB susceptibility modelling is apparent. The large amount of data types and the increasing availability of this data can have a large influence on the types of models that will be most applicable for predicting MPB susceptibility. For instance, Landsat has been collecting data for decades and it is readily available, while airborne based techniques, such as CASI, AVIRIS and LiDAR, are relatively new and require custom flight missions for data collection. As the advantage of remote sensing data analysis for determining forest attributes and MPB susceptibility is partly cost based, if data collection missions are required for large-scale assessments of MPB susceptibility the cost advantage of remote sensing may decrease when compared to field data collection programs. The need for a quick and simple method for mapping MPB susceptibility in the near future is great. The potential for ecosystems and economies to suffer drastic effects needs to be mitigated before the devastation witnessed in western Canada and the northern United States continues to spread eastward into the boreal forest ecosystem. Literature Cited ASRD (Alberta Sustainable Resource Development). 2005. Incidence of mountain pine

beetle in Alberta, 2005. Alberta Sustainable Resource Management, Forest Protection Division. Edmonton, AB.

ASRD. 2006a. General locations of mountain pine beetle attacks in Alberta, 2006. Alberta Sustainable Resource Management, Forest Protection Division. Edmonton, AB.

ASRD. 2006b. Mountain Pine Beetle Action Plan for Alberta. Alberta Sustainable Resource Management, Forest Protection Division. Edmonton, AB. 8 pp.

Avery T.E. and Burkhart H.E. 1994. Forest Measurements, 4th edn. McGraw-Hill, New York.

83


Badeck, F.-W., Bondeau, A., Böttcher, K., Doktor, D., Lucht, W., Schaber, J. and Sitch, S., 2004, Responses of spring phenology to climate change. New Phytologist, 162, 295 309.

Beck, P. S. A., Atzberger, C., Høgda, K. A., Johansen, B. and Skidmore, A. K., 2006, Improved monitoring of vegetation dynamics at very high latitudes: a new method using MODIS NDVI. Remote Sensing of Environment, 100, 321-334.

Beck, P. S. A., Jönsson, P., Høgda, K.-A., Karlsen, S. R., Eklundh, L., and Skidmore, A. K., 2007, A ground-validated NDVI dataset for monitoring vegetation dynamics and mapping phenology in Fennoscandia and the Kola peninsula. International Journal of Remote Sensing, 28, 4311-4330.

Bentz, B.J., G.D. Amman and J.A. Logan. 1993. A critical assessment of risk classification systems for the mountain pine beetle. Forest Ecological Management. 61: 349-366.

Bradley, T. 1989. Mountain pine beetle literature review. Silva Ecosystem Consultants Ltd. Revised 1993. 15 pp.

Bunnell, F.L., K.A. Squires and I. Houde. 2004. Evaluating the effects of large-scale salvage logging for mountain pine beetle on terrestrial and aquatic vertebrates. Mountain Pine Beetle Initiative Working Paper 2004-2. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Carroll, A.L., S.W. Taylor, J. Regniere and L. Safranyik. Effects of climate change on range expansion of mountain pine beetle in British Columbia. pp. 223-232 in T.L. Shore, J.E. Brooks and J.E. Stone, eds. Mountain Pine Beetle Symposium: Challenges and Solutions, October 30 to 31, 2003. Kelowna, BC. Information Report BC-X-399. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Congalton, R.G. and K. Green. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. Mapping Science Series. Taylor and Francis CRC Press, Boca Raton, Florida. 137 pp.

Cohen, W.B. and T.A. Spies. 1992. Estimating structural attributes of Douglas-fir /western hemlock forest stands from Landsat and SPOT imagery. Remote Sensing of Environment. 41: 1-17.

Ferguson G.A.1981. Statistical analysis in psychology and education (5th ed.). New York: McGraw-Hill

Franklin, S. E., G. B. Stenhouse, M. J. Hansen, C. C. Popplewell, J. A. Dechka, and D. R. Peddle, 2001: An integrated decision tree approach (IDTA) to mapping land cover using satellite remote sensing in support of grizzly bear habitat analysis in the Alberta yellowhead ecosystem. Canadian Journal of Remote Sensing, Vol. 27, No. 6, pp. 579-592.

Garshelis, D.L., Gibeau, M.L., and Herrero, S., 2005. Grizzly Bear Demographics in and Around Banff National Park and Kananaskis Country, Alberta. Journal of Wildlife Management 69, 277 – 297.

Gougeon, F.A. and D.G. Leckie. 2003. Forest information extraction from high spatial resolution images using an individual tree crown approach. Information Report BC-X-396. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Government of British Columbia. 2006. Mountain Pine Beetle Action Plan 2006-2011: Sustainable forests, sustainable communities. 21pp.

84


Hiratsuka, Y., D.W. Langor and P.E. Crane. 1995. A field guide to forest insects and diseases of the prairie provinces. Special Report 3. Natural Resources Canada, Northern Forestry Centre. Edmonton, Alberta. 297 pp.

Hobson, D., 2005, Denning of Grizzly Bears in the Foothills Research Institute. In Foothills Research Institute Grizzly Bear Research Program 1999-2003 Final Report, G. B. Stenhouse and K. Graham (Eds.), pp. 32-37.

Howard, J.A. 1991. Remote sensing of forest resources: Theory and application. Chapman and Hall. London, UK. 420 pp.

Hyyppa, J., H. Hyyppa, M. Inkinen, M. Engdahl, S. Linko and Y. Zhu. 2000a. Accuracy comparison of various remote sensing data sources in the retrieval of forest stand attributes. Forest Ecology and Management. 128:109-120.

Jennings, S.B., Brown, N.D., Sheil, D., 1999. Assessing forest canopies and understorey illumination: canopy closure, canopy cover, and other measures. Forestry 72, 59–73.

Jönsson, P. K. and Eklundh, L., 2002, Seasonality Extraction by Function-fitting to Time Series of Satellite Sensor Data. IEEE Transactions on Geoscience and Remote Sensing, 40, 1824-1832.

Jönsson, P. and Eklundh, L., 2004, TIMESAT – a program for analyzing time-series of satellite sensor data. Computers and Geosciences, 30, 833-845.

Kansas, J., 2002. Status of the Grizzly Bear (Ursus arctos) in Alberta. Alberta Sustainable Resource Development, Fish and Wildlife Division, and Alberta Conservation Association. Wildlife Status Report No. 37, Edmonton, AB. 43 pp.

Katovich, S.A. and R.J. Lavigne. 1986. The applicability of available hazard rating systems for mountain pine beetle in lodgepole pine stands of southern Wyoming. Canadian Journal of Forest Resources. 16: 222-225.

Kayitakire, F., C. Hamel and P. Defourny. 2006. Retrieving forest structure variables based on image texture analysis and IKONOS-2 imagery. Remote Sensing of Environment. 102: 390-401.

Koehler, G.M. and K.B. Aubry. 1994. Lynx. In The scientific basis for conserving forest carnivores: American marten, fisher, lynx, and wolverine in the western United States. L.F. Ruggiero, K.B. Aubry, S.W. Buskirk, I.J. Lyon and W.J. Zielinsky (eds.). General Technical Report RM-254. U.S. Department of Agriculture, Forest Service. Fort Collins, Colorado. 74-98.

Linke, J., Franklin, S.E., Huettmann, F., and Stenhouse, G.B., 2005. Seismic cutlines, changing landscape metrics and grizzly bear landscape use in Alberta. Landscape Ecology 20, 811 – 826.

Langford WT, Gergel SE, Dietterich, TG Cohen W (2006) Map misclassification can cause large errors in landscape pattern indices: examples from habitat fragmentation. Ecosystems 9: 474-488

Lewis, K.J. and I. Hartley. 2005. Rate of deterioration, degrade and fall of trees killed by mountain pine beetle: A synthesis of the literature and experiential knowledge. Mountain Pine Beetle Initiative Working Paper 2005-14. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Mace, R.D., Waller, J.S., Manley, T.L., Lyon, L.J., and Zuuring, H., 1996. Relationships Among Grizzly Bears, Roads and Habitat in the Swan Mountains Montana. The Journal of Applied Ecology 33 (6), 1395 – 1404.

85


Martin, K, A. Norris and M. Drever. 2006. Effects of bark beetle outbreaks on avian biodiversity in the British Columbia interior: Implications for critical habitat management. BC Journal of Ecosystems and Management. 7(3): 10-24.

Maselli, F., G. Chirici, L. Bottai, P. Corona and M. Marchetti. 2005. Estimation of Mediterranean forest attributes by the application of k-NN procedures to multitemporal Landsat ETM+ images. International Journal of Remote Sensing. 26(17): 3781-3796.

McDermid, G. J., S. E. Franklin, and E. F. LeDrew (2005). Remote sensing for large-area habitat mapping. Progress in Physical Geography, 29(4), 449-474.

McDermid, G. J., S. E. Franklin, and E. F. LeDrew, 2008: A multi-attribute approach to mapping vegetation and land cover over large areas in support of wildlife habitat mapping, Remote Sensing of Environment, in review.

McGarigal, K., Cushman, S. A., Neel, M. C., and Ene, E., 2002. FRAGSTATS: Spatial Pattern Analysis Program for Categorical Maps. Computer software program produced by the authors at the University of Massachusetts, Amherst. http://www.umass.edu/landeco/research/fragstats/fragstats.html accessed March 25, 2009.

Munro, R. H. M., Price, M. H. H. and Stenhouse, G. B., 2005, The Diet of Grizzly Bears, Ursus arctos, in West-Central Alberta, Canada. In Foothills Research Institute Grizzly Bear Research Program 1999-2003 Final Report, G. B. Stenhouse and K. Graham (Eds.), pp. 11-20.

Nelson R.F., Krabill W.B., and Maclean G.A. 1984. Determining forest canopy characteristics using airborne laser data. Remote Sensing of Environment. 15, 201-212

Nelson, T., B. Boots, M. A. Wulder and R. Feick. 2004. Predicting forest age classes from high spatial resolution remotely sensed imagery using Voronoi polygon aggregation. GeoInformatica. 8(2): 143-155.

Nielsen, S.E., Boyce, M.S., Stenhouse, G.B. (2006). A habitat-based framework for grizzly bear conservation in Alberta. Biological Conservation, 130, 217-229.

Nielsen, S.E., Boyce, M.S., Stenhouse, G.B., Munro, R.H.M., 2003: Development and testing of phonologically driven grizzly bear habitat models. Ecoscience, Vol. 10, pp. 1−10.

Philip M.S. 1994. Measuring Trees and forests, 2nd edn. CAB International, Wallingford, UK.

Randall, C.B. and G.T. Tensmeyer. 2000. Hazard rating system for mountain pine beetle in lodgepole pine using the Oracle database and the Forest Service IBM platform. Report No. 00-6. Forest Health Protection, United States Forest Service. Northern Region, Missoula, Montana.

Ritchie J.C., Everitt J, Escobar D, Jackson T, and Davis M. 1992. Airborne lidar measurements of rangeland canopy cover. Journal of Range Management. 45, 189–93

Safranyik, L. 1988. Mountain pine beetle: biology overview. pp. 9-13 in Symposium on the Management of Lodgepole Pine to Minimize Losses to the Mountain Pine Beetle. U.S. Forest Service General Technical Report INT-GTR-162.

Shore, T.L., P.A. Boudewyn, E.R. Gardner and A.J. Thompson. 1989. A preliminary evaluation of hazard rating systems for the mountain pine beetle in lodgepole pine in British Columbia. pp. 28-33 In Proceedings of a symposium on the management of lodgepole pine to minimize losses to the mountain pine beetle. July 12 to 14, 1989. Kalispell, Montana. USDA Forest Service General Technical Report No. INT-262.

86


Shore, T.L. and L. Safranyik. 1992. Susceptibility and risk rating systems for the mountain pine beetle in lodgepole pine stands. Information Report BC-X-336. Natural Resources Canada, Pacific Forestry Centre. Victoria, B.C.

Shore, T.L. and L. Safranyik. 2004. Mountain pine beetle management and decision support. pp. 97-105 In Mountain pine beetle symposium: Challenges and solutions. October 30 to 31, 2003. Kelowna, B.C. Information Report BC-X-399. Natural Resources Canada, Pacific Forestry Centre. Victoria, B.C.

Stenhouse, G.B., Boyce,M.S., Boulanger, J., 2003. Report on Alberta Grizzly Bear Assessment of Allocation. Alberta Sustainable Resource Development, Fish and Wildlife Division, Hinton, Alta.

Taylor, S.W. and A.L. Carroll. 2004. Disturbance, forest age and mountain pine beetle outbreak dynamics in BC: A historical perspective. pp. 41-51 in T.L. Shore, J.E. Brooks and J.E. Stone, eds. Mountain Pine Beetle Symposium: Challenges and Solutions, October 30 to 31, 2003. Kelowna, BC. Information Report BC-X-399. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Tuominen, S., S. Fish and S. Poso. 2003. Combining remote sensing, data from earlier inventories, and geostatistical interpolation in multisource forest inventory. Canadian Journal of Forest Resources. 33:624-634.

Whitehead, R.J. and G.L. Russo. “Beetle-proofed” lodgepole pine stands in interior British Columbia have less damage from mountain pine beetle. Information Report BC-X-402. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Wilson, S.M., Madel, M.J., Mattson, D.J., Graham, J.M., Burchfield, J.A., and Belsky, J.M., 2005. Natural landscape features, human-related attractants, and conflict hotspots: a spatial analysis of human-grizzly bear conflicts. Ursus 16 (1), 117 – 129.

Wilson, S.M., Madel, M.J., Mattson, D.J., Graham, J.M., and Merrill, T., 2006. Landscape conditions predisposing grizzly bears to conflicts on private agricultural lands in the western USA. Biological Conservation 130, 47 – 59.

Wulder, M.A., C.C. Dymond and B. Erickson. 2004a. Detection and monitoring of the mountain pine beetle. Information Report BC-X-398. Natural Resources Canada, Pacific Forestry Centre. Victoria, BC.

Wulder, M.A., R.J. Hall and S.E. Franklin. 2005. Remote sensing and GIS in forestry, Chapter 12. Pp. 351-356 in S. Aranoff ed. Remote Sensing for GIS Managers. ESRI Press. Redlands, California.

Zhang, X., Friedl, M. A., Schaff, C. B. and Strahler, A. H., 2004, Climate controls on vegetation phenological patterns in northern mid- and high latitudes inferred from MODIS data. Global Change Biology, 10, 1133-1145.

87

Chapter 3: Mountain Pine Beetle Remote Sensing Monitoring and Modeling

CHAPTER 3: GRIZZLY BEAR/MOUNTAIN PINE BEETLE INTERACTIONS: REMOTE SENSING MONITORING AND

MODELING YEAR END REPORT Mike Wulder6; Joanne White6, Nicole Seitz6, Nicholas Coops7, Thomas Hilker7, Chris Bater7, Greg McDermid4, Adrian Faraguna4, Gordon Stenhouse1 6Canadian Forest Service; 7University of British Columbia; 4University of Calgary; 1Foothills Research Institute

EXECUTIVE SUMMARY This report describes year one project activities undertaken at the Canadian Forest Service and University of British Columbia. Activities undertaken at the University of Calgary will be under a different chapter. All planned project activities were undertaken and completed. The establishment of software foundation for blending MODIS and Landsat data for reflectance, based on the Gao (2006) model, has been implemented and is summarized in this document. The creation of exploratory remote sensing change maps for MPB attack, harvesting, and mitigation activities in the prototype study area have also been completed and are summarized herein. The development of a sampling strategy to guide the use of field and airborne sampling activities in subsequent years has also been planned. Year 2 project activities will build upon these completed and developed Year 1 outcomes. Analysis Preparation and Support The Landsat path/row combination used was 46/22, near Grand Prairie, AB, for the year 2006 (See Figure 1). Two essentially cloud free images were available, Landsat TM for June 30, 2006, and Landsat ETM+ for September 2, 2006. The June TM image was orthorectified in an image-to-image registration to the September ETM+ image. After attaining a RMS error of less than half a pixel from 25 ground control points, 2 scripts were run on the June TM image; one to convert it to an ETM+ image, and one to correct for Top of Atmosphere reflectance.

88


Figure 1: Grizzly bear focus study area in the Kakwa region. Canada-wide MODIS imagery was collected for 2006. The MODIS data was made up of 7 band composites at 250m spatial resolution (as suggested by Trishchenko et. al. 2006) for the 1st, 11th, and 21st of each month (for a total of 36 dates). Nine dates were used: June 21, July 1, July 11, July 21, Aug 1, Aug 11, Aug 21, Sept 1, and Sept 11. For each of these dates, the 7 band composites were clipped down from the Canada wide extent to the provincial extents of British Columbia and Alberta (See Table 1). Table 1: Landsat TM and MODIS image acquisition dates for path 46, row 22, year 2006 for the Kakwa region.

Landsat TM Acquisition dates

Actions MODIS* Acquisition dates

Actions

January 1, 2006 January 11, 2006 January 21, 2006 February 1, 2006 February 11, 2006 February 21, 2006 March 1, 2006

89


90


Actions MODIS* Acquisition dates

Actions

March 11, 2006 March 21, 2006 April 1, 2006 April 11, 2006 April 21, 2006 May 1, 2006 May 11, 2006 May 21, 2006 June 1, 2006 June 11, 2006

June 30,2006

Orthorectified to Sept 2, 2006 image. Top of atmosphere correction. June 21, 2006

Clipped AB & BC into grizzly.mxd

July 1, 2006 Clipped AB & BC into grizzly.mxd



August 1, 2006 Clipped AB & BC into grizzly.mxd



September 2, 2006

Orthorectified. Top of atmosphere correction (U of C). September 1, 2006

Clipped AB & BC into grizzly.mxd

September 11, 2006 Clipped AB & BC into grizzly.mxd

September 21, 2006 October 1, 2006 October 11, 2006 October 21, 2006 November 1, 2006 November 11, 2006 November 21, 2006 December 1, 2006 December 11, 2006 December 21, 2006 * MODIS 250m available: 3 dates per month (1st, 11th, 21st) for 2006 (36 dates). **Also have Aug 11, 2004 Orthorectified to Sept 2, 2006. Top of atmosphere correction.

A second study area and year were decided upon in terms of availability of more cloud free imagery. Landsat TM imagery for path/row 47/24, around Williams Lake, British Columbia, for the year 2001 (See Figure 2) was used. Four dates were included: May 5, July 9, Aug 8,


and Sept 27. These images were orthorectified in an image-to-image registration to an ETM+ image for October 5, 2001. After an RMS error of less than half a pixel was attained with at least 30 ground control points, the same two scripts were run on the TM images to convert them to ETM+ images, and to correct them for Top of Atmosphere reflectance.

Figure 2: Grizzly bear test site in the Williams Lake area. As done with the 2006 study area, MODIS imagery for the year 2001 was collected for 250m spatial resolution 7 band composites. These Canada wide images were clipped down to the provincial boundaries of British Columbia and Alberta. This was done for each of the 36 dates produced for the year 2001 (See Table 2). Finally, a research paper is well underway to describe data-blending; one leg of this extensive project. This is a review paper on the possible approaches to the derivation of 30m reflectance from 250m data using data blending. Much of the basis for the paper was supported by work previously done by Gao et al (2006), with whom we are collaborating. We are also aware of other data sources available (Table 3), although a data blending approach is preferred. Various footprint sizes of the possible image sources are also presented in Figure 2.

91


Table 2: Landsat TM and MODIS image acquisition dates for path 47, row 24 year 2001 for the Williams Lake study area.


Actions MODIS Acquisition dates

Actions

January 1, 2001 Clipped AB & BC into grizzly.mxd



February 1, 2001 Clipped AB & BC into grizzly.mxd



March 1, 2001 Clipped AB & BC into grizzly.mxd



April 1, 2001 Clipped AB & BC into grizzly.mxd



May 6, 2001 Orthorectified to Oct 5, 2001 image. Top of

atmosphere correction.

May 1, 2001 Clipped AB & BC into grizzly.mxd



June 1, 2001 Clipped AB & BC into grizzly.mxd

June 11, 2001 Clipped AB & BC into grizzly.mxd

July 9, 2001 Orthorectified to Oct 5, 2001 image. Top of






92


93


Actions MODIS Acquisition dates

Actions

August 10, 2001 Orthorectified to Oct 5, 2001 image. Top of







September 27, 2001

Orthorectified to Oct 5, 2001 image. Top of



October 5, 2001 October 1, 2001 Clipped AB & BC into grizzly.mxd

October 11, 2001 Clipped AB & BC into grizzly.mxd

October 21, 2001 Clipped AB & BC into grizzly.mxd

November 1, 2001 Clipped AB & BC into grizzly.mxd



December 1, 2001 Clipped AB & BC into grizzly.mxd




Table 3: Select moderate spatial resolution image sources.

Sensor Number

of Scenes

Total Cost of Acquisition

Swath Width (km) or Scene Dimensions (km x km)

Return Period

Spatial Resolution

(m) Band Wavelengths

(μm) Radiometric Resolution Contact

IRS ResourceSat-1

AWiFS 1 $700-$1,000 740 ~ 2 weeks 56 0.52-0.59, 0.62-0.68,

0.77–0.86, 1.55-1.70 10 bit

IRS ResourceSat-1

LISS-III 5

$2,500-$3,575/scene =

$12,500 - $17,875

141 ~ 2 weeks 24 0.52-0.59, 0.62-0.68, 0.77–0.86, 1.55-1.70 7 bit

Customer Support ASRC Management Services

Ph: 877.480.6255 email: [email protected]

www.asrcms.com

SPOT-2 HRV

0.50-0.59, 0.61-0.68, 0.78-0.89, 0.50-0.73

(pan)*

*Must choose between MS and Pan

SPOT-4 HRVIR

9+

$500/scene, plus a tasking cost of $250*/scene =

$6,750

*Academic pricing

60 x 60

26 days; 1 and 4 days alternatively

when viewing off-

nadir

20

10 m pan 0.50-0.59, 0.61-0.68, 0.78-0.89, 1.58-1.75,

0.61-0.68 (pan)*

*Must choose between MS and Pan

8 bit

Trevor Armstrong Technical Services

Representative ATIC - Alberta Terrestrial

Imaging Center www.imagingcenter.ca

Ph: 403.317.9188 ext.22 email:

[email protected]

Landsat 5 TM $510/scene = $1,020

30

120 (thermal)

0.45-0.52, 0.52-0.60, 0.63-0.69, 0.76-0.90, 1.55-1.75, 10.4-12.5,

2.08-2.35

Landsat 7 ETM+

2

$720/scene = $1,440

185 16 days 30

60 (thermal)

15 (pan)

0.450-0.515, 0.525-0.605, 0.630-0.690, 0.750-0.900, 1.55-1.75, 10.40-12.50,

2.09-2.35, 0.50-0.90 (pan)

8 bit

Nicole Danyluk Client Service Representative

MDA - Geospatial Services Ph: 604.231.4970

email: [email protected]

94

mailto:[email protected]

http://www.asrcms.com/

http://www.imagingcenter.ca/




95

Figure 3: Select image footprints over Kakwa study area.


DATA BLENDING: BLENDING OF LANDSAT AND MODIS DATA FOR THE GENERATION OF HIGH SPATIAL AND TEMPORAL RESOLUTION IMAGERY

FOR CHANGE DETECTION IN VEGETATION Introduction Mapping of vegetation and habitat attributes such as land cover, crown closure, species composition, phenology, and change is an essential requirement for estimating wildlife habitats and biodiversity. Determination of such attributes requires parameterization of the continuous land surface, which over large areas is only feasible using remote sensing. A suitable data source for habitat mapping is moderate to high spatial resolution satellite imagery (Cohen and Goward, 2004), the limited availability of this source however, often requires concessions to be made between spatial and temporal resolution. Typically, high spatial resolution imagery has a smaller footprint, or spatial extent, and thus it takes the satellite longer to revisit the same location on earth (Coops et al 2006). Conversely, high temporal resolution sensors have a more frequent revisit rate and produce wide-area coverage, which is, however, often achieved at the cost of lower spatial resolution. One example of a moderate to high spatial resolution satellite product is Landsat. The instrument has a narrow field of view of 15o and a high spatial resolution of 30m, while each image covers an area of 185 x 185 km. These properties have proven extremely useful for monitoring land cover and land cover changes (Townshend et al, 1991, Vogelmann et al, 2001, Gao et al, 2006), the 16-day Landsat revisit cycle together with frequent cloud contamination, however, have limited the application of Landsat data in detecting rapid surface changes that are crucial to monitoring and detecting intra-seasonal ecosystem disturbance and change (Gao et al, 2006, Ju and Roy, 2007). One possible solution to improve the temporal resolution of Landsat is to combine it with the high temporal frequency of coarse-resolution sensors, such as NASA’s Moderate Resolution Imaging Spectroradiometer (MODIS). MODIS imagery is collected over the same place on Earth at least once per day at a spatial resolution of 250-500 nm in the visible and near infrared region (Channels 1-7). Gao et al (2006) introduced a temporal adaptive reflectance fusion model (STARFM) algorithm to blend Landsat and MODIS surface reflectance to produce a synthetic high frequency surface reflectance product at 30m spatial resolution. This part of the report describes the application of the STARFM algorithm to the study area for the purpose of change detection and grizzly bear habitat modeling. Study area and data description The study area is located in west central Alberta, Canada, near the towns of Grande Cache and Grande Prairie and extends further south into Willmore Wilderness Park (Figure 1). Landsat and MODIS data were acquired for the study site throughout the summer of 2006. A total of nine MODIS 10 day composites (06/21, 07/01, 07/11, 07/21, 08/01, 08/11, 08/21, 09/01, 09/11) and two clear Landsat scenes (06/01 and 08/11) were available and acquired for this time period and were used for data blending. MODIS data were extracted for the greater study area and clipped to the extent of the Landsat imagery to facilitate data blending.

96


Methods Improving MODIS spatial resolution While MODIS bands B1 and B2 are sampled at a spatial resolution of 250m, bands B3-7 provide a spatial resolution of only 500m, which can be limiting for land applications. It has been shown that the optimal spatial resolution should be of the order of a few hundred meters for operational crop monitoring and other land monitoring applications (Trishchenko et al 2006). As a result, a first step in data processing was to improve the spatial resolution of MODIS B3-7 to be able to synthesize Landsat data from a 250m basis by downscaling the 500m resolution MODIS imagery to 250m resolution (Trishchenko et al 2006). The approach is based on the generation of nonlinear regressions between bands B3-B7 and B1-B2, and the Normalized Difference Vegetation Index (Trishchenko et al, 2006) to produce a set of regression coefficients which, when being applied to the 250m data of B1 and B2, yield 5 synthetic channels (B3-B7) at 250m resolution. The generated 250m resolution images were then normalized to preserve radiometric consistency (Trishchenko et al, 2006). This method allowed the use of 500m spatial resolution land bands with terrestrial applications. Generating Landsat ETM data The quality and consistency of change detection using remote sensing depends on the imagery processing required to address issues related to image radiometry, normalization, and computation of the spectral indexes (Han et al, 2007). These processing steps are typically undertaken independently, thereby increasing the risk of computation errors and data inconsistencies. Han et al (2007) introduced an algorithm to efficiently approach change detection from Landsat-5 and -7 imagery based on tasselled cap transformation. The approach mitigates data inconsistencies between the different Landsat data types and reduces errors introduced by image processing, to allow threshold-based change detection (Han et al, 2007). The algorithm was used in this study to normalize Landsat 5 and 7 data to a common basis of top-of-atmosphere reflectance to facilitate a direct comparison of radiometric values with the 10-day MODIS composites. Data blending MODIS and TOA-corrected Landsat data were transformed to a common coordinate system (NAD 1983, UTM Zone 11) and MODIS data were resampled to a spatial resolution of 30 m using nearest neighbours. MODIS data were then clipped to the extent of the Landsat imagery to allow a direct comparison on a pixel by pixel basis. The six Landsat bands resulting from the tasselled cap transformation (Han et al, 2007) were associated with the closest available MODIS bands as shown in Table 4 and each pair of bands was processed individually.

97


Table 4: Association of Landsat and MODIS bands for STARFM prediction. Landsat MODIS

Band Spectral range Name Band Spectral range 1 450-520 nm Blue 3 459-479nm 2 520 – 600 nm Green 4 545-565nm 3 630-690 nm Red 1 620-670nm 4 760-900 nm Near IR 2 841-876nm 5 1550-1750 nm Mid IR 6 1628-1652nm 7 2080-2350 nm Mid IR 7 2105-2155nm

Data blending was performed for each band individually following Gao et al (2006). STARFM predicts values for an individual Landsat pixel (r,c) by comparing the spectral differences between

1. the Landsat and MODIS base images (at the time the MODIS and Landsat image was taken, t=1) and

2. the Landsat base image (t=1) and the MODIS image taken at the prediction time (t=2) within a 50 x 50 m window around r,c. The difference values are then weighted according to their spatial distance to r,c (Figure 4).

Difference image between Landsat and MODIS base imagery (t=1)

Difference image between Landsat base image (t=1) and MODIS image sampled at prediction time (t=2)

Output Landsat ETM prediction (t=2)

Pixel p(r,c) in input image 2

Window around pixel p(r,c) (50*50 pixels)

Pixel p(r,c) in input image 1

Pixel p(r,c) in output image

Window around pixel p(r,c) (50*50 pixels)

Figure 4: Simplified schema of the STARFM prediction algorithm. Validation of STARFM predictions Validation of the STARFM made predictions for high frequency high spatial resolution data was a challenge at the site, as the existing cloud contamination for most of the 16 day

98


Landsat scenes did not allow us to validate the STARFM based results by sub-sampling Landsat data and comparing the predictions to the observed values from the remaining images. As a result, a general validation approach of the STARFM predictions was performed and a second study site was chosen at a similar, also Mountain Pine Beetle (Dendroctonus ponderosae) infested area in central British Columbia, near the town of Williams Lake (N 52o 10’, W 122o 03’) in order to assess whether the STARFM based predictions were actually capable of tracking sub-seasonal changes, such as disturbance by Mountain Pine Beetle, in coniferous vegetation. At the validation site, 36 10-day MODIS composites were available and downloaded from the MODIS website. Additionally, a total of 5 cloud free Landsat (TM and ETM) scenes (05/06, 07/09, 08/10, 09/27, 10/05) were acquired for this site for the year of 2007. MODIS and Landsat scenes were transformed and pre-processed identically to the west central Alberta study site and the STARFM algorithm was applied to predict 36 Landsat scenes throughout 2007 based on the MODIS acquisition dates. Two different tests were performed, trying to predict year round Landsat data from only one Landsat input scene as well as from all available Landsat scenes. The predicted Landsat scenes from both tests were then compared to the observed Landsat data at the given dates to assess how well STARFM was able to predict high resolution observations. Results and Discussion A total of 9 Landsat scenes were predicted for the study site throughout the summer of 2006 (06/21, 07/01, 07/11, 07/21, 08/01, 08/11, 08/21, 09/01, 09/11), identical to the MODIS acquisition dates described in the Executive summary. Figure 5A-J shows the Landsat scenes for 5 acquisition dates for the validation site in the left column, whereas the right column shows the spectral difference (TOA reflectance) between this observed scene and the closest predicted Landsat scene (prediction dates in the figure header) for the example of Landsat ETM band 2. All Landsat scenes were predicted on the basis of the Landsat scene shown in Figure 5E.

99


A B Difference Landsat scene observed at 2001/05/06 and Landsatscene predicted for 2001/05/01. ρ=520-600nm (ETM band:02)

abs(mean): 3.76%abs(σ): 8.36%

-8%

-6%

-4%

-2%

0%

2%

4%

6%

8%

C D

Difference Landsat scene observed at 2001/07/09 and Landsatscene predicted for 2001/07/11. ρ=520-600nm (ETM band:02)

abs(mean): 1.53%abs(σ): 3.56%

-8%

-6%

-4%

-2%

0%

2%

4%

6%

8%

Acquisition date: May 6

Acquisition date: June 9

100


E F


abs(mean): 0.00%abs(σ): 0.00%

-8%

-6%

-4%

-2%

0%

2%

4%

6%

8%

G H


abs(mean): 3.13%abs(σ): 6.82%

-8%

-6%

-4%

-2%

0%

2%

4%

6%

8%

Acquisition date: August 10

Acquisition date: September 27

101


102

H J


abs(mean): 1.06%abs(σ): 1.72%

-8%

-6%

-4%

-2%

0%

2%

4%

6%

8%

Acquisition date: October 05

Figure 5 A-J: Comparison between observed Landsat scene and difference image between observed and predicted Landsat scenes. A B

Figure 6: TOA reflectance for Landsat scene observed on July 11 (A) and the same area predicted for July 21 (B) (both Landsat band 1). As it can be seen from Figure 5A-J, differences between observed and predicted Landsat scenes were largely due to clouds, the mean difference varied between 0% and 3.76% (this is including the clouds). Differences for the other Landsat ETM bands were in the same range as for band 02.


Figure 6 shows a direct comparison between an observed and a predicted Landsat scene for July 11 and July 21, respectively. Figure 5B reveals the great amount of detail that is still visible in the predicted data. Largest differences between observed and predicted images were found in high mountainous areas with no vegetation cover, while the vegetated land surface showed only little differences between prediction and observation. Although the quality assessment of change detection using synthetic Landsat imagery is still a work in progress, first results indicate that the selected STARFM algorithm is able to predict high spatial and temporal resolution Landsat data with a sufficient level of detail. Larger deviations between image prediction and observed scenes were found only in smaller, non-vegetated areas, such as high mountains passes, while the majority of the vegetated land surface was predicted with a high level of accuracy. Mean and standard deviation of the difference image between prediction and observation are small compared to the nominal data range (TOA reflectance). Literature Cited Cohen, W.B. and Goward, S.N. 2004. Landsat's Role in Ecological Applications of Remote

Sensing. Bioscience 54:535-545. Coops, N.C., Wulder, M.A. and White, J.C. 2006. Integrating Remotely Sensed and

Ancillary Data Sources to Characterize a Mountain Pine Beetle Infestation. Remote Sensing of Environment 105:83-97.

Gao, F., Masek, J., Schwaller, M. and Hall, F. 2006. On the Blending of the Landsat and Modis Surface Reflectance: Predicting Daily Landsat Surface Reflectance. Transactions on Geoscience and Remote Sensing 44: 2207-2218.

Han, T., Wulder, M.A., White, J.C., Coops, N.C., Alvarez, M.F. and Butson, C. 2007. An Efficient Protocol to Process Landsat Images for Change Detection With Tasselled Cap Transformation. Geoscience and Remote Sensing Letters 4: 147-151.

Ju, J., & Roy, D.P. 2007. “The Availability of Cloud-free Landsat EMT+ Data Over the Conterminous United States and Globally” Remote Sensing of Environment, in press.

Townshend, J., Justice, C., Li, W., Gurney, C. and Mcmanus, J. 1991. Global Land Cover Classification by Remote-Sensing - Present Capabilities and Future Possibilities. Remote Sensing of Environment 35:243-255.

Trishchenko, A.P., Luo, Y., Khlopenkov, K.V. 2006. “A Method for Downscaling MODIS Land Channels to 250m Spatial Resolution Using Adaptive Regression and Normalization” In: Remote Sensing for Environmental Monitoring, GIS applications, and Geology VI, edited by M. Ehlers, U. Michel. Canada Center for Remote Sensing. Ottawa, ON. vol. 6366.

Vogelmann, J.E., Howard, S.M., Yang, L.M., Larson, C.R., Wylie, B.K. and Van Driel, N. 2001. Completion of the 1990s National Land Cover Data Set for the Conterminous United States From Landsat Thematic Mapper Data and Ancillary Data Sources. Photogrammetric Engineering and Remote Sensing 67:650-+

103


CHANGE AND BEETLE MAPPING: MAPPING FOREST DISTURBANCE USING A LANDSAT-5 TM TIME-SERIES: INITIAL RESULTS

Background Numerous studies have shown that moderate spatial resolution remote sensing imagery can be useful for change detection and characterizing forest disturbance (Coppin and Bauer, 1994; Adams et al, 1995; Fraser and Li, 2002; Kennedy et al, 2007). Recently, Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper+ (ETM+) data have been employed to map mountain pine beetle (MPB) infestations in British Columbia (Skakun et al, 2003; Coops et al, 2006; Wulder et al, 2006). Typically based on indices derived from tasselled cap transformations (TCT), these studies have been successful in producing both discrete (categorical) and probabilistic infestation maps. According to an aerial MPB outbreak survey carried out in 2005, infestation in the area of interest was light and generally limited to less than five trees at any given location. Because it is not feasible to identify a single or few red attacked trees within a 30 x 30 m pixel, the goal of this research was to demonstrate the type of products that may be generated for mapping MPB-related outbreak in later years. Data description

Figure 7: T1 (left) and T2 (right) Landsat-5 images. Note the location of inset maps for Figure 10.

104


Remotely sensed data Landsat data were acquired in August 2004 (T1) and September 2006 (T2) (path 46, row 22) (Figure 7). The images contained significant amounts of cloud and shadow. Both had undergone atmospheric correction and were provided as top of atmosphere (TOA) reflectance units. Mountain pine beetle attack and mitigation data Mountain pine beetle mitigation data acquired in 2005 (n=257) was used to characterize beetle-related disturbance across the landscape. The data consisted of points describing the number of attacked trees at a given location, and the number that were then cut and burned in order to mitigate the infestation. Helicopter-based global positioning system (heli-GPS) survey data mapping MPB attack across the area of interest was acquired for 2007 (n=48,382). Because forest inventory data were not available, the heli-GPS data were used to identify stands containing pine. Methods Generating an enhanced wetness difference index The Landsat-5 TM scenes were processed following the methods outlined in Wulder et al (2005). First, the two scenes were co-registered using a nearest neighbour, second-order polynomial transformation. The T2 image was normalized to T1 by using mean TOA reflectance values collected from bright and dark pseudo-invariant targets from each image. Wetness indices were calculated using a tasselled cap transformation (TCT); the coefficients applied to the individual bands are described in Table 5. An enhanced difference wetness index (EWDI) was then generated by subtracting T2 wetness from T1. Table 5: Coefficients used to generate wetness indices from Landsat 5 TM data (from Liang, 2004).

Landsat-5 TM Bands Band 1 Band 2 Band 3 Band 4 Band 5 Band 7

Coefficient 0.1446 0.1761 0.3322 0.3396 -0.6210 -0.4186

Processing mountain pine beetle mitigation data The MPB mitigation data collected in 2005 were used to characterize areas that were possibly or lightly disturbed. Points located in cloud or shadow were eliminated, leaving 168 for analysis. Enhanced difference wetness values were then extracted for each point. In order to characterize healthy stands for comparison with those undergoing infestation, a forest inventory-based pine mask is often utilized to locate candidate areas. Because one was not available for the area of interest, 2007 MPB attack data were employed. First, all points with more than one tree identified as attacked in 2007 were eliminated, as were those within 90m of a neighbouring point or road, leaving 2,179 points. The EWDI and T2 TCT greenness index values were then extracted. The greenness component of the TCT represents the abundance and vigour of vegetation cover; thus, to ensure that the 2007 data represented

105


healthy pine forest, the 1,000 points with the highest greenness values were extracted, and from those, 168 points were randomly selected to represent non-attack locations. Additional disturbance data In order to capture the range of disturbance across the area of interest, EWDI values were sampled from locations that had undergone a stand-replacing disturbance such as harvest (n=109), and sites where vegetation was in a state of regeneration (n=106). These points were selected by visually interpreting the T1 and T2 images. Thresholding EWDI values The EWDI distributions for the MPB mitigation, non-attacked, disturbed and regenerating samples were plotted, and thresholds were selected to separate the classes. A thematic map was then created by applying those thresholds to the EWDI image. Results Mountain pine beetle infestation in 2005 was light, with the majority of mitigation sites having less than ten attacked trees (Figure 8). The low intensity of MPB attack and mitigation resulted in very little separation between the EWDI values for the non-attack

Figure 8: Distribution of the number of stems removed at the MPB mitigation sites. Only five sites had more then ten stems eliminated. and mitigation points (Figure 9). Areas that have undergone recent stand replacing disturbance, or that were regenerating following one, showed greater separation, with higher and lower EWDI values than found in healthy stands, respectively (Figure 9). The lack of separation between the mitigation sites and healthy forest was a confounding factor for threshold determination.

106


Figure 9: Median, upper and lower quartiles, and non-outlier range of EWDI values for each class. The mitigation data, categorized as “possibly disturbed,” shows virtually no separation from the healthy forest sites, whilst areas of more severe disturbance are more clearly defined.

Figure 10 shows the T1 and T2 images, and a visualization of the EWDI. The EWDI visualization was produced by assigning the T1 wetness index to the red colour gun, and the T2 wetness index to the blue and green colour guns. Thus, red indicates a reduction in moisture between image dates, while blue indicates an increase. Finally, Figure 10 presents an example of a thematic map produced by thresholding the EWDI data. While some mitigation sites fall within the possible disturbance class, the majority are committed to the healthy forest category, as very little separation exists between the two EWDI distributions.

107


Figure 10: T1 and T2 Landsat-5 TM images, a visualization of the EWDI, and an example of a threshold-based classification with actual MPB mitigation data. The location of the maps is shown in Figure 7.

108


Discussion While the capacity to map MPB infestation using moderate resolution remote sensing data is well-established, success depends on the timing and severity of attack. For instance, Skakun et al (2003) found significant accuracy improvements for sites containing 30-50 red attack trees versus those containing 10-29. The data used in this study also suggests that our ability to discriminate red attack will improve as infestation rates increase. Figure 11 shows a graph of EWDI values grouped by the number of stems that were removed from mitigation sites in 2005, and by sites representing healthy forest (mitigation amount = 0). Although the medians are similar, the upper quartiles increases from approximately 14 for healthy forest, to 22 for sites where more than ten stems were removed. Figure 11 also presents generalized curves of the EWDI distributions for each disturbance class, and a hypothetical curve representing what may be expected as infestation amounts increase across the in the area of interest.

Figure 11: Mitigation amounts and associated EWDI values (top), and idealized EWDI distributions for the classes (bottom). Note the inclusion of the distribution expected for increased levels of red attack.

109


Literature Cited Adams, J.B., Sabol, D.E., Kapos, V., Almeida, R., Roberts, D.A., Smith, M.O., and Gillespie,

A.R. 1995. Classification of multispectral images based on fractions of endmembers - application to land-cover change in the Brazilian Amazon. Remote Sensing of Environment 52(2): 137-154.

Coops, N.C., Wulder, M.A. and White, J.C. 2006. Integrating remotely sensed and ancillary data sources to characterize a mountain pine beetle infestation. Remote Sensing of Environment 105:83-97.

Coppin, P. R., and Bauer, M. E. 1994. Processing of multitemporal Landsat TM imagery to optimize extraction of forest cover change features. IEEE Transactions on Geoscience and Remote Sensing 32 (4): 918-927.

Fraser, R. H., and Li, Z. 2002. Estimating fire-related parameters in boreal forest using SPOT VEGETATION. Remote Sensing of Environment 82 (1):95-110.

Kennedy, R.E., Cohen, W.B., and Schroeder, T.A. 2007. Trajectory-based change detection for automated characterization of forest disturbance dynamics. Remote Sensing of Environment 110 (3):370-386.

Liang, S. 2004. Estimation of land surface biophysical variables. In: Quantitative Remote Sensing Methods, edited by J.A. Kong. Hoboken: John Wiley & Sons:246-309.

Skakun, R.S., Wulder, M.A, and Franklin S.E. 2003. Sensitivity of the thematic mapper enhanced wetness difference index to detect mountain pine beetle red-attack damage. Remote Sensing Environment 86:33-443.

Wulder, M. A., White, J. C., Bentz, B., Alvarez, M. F., & Coops, N. C. 2006. Estimating the probability of mountain pine beetle red-attack damage. Remote Sensing of Environment 101 (2): 150-166.

Wulder, M.A., White, J.C., Coops, N.C., Han, T. and Alvarez, M.F. 2005. A protocol for detecting and mapping mountain pine beetle damage from a time-series of Landsat TM or ETM+ data. Version 1.0. Report prepared for the British Columbia Ministry of Forests and Range, October 31, 2005. Natural Resources Canada, Canadian Forest Service, Pacific Forestry Centre, Victoria, BC. 70 p.

110

Chapter 4: Food Model

CHAPTER 4: POTENTIAL AND REALIZED GRIZZLY BEAR HABITAT BASED ON FOOD RESOURCES

Scott Nielsen8

8 University of Saskatchewan, Canadian Cooperative Wildlife Health Centre

Introduction Grizzly bears are habitat generalists occupying numerous environments throughout the northern hemisphere (Pasitschniak-Arts 1993). The broad distribution of the species is partly due to a diet that is both diverse and omnivorous in nature providing plasticity in habitat requirements and adaptability to novel conditions. To date, habitat assessments and mapping for grizzly bears have largely relied on radiotelemetry data (Apps et al 2004, Nielsen et al 2004, 2006, Ciarniello et al 2007). Although radiotelemetry data is useful for informing habitat models, such information does not fully consider the importance of animal activity (bedding versus foraging) or fine scale temporal ‘switches’ in resources. Methods that explicitly target critical behaviours, such as foraging, or critical limiting factors, such as food resources, may be of greater value than naïve behaviour models that rely on indirect resource factors as predictors. Recently, Nielsen et al (unpublished) suggested that food-based habitat quality models be considered as an alternative to or complimentary to radiotelemetry-based resource selection functions (RSFs) for grizzly bears in Alberta. Given knowledge about local grizzly bear diet and the spatial distribution food resources, maps of individual food resources, as well as the overall quantity of food resources (habitat quality), are feasible. As well as more directly indexing habitat quality, food-based habitat mapping is much more cost effective than RSF-based approaches since capturing and collaring of animals is unnecessary. Here I describe the past development and recent application of food-based habitat models for west-central Alberta. Methods Distribution of individual food resources The distribution of 12 critical food resources (Table 1) was estimated using logistic regression for the region surrounding Robb, Alberta (Figure 1) based on 642 stratified (by landcover) random field plots (296 in un-harvested forests, 247 in clearcuts of various ages, and 99 open sites) and a suite of environmental GIS predictors. A more detailed description of field methods can be found in Nielsen et al (2004). Environmental predictors included landcover type (e.g., for forests: deciduous forest, mixed forest, closed conifer forest, etc.), climate (growing degree days, frost free period, mean annual temperature, mean annual precipitation, etc.), terrain (compound topographic index, topographic position, and solar radiation), and forest stand characteristics (age, canopy, leaf area index, and distance to edge) (Table 2). Univariate analyses were used to rank the importance of individual linear or non-linear factors and hypothesized interaction terms with the highest ranked uncorrelated factors introduced individually into a multivariate model until all significant factors were retained

111


(Hosmer and Lemeshow 2000). Model significance was assessed from a likelihood ratio χ2 test, while the receiver operating characteristics (ROC) area under the curve value (Swets 1988) was used to assess predictive accuracy. ROC values above 0.9 were considered to have high model accuracy, 0.7 to 0.9 good model accuracy, and <0.7 low model accuracy (Swets 1988, Manel et al 2001). Sensitivity and specificity values were used to predict the optimal cutoff probability for predicting species presence-absence by minimizing the absolute value of the difference between sensitivity and specificity curves (Liu et al 2005). Using model coefficients and associated GIS predictors, probability of occurrence was estimated for each species in a GIS. Probabilities were reclassified into binary presence-absence maps using the cut-off probability estimated from the sensitivity-specificity analysis. Diet-based weighting of food presence-absence maps To assign a seasonal contribution of each food resource, weights were applied to binary presence-absence maps for each of 10 bi-monthly periods beginning on 1 May and ending on 31 September. Weights were based on diet volumes of each food resource for the foothills and mountain environments during each of the 10 bi-monthly periods as reported in Munro et al (2006) and summarized in Table 3. Natural region boundaries (Natural Regions Committee 2006) were used to specify whether study pixels were in a mountainous or foothill environment. Following food weighting, food-habitat scores for each pixel and bi-monthly period were estimated as the sum of individual food-weights. As well as estimating bi-monthly habitat scores, multi-seasonal food-habitat scores were estimated as the sum of all bi-monthly food-habitat scores. Evaluating selection for food resource models Field-visited radiotelemetry locations from 9 sub-adult and adult female grizzly bears in west-central Alberta (Munro et al 2006) were used to evaluate the selection and performance of food resource models for predicting observed use of those foods. At each of 1,032 radiotelemetry locations, animal activity was classified as bedding, foraging (ants, berries, roots, and ungulate carcass), or no obvious sign. Using radiotelemetry locations where individual food resource use was recorded for ants, sweet vetch, cow parsnip, buffaloberry, and ungulates, probability of occurrence for each of the five food resources were contrasted against either radiotelemetry locations with no obvious sign (non-foraging or bedding sites) or 1,032 available locations randomly generated within the study area using logistic regression. A random effect on individual bear was used to account for differences in sample size among bears (Gillies et al 2006). Evaluating selection for food-habitat models

dividual food resource models, a larger zly

s

nd

asonal

Rather than assessing predictive capacity of inradiotelemetry dataset (29,255 radiotelemetry) from 31 female (sub-adult or adult) grizbears was used to assess seasonal selection of final food-habitat models. Here, no effort wamade to distinguish activity of animal and thus is an under-estimate of actual selection. Regardless, a positive relationship would be expected between radiotelemetry locations afood-habitat predictions, since bears should favor areas with higher food resource values andwould necessitate that some of the observations represented foraging activities. To determine whether food-habitat models predicted animal use (degree of selection), se

112


food-habitat values at radiotelemetry locations were compared to random available resources for each animal using seasonal (bi-monthly) resource selection functions (RSFs, Manly et al 2002). Available resources were defined for each animal’s minimum convex polygon (MCP) home range (design III, Thomas and Taylor 1990) using 1,000 random locations. To account for temporal variation in food-habitat values, all available locations were randomly assigned a bi-monthly period (bi-monthly period was already known for radiotelemetry locations). Given animal-specific estimates of selection, population-level significance was estimated using a one-sample t-test. Realized food-habitat values and a regional habitat loss index Because food-habitat models do not directly use radiotelemetry information or consider historic displacement of grizzly bear habitat by human activity, model estimates represent potential habitat (e.g., over-estimated in areas where grizzly bear habitat loss has occurred). To estimate realized habitat, I used a landscape-scale female grizzly bear occupancy model for Alberta from Nielsen (unpublished, Table 4, Figure 2) to ‘down-scale’ habitat values where occupancy was restricted. More specifically, realized habitat was estimated as the product of potential habitat and female range occupancy for each bi-monthly period. Having estimates of both realized and potential habitat, I estimated a regional habitat loss index based on the percent change in the realized habitat from potential habitat. Here I only report realized habitat and habitat loss for the multi-seasonal model. Both of these indices do not consider local, within range loss of habitat due to specific, localized human activities, such as road building, mines, etc. Regardless, models should prove useful for understanding broad-scale patterns of realized habitat value and habitat loss. Correlations between food-habitat models and RSFs Using seasonal RSF grizzly bear habitat models for the FRI core grizzly bear population unit (Nielsen, unpublished) and bi-monthly realized and potential habitat values for the same region, comparisons were made between RSF models and food-based habitat models for the mountain and foothill natural regions. The FRI core population unit was sampled with an intensity of one random location per 1-km2 to sample values of RSF and food-habitat predictions. Water, snow, ice, and barren areas in the alpine natural sub-region were excluded from the analysis, since they were masked out in RSF models and would inflate estimates of correlation coefficients. Pearson correlations were estimated for RSF and food-habitat models for each of the foothill and mountain regions. Relationship between food-habitat models and bear occupancy from DNA studies To estimate whether food-habitat models predict population occupancy, I used realized and potential habitat values surrounding DNA mark-recapture hair snag sites from the 2004 population survey south of Highway 16 and north of Highway 11 (Boulanger et al 2005) to predict occupancy of grizzly bears. Specifically, logistic regression was used with the presence or absence (detected/non-detected) of grizzly bears at bait sites used as the response variable and realized or potential habitat within a 900m buffer (largest size of buffer before adjacent sites began to overlapping) included separately as predictors. Zonal statistics of food-habitat values were used to estimate average food-habitat value for each bait site and values standardize (in standard deviation units) to assess strength of relationship.

113


Results Food resource distribution models Likelihood ratio model χ2 statistics and ROC values indicated that the majority of food resource distribution models were significant and having high predictive accuracy (Tables 5, 6 & 7). Non-linear relationships, particularly Gaussian responses, were common for climate variables (especially temperature related factors), canopy and canopy-related derivatives (i.e., leaf area index), and to a lesser extent terrain-induced variables, such as solar radiation. Maps predicting food occurrence illustrated that the distribution of food resources was highly variable, often being substantially different between mountain and foothill natural regions. For instance, ants were most prevalent in the foothills, while sweet vetch was more common in mountain environments. Important fruiting species, such as buffaloberry, were more common in the foothills, but also occurred extensively in valleys throughout the mountains. Food-habitat models Bi-monthly food-habitat values varied spatially and temporally reflecting wide-spread differences in the distribution of food resources and temporal changes in grizzly bear diets. Considering all bi-monthly periods together (multi-seasonal food-habitat index), food-habitat values were highest in low elevation mountain valleys and young forest stands in the upper foothills (Figure 3). Although some areas contained lower multi-season food-habitat scores, variability was less pronounced. Few if any areas contained all possible sources of food making the final multi-seasonal predictions for nearly all areas low compared to the theoretical high value of 1000 (e.g., maximum of 100 for each bi-monthly period x 10 bi-monthly periods). Evaluating selection for food resource models Compared to random locations, ants, sweet vetch, cow parsnip, and buffaloberry foraging radiotelemetry locations were 12.1 to 18.5 times more likely to occur in sites predicted to contain those food resources (Table 8). Ungulate carcass sites, however, were less well predicted by the ungulate habitat model with a non-significant odds ratio of 2.1. Comparing radiotelemetry foraging locations to radiotelemetry locations with no obvious sign, results were more variable, but patterns similar. Selection for ants and sweet vetch models occurred at lower odds than random locations with an odds ratio of 6.1 and 9.1 respectively (Table 8). Selection for buffaloberry and especially cow parsnip habitat was higher than at random locations at an odds ratio of 14 and 49.4 respectively. Finally, ungulate habitat was non-significant, although still positive with an odds ratio of 3.4. Evaluating selection for food-habitat models Bi-monthly food-habitat models predicted the occurrence of grizzly bear radiotelemetry observations best in mid-summer and more generally for bears residing in mountainous terrain. Seven of 10 bi-monthly habitat models were significant for mountainous regions, while 5 of 10 bi-monthly habitat models were significant in foothill regions. Habitat use in the mountains increased by a factor of 7 (odds ratio; exp(2)) for each one unit increase (values ranges from 0 to 100) in the July and early August food-habitat models, (Figure 4). In contrast, during the same period in the foothills, odds of habitat use ranged between 2 and 3. Poor predictive capacity was especially notable in the foothills during early May and early June and may reflect use of food sources not well described or the fact that activities other

114


than foraging (mating, extensive movements, etc.) often dominate during this time of year. Regardless, given the fact that radiotelemetry locations do not differentiate foraging locations from other activities, the general pattern of agreement between food-habitat models and animal locations suggests that food-habitat models within grizzly bear range can be a useful indicator of animal use. Realized food-habitat values and a regional habitat loss index Realized food-habitat predictions were highest in the mountains and the upper foothills (Figure 5). In contrast, realized food-habitat values were lowest in major mountain valleys (Montane natural sub-region) and lower foothills environments where bear occupancy restricted potential food-habitat values from being realized. The resulting difference reflected regional loss of habitat and is depicted in Figure 5. Losses are most substantial in large montane values of the mountains and especially lower foothill environments where bear occupancy has been reduced. Correlations between food-habitat models and RSFs Multi-seasonal food-habitat predictions were positively correlated with multi-seasonal RSF predictions, although the weakest correlation (r = 0.102) occurred with potential food-habitat prediction and RSF values in the foothills natural region. As potential food-habitat predictions did not consider current and historic displacement of grizzly bear range in the eastern parts of the study area, over-estimates of habitat resulted in a lower correlation. Once restrictions in the range of grizzly bears were considered with realized food-habitat predictions, correlations among RSF models and food-habitat models in the foothills were more substantial at 0.445. Differences among potential and realized food-habitat values and RSF models were less evident in the mountains with correlations estimated at 0.416 and 0.439. Relationship between food-habitat models and bear occupancy from DNA studies Grizzly bear occupancy was significantly predicted by average food-habitat scores surrounding bait sites (Table 9). Per one standard deviation increase in mean food-habitat value within a 900m radius of bait sites, grizzly bear occupancy increased by a factor of 2.4 (SE = 0.3) with the realized food-habitat model and 1.9 (SE=0.2) with the potential food-habitat model (Table 9). Although the relationship between the realized food-habitat model and bear occupancy was stronger than with the potential food-habitat model, differences were smaller than expected and are likely due to the fact that analyses were restricted to where bear occupancy was known (DNA survey locations used for this analysis were all within grizzly bear range). Larger spatial extents would likely increase differences in performance between realized and potential habitat models. Discussion Spatial predictions of critical grizzly bear food resources and subsequent food-based habitat maps show promise as an alternative to radiotelemetry-based habitat mapping. Using field-visited radiotelemetry observations for evaluation of food-resource models, models for ants, sweet vetch, cow parsnip, and buffaloberry were all positive and significantly associated with locations where foraging on those food items occurred. Although a positive relationship was evident for the ungulate carcass model, it was not significant suggesting that further

115


refinement is necessary. Selection for areas of higher food-habitat value−especially during mid-summer−was also evident for a larger set of radiotelemetry locations that did not consider activity of the animal. Despite differences in methods with RSF-based approaches, evaluation of models demonstrates positive correlations between RSF and food-habitat scores. Correlations with RSF values were highest for the realized food-habitat model and particularly for mountain environments. Potential food-habitat scores and RSF values in the foothills, although positively correlated, were weak demonstrating the need to consider large-scale restrictions in the range/occupancy of grizzly bears outside of mountain environments before using potential food-habitat models as predictors of habitat use. Food-habitat values were also significantly related to occupancy of grizzly bears at DNA mark-recapture bait sites. Although not described, additional positive relationships were evident for counts of unique grizzly bears at bait sites suggesting a relationship with population density. These results suggest that food-habitat models show promise in providing additional estimates of habitat and may more directly relate to habitat quality than RSF approaches that do not consider activity of animal. Some possible advantages of food-based habitat models over RSF-based approaches are: (1) maps are more likely to directly link to resources that influence reproduction and animal density; (2) animal activities, such as bedding and movement, are not emphasized as critical habitat like they are in RSF designs; (3) seasons are defined a priori based on diet; (4) temporal detail is retained and depicted at bi-monthly periods, which are likely to be necessary for showing phenological changes within an area and between the mountains and foothills; and (5) the ability to identify potential (i.e., assuming no human disturbance) versus realized habitat conditions. Development of food-habitat models for larger spatial extents are currently underway, given that this example concentrates on a region surrounding (within about 100km) Robb, Alberta.

116


Literature Cited Apps, C.D., B.N. McLellan, J.G. Woods, and M.F. Proctor. 2004. Estimating grizzly bear

distribution and abundance relative to habitat and human influence. Journal of Wildlife Management, 68:138–152

Boulanger, J., G. Stenhouse, M. Proctor, S. Himmer, D. Paetkau, and J. Cranston. 2005. 2004 Population inventory and density estimates for the Alberta 3B and 4B grizzly bear management area. Report prepared for Alberta Sustainable Resource Development, Fish and Wildlife Division, May 2005 (with updates November 2005). Edmonton. 28 pp.

Ciarniello, L.M., M.S. Boyce, D.C. Heard, and D.R. Seip. 2007. Components of grizzly bear habitat selection: density, habitats, roads, and mortality risk. Journal of Wildlife Management, 71:1446–1457.

Gillies, C.S., Hebblewhite, M., Nielsen, S.E., Krawchuk, M.A., Aldridge, C.L., Frair, J.L., Saher, D.J., Stevens, C.E., Jerde, C.L., 2006. Application of random effects to the study of resource selection by animals. Journal of Animal Ecology 75, 887-898.

Hosmer, D.W., & Lemeshow, S., 2000. Applied logistic regression. John Wiley & Sons, Inc., New York, New York.

Liu, C., Berry, P.M., Dawson, T.P. & Pearson, R.G. (2005) Selecting thresholds of occurrence in the prediction of species distributions. Ecography, 28, 385–393.

Manel, S., Williams, H.C. & Ormerod, S.J. (2001) Evaluating presence-absence models in ecology: the need to account for prevalence. Journal of Applied Ecology, 38, 921–931.

Manly, B.F.J., McDonald, L.L., Thomas, D.L., McDonald, T.L., Erickson, W.P., 2002. Resource selection by animals: Statistical design and analysis for field studies. 2nd ed. Kluwer Academic Publishers, Dordrecht, the Netherlands.

Munro, R.H.M., Nielsen, S.E., Price, M.H., Stenhouse, G.B., & Boyce, M.S., 2006. Seasonal and diel patterns of grizzly bear diet and activity in west-central Alberta. Journal of Mammalogy 87, 1112–1121.

Natural Regions Committee 2006. Natural Regions and Subregions of Alberta. Compiled by D.J. Downing and W.W. Pettapiece. Government of Alberta. Pub. No. T/852.

Nielsen, S.E., Boyce, M.S., and Stenhouse, G.B., 2004. Grizzly bears and forestry I: selection of clearcuts by grizzly bears in west-central Alberta, Canada. Forest Ecology and Management 199, 51–65.

Nielsen, S.E., Munro, R.H.M., Bainbridge, E.L., Stenhouse, G.B., Boyce, M.S. 2004b. Grizzly bears and forestry II: distribution of grizzly bear foods in clearcuts of west-central Alberta, Canada. Forest Ecology and Management, 199:67–82.

Nielsen, S.E., Boyce, M.S., Stenhouse, G.B., 2006. A habitat-based framework for grizzly bear conservation in Alberta. Biological Conservation, 130, 217-229.

Pasitschniak-Arts, M. 1993. Ursus arctos. Mammalian Species, 439, 1–10. Swets, J.A. (1988) Measuring the accuracy of diagnostic systems. Science, 240, 1285–1293. Thomas, D. & Taylor, E., 1990. Study designs and tests for comparing resource use and

availability. Journal of Wildlife Management 54, 322–330.

117


Table 1: Critical food resources used to define habitat in west-central Alberta, Canada. Food resource, abbreviation (code), feeding activity, and general season of use are described.

Food resource Abbrev.Feeding activity Season of use Ants ANTS myrmecophagy summer Equisetum spp. (horsetails) EQUI herbaceous spring Forbs FORB herbaceous spring − summer graminoids GRAS herbaceous spring − summer Hedysarum spp. (sweet vetch) HEDY root/tuber digging spring and fall Heracleum lanatum (cow parsnip) HELA herbaceous summer Shepherdia canadensis (buffaloberry) SHCA frugivory late summer − fall Trifolium spp. (clover) TRIF herbaceous spring − summer

ungulate carcass UNGL carnivorous/scavenging late spring − early summer

Vaccinium caespitosum (bilberry) VACA frugivory late summer and fall Vaccinium membranaceum (huckleberry) VAME frugivory late summer and fall Vaccinium myrtilloides (Cnd. blueberry) VAMY frugivory late summer and fall

118


Table 2: Description and characteristics of environmental variables used to model the probability of occurrence of individual grizzly bear food resources in west-central Alberta, Canada.

Variable group Variable name Abbrev. Res. Units Data range Landcover deciduous forest DFOR 30 m category 0 or 1 mixed forest MFOR 30 m category 0 or 1 open conifer forest OCON 30 m category 0 or 1 treed bog TBOG 30 m category 0 or 1 anthropogenic HUMN 30 m category 0 or 1 open bog OBOG 30 m category 0 or 1 shrub SHRB 30 m category 0 or 1 clear-cut CUT 30 m category 0 or 1 Stand characteristics forest canopy CNPY 30 m percent 1−100 forest age AGE 30 m 10 years 0−31.5 leaf area index (June 10) LAIa 30 m cm3/m2 0.62−6.01 leaf area index (August 13) LAIb 30 m cm3/m2 0.20−9.31 % change from LAIa to LAIb ΔLAI 30 m percent -93−328 distance to edge EDGE 30 m 100 m 0−14.1 Terrain compound topographic index CTI 30 m unitless 3.97−23.5 topographic position TOPO 30 m unitless -1075−242 solar radiation SOLR 30 m kJ/m2/1000 60.2−91.8 Climate annual moisture index AMI 500 m unitless 0.29−2.37 degree days-base 0 C DD0 500 m degree days 1134−1918 degree days-base 0 C DD5 500 m degree days 275−1179 frost free period FFP 500 m days 44−87 growing season precipitation GSP 500 m mm 308−499 mean annual precipitation MAP 500 m mm 483−944 mean annual temperature MAT 500 m °C -2.7−2.8 summer moisture index SMI 500 m unitless 0.57−3.71

119


Table 3: Weights used to score food-habitat values (weight × food item presence- 0 or 1) for bi-monthly periods based on reported diet volumes from Munro et al (2006). Final food-habitat values were standardized between 0 and 100 (maximum possible sum of weights in a bi-monthly period) by multiplying the bi-monthly sum by a scaling factor determined by dividing the sum into 100.

Season number and mid-point date

Food item 1

May 7 2

May 21 3

Jun 7 4

Jun 21 5

Jul 7 6

Jul 21 7

Aug 7 8

Aug 21 9

Sep 7 10

Sep 21

a. Foothills ANTS 0 0 0 1 2 3 1 3 0 2 EQUI 7 7 6 3 3 0 1 0 0 0 FORB 0 3 5 15 20 19 12 9 6 6 GRAS 2 12 28 37 45 28 24 10 18 18 HEDY 75 53 10 0 1 0 8 22 12 47 HELA 0 0 0 3 3 16 8 6 3 0 SHCA 0 0 0 1 0 14 25 28 12 17 TRIF 0 0 0 3 5 2 1 1 0 0 UNGL 5 15 37 23 14 4 5 3 2 4 VACA + VAMY 0 0 0 0 0 0 4 8 0 1 VAME 0 0 0 0 0 1 1 3 34 0

Sum = 89 90 86 86 93 87 90 93 87 95

b. Mountains ANTS 0 1 0 0 0 5 3 2 1 0 EQUI 0 0 3 1 10 6 2 2 1 0 FORB 0 2 1 11 30 16 14 6 3 0 GRAS 7 13 5 40 24 23 17 7 3 0 HEDY 90 55 53 34 10 19 21 27 3 3 HELA 0 0 0 0 11 23 25 0 0 0 SHCA 0 0 0 0 0 0 11 45 79 94 TRIF 0 0 0 0 6 0 0 0 0 0 UNGL 0 25 33 8 5 0 3 5 2 0 VACA + VAMY 0 0 0 0 0 0 0 0 0 0 VAME 0 0 0 0 0 0 0 0 0 0

Sum = 97 96 95 94 96 92 96 94 92 97

120


Table 4: Coefficients describing the probability of female grizzly bear occupancy (presence of a home range) for six population units in Alberta. All factors represent proportion within a 10-km radius window, except agriculture in the Livingstone and Waterton (phase 4) units where agriculture was measured as proportion within 1-km. Lower foothills were used as the reference category for natural sub-region. Natural sub-region categories with ‘avoid’ had perfect avoidance and were assumed a probability of 0.

95% CI Variable β S.E. p Lower Upper Boreal Forest Central Mixedwood -1.679 0.386 <0.001 -2.436 -0.922 Dry Mixedwood avoid NA NA NA NA

Grassland Foothills Fescue -0.226 0.673 0.737 -1.545 1.093 Mixed Grassland avoid NA NA NA NA

Foothills Upper Foothills 1.536 0.191 <0.001 1.163 1.910 Lower Foothills [ natural sub-region indicator category ]

Parkland Central Parkland avoid NA NA NA NA Foothills Parkland -0.099 0.672 0.882 -1.417 1.218 Peace River Parkland avoid NA NA NA NA

Rocky Mountains Alpine 5.208 0.722 <0.001 3.793 6.623 Montane 1.685 0.266 <0.001 1.164 2.206 Sub-alpine 4.399 0.402 <0.001 3.611 5.188 Agriculture -7.982 1.718 <0.001 -11.350 -4.614 Constant -0.865 0.121 <0.001 -1.103 -0.627

121


Table 5: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of 4 letter codes) in non-harvested forest stands near Hinton, Alberta.

Variables ANTS EQUI FORB GRAS HEDY HELA SHCA TRIF UNGL VACA VAME VAMYHabitat type§: decid. forest 0.963 -0.661 0.671 0.712 -0.245 -1.209 0.745 0.898 0.699 -0.847 -1.118 0.658 mixed forest 0.457 0.355 -0.378 1.831 0.156 -0.154 1.884 0.732 0.596 -0.821 -0.523 0.881 open conifer 1.074 -0.149 -0.365 0.881 1.361 -2.836 0.478 2.055 -0.025 0.156 -0.577 -0.012 treed bog 0.899 2.253 ref -1.232 1.103 none -1.772 0.646 0.030 -1.429 none none Climate: AMI -1.008 DD0 -0.073 DD0_2 x10k 0.235 FFP -0.145 5.735 2.172 FFP2 -0.041 -0.016 GSP -0.017 0.406 -0.409 0.046 GSP2 x 1k -0.551 0.465 MAT 0.77 MAT2 -0.297 MAP 0.112 0.421 MAP2 x 1k -0.084 -0.327 SMI -1.366 Stand characteristics: AGE -0.094 -0.076 0.536 0.153 AGE2 -0.032 -0.007 CANOPY -0.488 0.757 -0.352 0.818 DIST-EDGE -1.123 june10-LAI -3.369 -2.865 june10-LAI2 0.448 0.474 aug13-LAI 0.604 0.173 0.724 2.780 aug13-LAI2 -0.106 -0.259 change LAI -0.033 Terrain: CTI -0.323 0.297 -1.117 -0.486 CTI2 0.055 0.020 SOLAR -0.873 -11.288 -3.736 0.233 0.069 SOLAR2 0.0058 0.069 0.024 TOPO -0.005 TOPO2 Interactions: CTI x canopy 0.052 -0.093 -0.091 constant 34.88 -5.798 10.04 478.4 59.11 -58.86 -95.43 -74.00 93.13 -56.01 -28.31 -137.7Model evaluation: LR χ2 41.8 54.7 39.1 60.8 78.1 41.1 109.4 20.3 40.9 60.2 31.5 11.92 ROC 0.726 0.730 0.729 0.846 0.888 0.877 0.877 0.867 0.695 0.757 0.789 0.713 cutoff prob. 0.335 0.294 0.710 0.788 0.163 0.051 0.253 0.062 0.462 0.360 0.144 0.101 §closed conifer used as reference category in indicator contrasts of forest habitat types

122


Table 6: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of 4 letter codes) in harvested forest stands near Hinton, Alberta.

Variables ANTS EQUI FORB GRAS HEDY HELA SHCA TRIF UNGL VACA VAME VAMYClimate: AMI 10.592 -46.21 AMI2 13.79 DD0 -0.006 -0.293 DD02 x1k 0.112 DD5 0.096 0.008 DD52 x10k -0.493 FFP -0.115 -1.832 -2.142 -0.085 6.451 FFP2 0.013 0.016 -0.044 GSP -0.061 MAP 0.868 -0.506 1.359 MAP2 x 1k -0.707 0.415 -1.104 Stand characteristics: AGE 2.581 -1.289 0.214 AGE2 -0.487 0.140 CANOPY 0.210 june10-LAI 0.943 -0.573 13.775 june10-LAI2 -1.730 aug13-LAI 1.173 -0.367 aug13-LAI2 -0.123 change LAI 0.014 Terrain: CTI -12.672 0.382 0.209 1.200 1.469 CTI2 0.744 -0.062 -0.073 SOLAR 0.294 -0.140 TOPO 1x0 -0.200 0.045 -0.162 TOPO2 x1k 0.181 2.240constant 4.21 -45.13 73.14 -209.4 211.2 -28.85 44.89 149.8 4.57 -420.7 20.63 -250.6Model evaluation: LR χ2 48.0 48.0 19.1 16.3 62.3 14.0 39.0 19.5 23.0 46.5 19.6 92.7 ROC 0.738 0.754 0.682 0.872 0.938 0.718 0.872 0.675 0.692 0.719 0.788 0.905 cutoff prob. 0.661 0.415 0.699 0.950 0.108 0.026 0.110 0.138 0.341 0.473 0.076 0.225

123


Table 7: Estimated coefficients describing the occurrence of 12 grizzly bear food resources (see Table 1 for definitions of the 4 letter codes) for open habitats near Hinton, Alberta. Forbs occurred in every open habitat type (ubiquitous presence), while HELA, VAME, and VAMY where absent from all open sites.

Variables ANTS EQUI GRAS HEDY SHCA TRIF UNGL VACA Habitat type§: anthropogenic -1.291 0.842 all -2.322 2.814 0.056 0.366 1.082 open bog -0.977 1.171 -3.033 none 0.562 none 3.269 1.672 shrub -1.808 0.360 -1.264 0.130 none none 1.337 2.335 Climate: DD0 -0.017 DD5 0.002 GSP -1.327 GSP2 x 1k 1.475 MAT 0.559 MAP 0.590 MAP2 x 1k -0.410 SMI 12.32 25.88 -2.624 SMI2 -2.722 -5.81 Stand characteristics: DIST-EDGE -0.919 june10-LAI 3.987 june10-LAI2 -0.592 aug13-LAI -1.741 0.543 -3.138 aug13-LAI2 0.256 0.308 change LAI 0.028 Terrain influences: CTI 0.136 -1.034 -0.790 CTI2 0.044 0.031 SOLAR -2.975 -0.412 SOLAR2 0.020 TOPO 0.006 TOPO2 x 1k 0.076 constant 102.0 -3.962 0.699 2.724 60.62 -28.82 298.9 -213.7 Model evaluation: LR χ2 44.8 22.4 14.4 24.0 15.3 25.1 40.3 19.8 ROC 0.874 0.770 0.822 0.889 0.904 0.905 0.842 0.895 cutoff prob. 0.405 0.390 0.790 0.211 0.089 0.355 0.396 0.159

§herbaceous used as reference category in indicator contrasts of habitat types

124


Table 8: Evaluation of individual food probability models based on a comparison between animal locations visited in the field and food use identified (number of positive sample sites out of 1,032 indicated as n) against either random locations in the study area or sites without any bear sign and assumed to be locations where bears were travelling or foraging on other minor foods such graminoids. Coefficients, odds ratio (OR), standard errors of coefficients, and significance are reported.

foraging vs. random foraging vs. no sign Food resource n β (OR) SE p β (OR) SE p ants (ANTS) 204 2.92 (18.5) 0.30 <0.001 1.81 (6.1) 0.41 <0.001sweet vetch (HEDY) 253 2.51 (12.3) 0.87 0.004 2.21 (9.1) 1.00 0.027 cow parsnip (HELA) 25 2.89 (18.0) 1.33 0.031 3.90 (49.4) 1.72 0.024 buffaloberry (SHCA) 82 2.49 (12.1) 0.37 <0.001 2.64 (14.0) 0.58 <0.001ungulate carcass (UNGL) 51 0.76 (2.1) 0.61 0.210 1.21 (3.4) 1.05 0.250

Table 9: Estimated odds ratio (Odds) of grizzly bear occupancy at 705 DNA mark-recapture bait sites in the FRI core population unit (Boulanger et al 2005) as described by a 1 standard deviation change in potential or realized food-habitat values within a 900-m radius of bait sites. Realized food-habitat Potential food-habitat Variable Odds SE p Odds SE p mean food-habitat value† 2.379 0.298 <0.001 1.880 0.230 <0.001

† expressed as standardized (standard deviation) variables

125


Figure 1: Study area map depicting elevation, populated places, study area where field plots were visited, and the area modelled and tested.

126


Figure 2: Probability of female grizzly bear occupancy as a function of natural sub-region identity and agriculture. Population units are numbered (1-Waterton; 2-Livingstone; 3-Clearwater; 4-FRI core; 5-Grande Cache; 6-Swan Hills; 7-Northern Alberta).

127


Figure 3: Predicted multi-seasonal (1 May to 31 September) distribution of food resources (index from 0 to 1,000) in west-central Alberta, Canada. Enlarged area shown in the lower left of figure is centered on Maligne Lake in Jasper National Park.

128


* p < 0.05

May 7 May 21 Jun 7 Jun 21 Jul 7 Jul 21 Aug 7 Aug 21 Sep 7 Sep 21

β fo

od-h

abita

t

0.0

0.5

1.0

1.5

2.0

2.5Mountain Foothills ** p < 0.01

*** p < 0.001

***

*

***

*

******

***

*

*

**

*

***

Figure 4: Mean (±SE) habitat selection (β) of radio-collared female grizzly bears by natural region (mountain in black and foothills in gray) for food-based scores of habitat quality. Significance of one-sample t-tests (H0 = mean > 0, e.g., selection for the food-habitat model) indicated by star symbols for each season and natural region.

129

Chapter 4: RSF Modeling

Figure 5: Predicted patterns of realized food-habitat values (a.) and a regional habitat loss index (b.) (0- no loss to 1000- complete loss) based on the difference between realized and potential habitats.

130

Chapter 5: Graph Theory Movement Corridors

CHAPTER 5: GRAPH THEORY AND RSF GENERATED

MOVEMENT CORRIDORS FOR GRIZZLY BEARS IN ALBERTA, CANADA (PHASE 3 AND PHASE 4 RESULTS COMBINED)

Barbara L. Carra9

9 Wilfrid Laurier University

Introduction Maintaining connections for movement across fragmented landscapes is important for the long-term conservation of grizzly bear populations. As such, it is important to address movement within and across habitats as grizzly bears utilize the entire landscape and respond to gradients of habitat quality. In Alberta, the surrounding spatial environment facilitates or impedes movement between resource patches and therefore is considered a vital modeling component. The graph-theoretic model described here identifies resource selection function (RSF) preferred movement corridors between habitat patches across multiple spatial scales (home range to landscape levels) (Schwab 2003). This approach further provides a platform for measuring and comparing connectivity indices across heterogeneous landscapes characterized by varying levels of human development. Methods Study Area and Graph Creation The study area encompassed approximately 127,000 km2 of mountainous and foothills habitats situated along the eastern slopes of the Rockies ranging from Grande Cache in the north to the Montana, USA border in the south. Due to the size of the study region, graph theory analyses were completed by watershed unit and rejoined to create provincial scale maps. Major data inputs included a RSF grizzly bear habitat model defining patches (basis for nodes), GIS grid-based cost surfaces (basis for least-cost path corridor creation), and GPS data for habitat model development and validation. See previous reports for detailed descriptions on graph theory (GT) model development and implementation (e.g. Schwab 2004), while additional information regarding RSF-based habitat models can be found in Nielsen (2002). Patches were chosen based on pixels where the relative probability of adult female animal occurrence exceeded a threshold of +1.5 standard deviations. The centroids of each patch were used for node creation with specific patch attributes recorded for later use in graph analyses. Patches smaller than 5.0 ha were not selected as nodes but were maintained in graph analyses as suitable low-cost habitat within the cost surface and implicitly included in edge creation as stepping-stones (Bunn et al 2000). We used a habitat-based cost surface to define ‘edges’ or connections between patches which formed the underlying landscape graph structure (Schwab 2003). The original RSF habitat model, which was scaled between 1 and 10 to represent low to high ranking in relative

131


occurrence, was inversed to represent resistance to movement across the landscape. More specifically, low resistance (low movement cost) patches represented attractive bear habitats, while high resistance (high movement cost) patches represented low ranked or non-attractive habitats. We further re-sampled the cost surface to 500 m and least-cost path (LCP) connections or edges were generated to represent all possible connections between identified nodes or patches. Each LCP edge approximated the actual landscape distance traversed by a grizzly bear as it moved from one patch to the next in a heterogeneous landscape (Boone and Hunter 1996; Bunn et al 2000). Figures 1 and 2 demonstrate the RSF-based habitat patch layer and resulting node or centroid surface which provided the basis for LCP generation. Edges within the graph structure were further modified by a functional distance D. We explored the effect of distance on maintaining connections between patches or nodes. High distance thresholds allowed for all patches across the landscape to be connected. Conversely, low distance thresholds limited the number of connections existing between identified patches. Resulting graph structures were analyzed to quantify the subsequent change to connectivity resulting from changes in potential travel distance (see Figures 3 and 4 for examples). Detailed Graph Analyses Graph habitat and edge analyses were completed using FORTRAN modules independent of GIS to evaluate the importance of individual elements (edges and nodes) for the entire graph structure (Bunn et al 2000; Urban and Keitt 2001). Edge importance was evaluated using a technique termed edge thresholding (Bunn et al 2000). Edge thresholding is the iterative removal of LCP edges based on movement or potential travel distance. After each iteration, the resulting graph structures were reevaluated to assess responding levels of connectivity and identify where graphs begin to disconnect or fragment into subgraphs. Habitat importance was identified through node removal where the importance of each habitat patch was assessed to the overall graph structure (Bunn et al 2000). As nodes were removed, all edges incident to each node were also removed (Bunn et al 2000). The overall area-weighted dispersal flux (F) calculated the relative contribution of individual nodes to total landscape connectivity where si is the size (total area) of node i and pij the dispersal probability for nodes i and j. To determine F for each habitat patch, we calculated F for the entire landscape before and after node removal. As such, F indicates the strength or contribution of the removed habitat patch to graph or landscape structure. Figures 5 and 6 demonstrate patch importance to overall connectivity of the landscape. Results Analysis was completed across 32 watershed units to assess connectivity. Changes in functional edge distance resulted in large changes to landscape structures. Graph edges were defined by daily movement rate (Figure 3) and 95th percentile distance (Figure 4) thresholds. As distance thresholds decreased, the graph structure appeared less connected with an appearance of sub graphs occurring in the northeast. It is apparent from visual interpretation of the resulting graph structures above that fragmentation occurred first in the foothills. This

132


coincides with increased human disturbance features such as road structures and decreased numbers of large, contiguous habitat patches. Node removals resulted in different spatial patterns for minimum spanning tree (MST) patch importance and area-weighted patch dispersal flux (F). Results for MST node removals indicate smaller, high-quality habitat patches evenly distributed throughout the province to be important for maintaining the ‘backbone’ of landscape connectivity. The removal of these patches will reduce connectivity by increasing the presence of subgraphs. Currently, these patches are important stepping stones for movement within and across the landscape (Figure 5). Iterative node removal for area-weighted patch dispersal highlighted a linear portion of high-quality habitat patches integral to maintaining overall connectivity along the eastern slopes of the mountains (Figure 6). These patches are larger, contiguous habitat patches located along the transition zone between foothill and mountain environments and represent habitats needed to maintain populations outside national parks. Resulting graph edges were further assessed to understand the relative spatial configuration of movement patterns ranked by importance, probability for movement and overall level of connectivity. We created a grid-based version of corridor importance for maintaining landscape connectivity using the kernel density estimation (KDE) function in ArcGIS (Figures 7 and 8). All resulting high probability corridors were weighted by probability for movement as well as total number of paths. Figure 7 demonstrates finer scale movement paths across watersheds while Figure 8 highlights larger scale movement corridors across regions. Discussion Model developments for wildlife studies are increasingly combining the functional capabilities of remote sensing, GIS and GPS (Anderson and Danielson 1997; Roberts et al 2000). The graph-theoretic approach proposed here takes advantage of the power of combining spatial analysis abilities of GIS and GPS grizzly bear movement data allowing for the identification of spatial movement patterns in relation to habitat RSF models. Assessing grizzly bear movement and habitat use through GIS-based methods has the potential to assist resource managers with land-use decisions related to the conservation of grizzly bears (Dugas and Stenhouse 1999). Iterative removal of nodes demonstrated the affect of losing habitat patches to both spatial dispersal patterns and resulting connectivity rates. For grizzly bears, node removal techniques allow land use managers to envision the quantity of habitat loss acceptable to grizzly bear movements based on graph structure. Furthermore, the use of empirical habitat-based nodes representing high-quality habitat patches or food resources specific to grizzly bears (Nielsen et al 2002) strengthens the utility of our modeling results. We recommend that high-quality habitat patches be maintained to supply both resources to bears and safe movement, while medium and low-quality classed patches are reserved to provide stepping-stones (Figures 5 and 6). We further recommend that high quality movement paths be maintained for travel between resource patches at both the local and regional levels.

133


Figure 1: Phase 3 and Phase 4 study region demonstrating RSF-based habitat patches used in graph theory analyses.

134


Figure 2: Phase 3 and Phase 4 study region showing RSF patch centroids or nodes which provided the basis for least-cost path creation.

135


Figure 3: Phase 3 and Phase 4 least-cost movement paths further defined by the average daily movement distance threshold (4942 m).

136


Figure 4: Phase 3 and Phase 4 least-cost movement paths further defined by the 95th percentile distance threshold (6247 m).

137


Figure 5: Phase 3 and Phase 4 identified RSF-based habitat patches important to maintaining overall connectivity based on minimum spanning tree node removal iterations.

138


Figure 6: Phase 3 and Phase 4 identified RSF-based habitat patches important to maintaining overall connectivity according to area-weighted node removal iterations.

139


Figure 7: Phase 3 and Phase 4 fine-scale graph theory generated corridors important to maintaining local connectivity (raster format). Dark represents high importance combined with higher path quantities and light represents less importance and lower path quantities available for travel.

140


Figure 8: Phase 3 and Phase 4 large-scale graph theory generated corridors important to maintaining regional connectivity (raster format). Dark represents high importance combined with higher path quantities and light represents less importance and lower path quantities available for travel.

141


Literature Cited Anderson, G. S., and B. J. Danielson. 1997. The effects of landscape composition and

physiognomy on metapopulation size: the role of corridors. Landscape Ecology 12:261-271.

Boone, R. B., and M. L. Hunter. 1996. Using diffusion models to simulate the effects of land use on grizzly bear dispersal in the Rocky Mountains. Landscape Ecology 11:51-64.

Bunn, A. G., D. L. Urban, and T. H. Keitt. 2000. Landscape connectivity: A conservation application of graph theory. Journal of Environmental Management 59:265-278.

Dugas, J., and G. B. Stenhouse. 1999. Grizzly bear management: validating existing cumulative effects models. Pages 157-160 in Proceedings of Thirteenth Annual Conferences on Geographic Information Systems: 157-160.

Roberts, S. A., G. B. Hall, and P. H. Calamai. 2000. Analysing forest fragmentation using spatial autocorrelation, graphs and GIS. International Journal of Geographical Information Science 14:185-204.

Schwab, B.L. 2004. Graph Theoretic Methods for Examining Landscape Connectivity and Spatial Movement Patterns: An Update, in Stenhouse, G.B., Munro R.M, and K. Graham (eds). Foothills Research Institute Grizzly Bear Research Program 2003 Annual Report. Hinton, Alberta, 87 pp.

Schwab, B. L. 2003. Graph theoretic methods for examining landscape connectivity and spatial movement patterns: application to the FRI Grizzly Bear Program Masters of Science Thesis, University of Calgary, Calgary.

Nielsen, S. E., M. S. Boyce, G. B. Stenhouse, and R. H. M. Munro. 2002. Modeling grizzly bear habitats in the Yellowhead Ecosystem of Alberta: taking autocorrelation seriously. Ursus 13:45-56.

Urban, D., and T. Keitt. 2001. Landscape connectivity: a graph-theoretic perspective. Ecology 82:1205-1218.

142

Chapter 6: Wildlife Health

CHAPTER 6: WILDLIFE HEALTH

Dr.Cattet8, Dr.Vijayan10, and Dr.Janz11 8Canadian Cooperative Wildlife Health Centre; 10University of Waterloo; 11University of Saskatchewan Introduction Accounts of associations between large-scale, human-caused environmental change and wildlife population declines make newsworthy items, recent examples including reports in the popular press on global warming and polar bear (Ursus maritimus) population declines (see www.washingtonpost.com/wp-dyn/content/article/2006 /12/26/ AR2006122601034.html), or descriptions of pollution, habitat loss, and mass localized extinction of amphibian populations (see www.sciencedaily.com/releases/2006/07/ 060707094220.htm). Specific to this study, recognition in Alberta of possible links between habitat change as a result of human land use activities and declining grizzly bear (U. arctos) populations is growing (Alberta Grizzly Bear Recovery Team, 2005). A fundamental problem with these and other examples, however, is the knowledge and tools needed to establish causal relationships are often lacking (Figure 1) (Stevenson et al, 2005). An added impediment is considerable time can lapse between occurrences of environmental change and detection of affected populations (Findlay and Bourdages, 2000). It follows, without understanding causal mechanisms, we cannot predict with accuracy effects of environmental change on wildlife populations before they occur (Wikelski and Cooke, 2006).

143


Although populations are usually a major focus of wildlife monitoring and conservation strategies, the responses of individual animals to human alteration of the environment is what causes or contributes to poor performance of resident wildlife populations (Figure 2) (Wikelski and Cooke, 2006). Wild animals are exposed to many “stressors” throughout their life, but often cope successfully through a suite of physiological and behavioral mechanisms, collectively known as the “stress response” (Reeder and Kramer, 2005). However, the duration of a stressor is important. Whereas short-term stress (lasting seconds to hours) encountered during the normal activities and experiences of daily life rarely pose a threat to healthy animals, long-term stress (lasting days to months) as can occur with human alteration of the environment, can exceed an animal’s ability to cope (Moberg and Mench, 2000). When “stressed” for weeks to months, an animal loses its capacity to sustain normal biological function (i.e., growth, reproduction, immunity, activity) and gradually develops signs of impaired health (termed “distress”) including reduced growth, impotency, infection, and sometimes premature death. Whether or not population-level effects occur depends on the proportion of the population that is distressed. If the proportion is small, population performance (i.e., reproductive output, survival rates, abundance) should not change, but if it increases, a reduction in performance is more likely to occur (WWF International Arctic Programme and WWF-DetoX, 2006). Because the time lapse between human alteration of the environment and reduction in wildlife population performance can take many years (Findlay and Bourdages, 2000), broadening the focus of wildlife monitoring and conservation strategies to include assessment of long-term stress and biological function in individual animals provides opportunity to alleviate environmental stressors before population performance is affected. Additionally, this approach can: ♦ Enable more sensitive evaluation of the efficacy of conservation actions than is provided

solely by measuring population-level parameters; ♦ Provide knowledge that is currently lacking to establish cause and effect between large-

scale, human-caused environmental change and wildlife population declines; and ♦ Provide baseline background data needed to (i) quantify the impact of human land use

activities, and to (ii) develop predictive models quantifying the response of wildlife populations in altered environments to understand how future problems can best be avoided.

144


Wildlife Health Research Objectives 1. Develop and validate serum-based indicators of long-term stress in grizzly bears. 2. Develop and validate a tissue-based protein array to detect long-term stress in grizzly

bears. 3. Develop health profiles for grizzly bears and their populations. 4. Evaluate the relationship between long-term stress and health in grizzly bears.

145


PROGRESS IN STRESS RESEARCH Efforts on this front pertain to research objectives 1 and 2. Development and Validation of Serum-Based Indicators of Long-Term Stress We evaluated the usefulness of serum cortisol and heat shock proteins (hsp) 60 and 70 as indicators of long-term stress experienced by grizzly bears. This was accomplished by measuring serum concentrations in bears collected over the past 9 years (between spring 1999 and 2007; approximately 140 bears). Serum cortisol levels were elevated by capture stress and, therefore, did not appear to be a good measure of stressors experienced by the animal prior to capture. Serum hsp60 levels did not show any significant differences among biological factors, such as age, sex, and reproductive status, or among regions of grizzly bear distribution in Alberta (see Figure 5). However, hsp70 levels differed between regional groupings of bears suggesting this measure shows some potential as an indicator of long-term stress. To confirm the usefulness of hsp70, we are continuing to evaluate it in grizzly bears in relation to other health and landscape measures, while also investigating its dynamics in polar bears (in conjunction with the Ontario Ministry of Natural Resources’ Climate Change Program) in relation to seasonal availability of sea-ice in southern Hudson Bay. In addition to measuring more conventional indicators of stress, we are also currently identifying and developing other serum indicators of stress in grizzly bears. One candidate molecule is the corticosteroid-binding globulin (CBG), a protein carrier to which cortisol is bound in circulation, which is shown to change in response to long-term stress but not to short-term stress in animals. Our preliminary studies attempted to purify this protein from bear serum. While we did purify the protein, the CBG concentration was not sufficient enough to generate antibodies for this protein. We did make polyclonal antibodies by injecting the purified CBG from bear into rabbits. However, the antibody was not specific to CBG, likely due to contamination by other proteins in the purified CBG fraction. We did not have sufficient proteins to repeat the injections. Consequently, we decided to undertake a biotechnology approach to make recombinant bear CBG. Currently, bear CBG mRNA has been partially sequenced using primers designed from dog CBG. When this is complete, the full-length bear CBG cDNA will be cloned into an expression vector for production of protein in recombinant bacteria. Purified recombinant CBG will then be used to produce polyclonal antibodies for developing techniques, including enzyme-linked immunosorbent assay (ELISA) for measuring CBG concentration in bear serum. As another aspect of our research, we are using proteomic technology to discover novel proteins that may be indicative of the stress state of the animal. For this, we are comparing animals with “low” and “high” stress scores (see Development of Grizzly Bear Health Profiles) to identify differentially expressed proteins in bear serum. We compared four independent groups of animals and we have identified 15 serum proteins that were significantly different in the bears with the high stress scores relative to the low stress scores. These proteins spots have been excised and sent for mass spectrometry at the University of Western Ontario proteomics facility, to identify proteins of interest. We are currently awaiting the results. In the mean time, we are also comparing the impact of repeated sampling on serum protein levels to identify markers of long-term stress. The procedure is

146


exactly the same as mentioned above but the comparison will involve serum of the same bear between first and subsequent captures. Development and Validation of Tissue-Based Indicators of Long-Term Stress We have developed an antibody-based protein microarray (“bear stress chip”) to simultaneously determine expression of 31 stress-associated proteins in skin and muscle biopsy samples collected from grizzly bears (Table 1). We are currently completing the initial determinations of these stress proteins in 158 grizzly bear skin samples obtained from the Foothills Research Institute Grizzly Bear Program (FRIGBP), and expect to complete this milestone by March 2008. To date we have determined the stress profiles for 30 bears and are conducting a preliminary data analysis to compare stress protein expression patterns (1) among bears of different sex, age class and body condition, (2) with serum stress indicators determined in the laboratory of Dr. Vijayan, and (3) among bears inhabiting different locations. As described in the previous interim report, our progress in 2007 involved completing the identification of commercially available antibodies for their cross-reactivity with bear stress proteins (i.e., the 31 proteins mentioned above), optimizing the processing of tissue samples, and optimizing the labeling of proteins in samples with fluorescent dyes. Once these milestones were achieved, prototype chips were printed with antibodies for initial validation and optimization experiments (e.g., different antibody dilutions, protein dilutions and blocking/washing procedures). Following this, the final chips were printed with antibodies by an independent company (FirstPhase Technologies, Tempe, AZ).

147


Table 1: Stress-associated proteins detected by the “bear stress chip”, a long-term stress detection tool developed for use in the conservation of grizzly bears. The stress proteins are arranged in general categories associated with their role in the stress response, although it should be noted that there is considerable overlap and integration among these categories.

Category of Stress Stress Protein 1. Hypothalamic-pituitary-adrenal axis Adrenocorticotropic hormone (ACTH)

Glucocorticoid receptor Corticotropin-releasing factor receptor 1/2 Proopiomelanocortin (POMC) Prolactin Arginine vasopressin receptor V1a

2. Oxidative stress and inflammation Superoxide dismutase-1 Superoxide dismutase-2 Peroxiredoxin-3 Chemokine receptor-5 Inducible nitric oxide synthase (iNOS) Endothelial nitric oxide synthase (eNOS) Heme oxygenase-2 Cyclooxygenase-2

3. Cellular stress and proteotoxicity Heat shock protein-27 (HSP27) HSP40 HSP60 HSP70 HSP70 (inducible) HSP90 HSP110 Glucose-regulated protein-78 (GRP78/BIP)

4. Apoptosis and mitosis Apoptosis-inducing factor (AIF) Annexin II Annexin IV Caspase 1 Caspase 2 Caspase 3 Caspase 6 E-cadherin GAPDH

148


PROGRESS IN WILDLIFE HEALTH RESEARCH Efforts on this front pertain to research objectives 3 and 4. Development of Grizzly Bear Health Profiles With approximately 110 variables relating to grizzly bear health, it is a challenge to summarize the health status of bears without conducting detailed and complex statistical analyses (See Chapter 6). This difficulty is further compounded by observations that: ♦ The grizzly bear health database contains records from repeated captures of individual

bears, i.e., repeated measures effects. ♦ Some health variables are influenced by method of capture. ♦ Some health variables are influenced by sex and age of bear. ♦ Some health variables are influenced by date of capture. ♦ Not all health records are complete, i.e., missing values. To overcome these difficulties, we developed a health function score system to enable quick summary of health profiles for individual bears, to identify bears with impaired health, and to explore associations between stress and health and between health and environment at a coarse level. The system is based on five health functions – age, growth, immunity, movement, and stress. We also intended to include reproduction as a function, but lacked sufficient data to calculate a meaningful score for many bears. We calculated scores for four health functions based on raw values for 2-7 constituent variables (Table 2), selected on the basis of biological knowledge and data reduction methods (e.g., principal components analysis), which collectively best represented a specific health function. We calculated health function scores based on health data collected from 164 grizzly bears during 280 captures occurring throughout grizzly bear range in Alberta from 1999 to 2007. The scores represent averages of percentile values for constituent variables ranging in value between 0.00 and 1.00. Where values for constituent variables differed between sexes and / or between capture methods, percentile values were calculated within categorical classes before averaging. We did not calculate a score for age of bear, but instead used age in years. Table 2: Constituent variables used to calculate health function scores for grizzly bears captured for the Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007).

Health Function Scores Constituent Variables

Stress score Serum concentrations of gamma-glutamyltransferse (GGT), total cortisol, heat-shock proteins 60 & 70, and glucose

Growth score Body mass, straight-line body length, axillary girth, body condition index (BCI), and serum concentration of alkaline phosphatase

Immunity score White blood cell count, lymphocyte count, proportion of neutrophils and monocytes, serum concentrations of total protein and globulin, and serum albumin-to-globulin ratio

Movement score Mean daily movement rate during breeding season (May 16 – July 31) and mean daily movement rate during all other times of year

149


The transformation of constituent variable values into percentile scores normalized data distributions and resulted in mean scores close to 0.50 (Table 3). Although 95% confidence intervals were narrow for health function scores, minimum and maximum values were broader (stress score = 0.22-0.76, growth score = 0.13-0.87, immunity score = 0.20-0.71, movement score = 0.04-0.97). Male grizzly bears had higher mean scores than females for growth and immunity, but were similar in other scores as well as age. Although Table 1 represents data collected from individual bears at first capture only, these patterns were also similar when analyzing data collected from all captures. Table 3: Descriptive statisticsA for health function scores and ages of grizzly bearsB captured for the Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007).

Health Functions Female bearsC Male bearsC All bears

Stress score 0.47 (0.43-0.50) [55] 0.49 (0.46-0.52) [68] 0.48 (0.46-0.50) [123]

Growth score 0.42** (0.38-0.47) [54] 0.47** (0.42-0.52) [52] 0.45 (0.41-0.48) [106]

Immunity score 0.46* (0.40-0.48) [53] 0.50* (0.48-0.53) [72] 0.48 (0.47-0.50) [125]

Movement score 0.56 (0.46-0.67) [21] 0.48 (0.38-0.59) [19] 0.52 (0.45-0.60) [40]

Age (years) 7.2 (5.8-8.5) [67] 6.2 (5.1-7.2) [80] 6.6 (5.8-7.4) [147] A Presented as mean, 95% confidence interval in round brackets, and sample size in square brackets. B Score and age statistics calculated from data collected at first capture only. C Differences between sexes indicated by ‘*’ for P ≤ 0.05 and ‘**’ for P ≤ 0.01. Individual bears captured multiple times throughout the project duration showed considerable variation in health function scores from one year to the next, but nonetheless displayed changes that paralleled multi-year patterns observed at the population level (Figure 3). For example, an adult male grizzly bear G017 had higher stress, growth, and immunity scores than most bears captured in the Foothills Research Institute (FRI) core study area (see map on Figure 5) from 1999 to 2003. However, the direction and magnitude of change in G017’s scores from year-to-year were similar to the pattern of population differences between years. This suggests that factors, presumably environmental in nature, influence health at multiple levels (individual and population) simultaneously.

150


Figure 3: Change in health function scores of four adult grizzly bears captured multiple times in the core study area of the Foothills Research Institute Grizzly Bear Program from 1999 to 2003. The green-shaded area represents the 95% confidence interval for the population mean in each year. Movement score values were too few to depict annual changes for individual bears.

Many of the health function scores shown in Figure 3 are based on data collected during sequential captures of four individual bears with an interval of approximately one year between captures. An exception, however, is the stress and immunity scores for G029 measured over a 2-week interval in 2001. These are noteworthy given the large degree of change – a 10% increase in stress score and a 12% decrease in immunity score – over a short span of time. Although we did not determine specifically the cause of these marked changes, we have identified short- and long-term effects of capture and handling over the course of the FRIGBP that had not previously been recognized (Cattet et al 2003, Cattet et al 2008).

151


Mean health function scores for the population of bears inhabiting the FRI core study area showed significant variation from year-to-year (Figure 4). Trends in stress and immunity scores were similar – lowest in 2001 and increasing over the next two years. Although we didn’t find statistically significant differences between years in mean growth scores, there was a consistent pattern of increasing mean scores from 1999 to 2003. We have not determined the factors influencing these changes, but this again serves to support our notion that health is dynamic at the population level, as well as the individual level. A next and obvious step in these analyses will be to evaluate associations between annual population mean health scores and annual measures of population performance, i.e., demographic rates.

Figure 4: Change in mean population health function scores for grizzly bears captured in the core study area of the Foothills Research Institute Grizzly Bear Program from 1999 to 2003. Mean scores and 95% confidence intervals were adjusted for a 7-year old bear and calculated from data collected at first capture only. Statistically significant differences (P ≤ 0.05) indicated by letters where a > b > c. Error bars with ‘ab’ or ‘bc’ above are intermediate in value to, but do not differ significantly from ‘a’ and ‘b’, or ‘b’ and ‘c’. Annual sample sizes are provided in round brackets below error bars.

152


We did not have adequate sample sizes based on data collected at first capture only to compare mean health scores among grizzly bear populations in Alberta. Nevertheless, we did conduct a regional analysis based on 3 regions (Figure 5) with the following population composition: (i) North of Highway 16 [N16] – populations 5, 9, 11, 12; (ii) FRI core study area [FRI] – population 1; and (iii) South of Highway 11 [S11] – populations 2, 3, and 4. We decided not to include data collected from bears in the Swan Hills area (population 6) because the sample size was too small, and because landscape conditions (natural and human-caused) in this area are distinctly different from other areas included in N16. We found significant differences between regions which imply health of grizzly bears is not uniform across their range in Alberta. Bears inhabiting areas north of Highway 16 (N16) have the highest growth scores, but also have the highest stress and immunity scores in the province. In contrast, bears inhabiting the FRI core study area (FRI) have the lowest growth and immunity scores, and bears inhabiting areas south of Highway 11 (S11) have the lowest stress scores. Our ongoing detailed statistical analyses of grizzly bear health and landscape structure / change data will help to reveal the basis for these differences. Associations between Age, Stress, and Other Health Function Scores Stress and growth scores of grizzly bears were directly associated with age tending to be higher in older bears and lower in younger bears (Figures 6a-b). We also found evidence for these patterns in individual bears captured multiple times suggesting stress scores (Repeated measures ANOVA – P = 0.031, n = 74) and growth scores (P = 0.098, n = 74) increased as individuals aged. Stress scores were significantly associated with all other health function scores (Figures 6a-d). Our finding of a positive association between stress and growth is counterintuitive as it suggests bears exhibiting the greatest growth were also the most stressed. However, some insight is offered by observations that bears with greatest growth also inhabit areas where road density is greatest whereas bears with lowest growth inhabit areas of higher elevation where human access is less (See Chapter 6). We suggest this pattern may be due to availability of better food resources in disturbed areas (as reflected by road density) where bears are more likely to be stressed by human activities and/or landscape conditions, than in higher elevation areas where human activity is less, but so too is food quality and availability.

153


Figure 5: Mean health function scores for grizzly bears that inhabit different regions of Alberta and were captured for the Foothills Research Institute Grizzly Bear Program (1999-2007). Regions on a north-south gradient are N16 – between Highways 2 and 16, FRI – core study area between Highways 16 and 11, and S11 – between Highway 11 and Montana border. Mean scores and 95% confidence intervals are adjusted for a 6-year old bear and were calculated from data collected at first capture only. Statistically significant differences (P ≤ 0.05) indicated by letters where a > b > c. Error bars with ‘ab’ above are intermediate in value to, but do not differ significantly from ‘a’ and ‘b’. Sample sizes are provided in round brackets below error bars. We found a positive association between stress and immunity scores in grizzly bears (Figure 6c). Long-term stress is believed to compromise immune function in animals (Schwab et al 2005) which suggests grizzly bears with high stress scores should also have low immunity scores, i.e., an inverse association. However, to date most of the constituent variables used in the grizzly bear stress score reflect short-term (acute) stress instead of long-term (chronic) stress. Dhabhar (2000) found acute stress to enhance immune function in mice while chronic stress had a suppressive effect. This could also explain the pattern observed in grizzly bears. We will continue to investigate the association between stress and immunity scores as we expand our suite of long-term stress biomarkers, e.g., cortisol binding globulin, stress-associated proteins in skin and muscle.

154


Figure 6: Associations between age, stress score, and other health function scores in grizzly bears captured for the Foothills Research Institute Grizzly Bear Program in western Alberta (1999-2007). Scatter-plots represent correlations between a) age and stress score, b) age, growth and stress scores, c) stress and immunity scores, and d) stress and movement scores calculated from data collected at first capture only. We found a weak inverse correlation between stress and movement scores (Figure 6d) suggesting bears with higher stress levels moved less. Again, as with the association between stress and growth, this may tie into where “more stressed” vs. “less stressed” bears live and relative differences in habitat characteristics, or more specifically food availability. If “more” stressed bears occupy lower elevation areas with greater food availability, they may move shorter distances and less frequently to find food when compared to “less stressed” bears living at higher elevations (See Chapter 6). At this point, we would not suggest these associations identified between stress and other health functions provide evidence that stress causes alterations in health. Nonetheless, these findings do corroborate results emerging from our more detailed statistical analyses of grizzly bear health and landscape structure / change data.

155


Literature Cited Alberta Grizzly Bear Recovery Team. 2005. Draft Alberta grizzly bear recovery plan.

Alberta Sustainable Resource Development, Fish and Wildlife Division, Alberta Species at Risk Recovery Plan No. 15. Edmonton, AB. 134 pp.

Cattet, M., J. Boulanger, G. Stenhouse, R.A. Powell, and M.J. Reynolds-Hogland. 2008. An evaluation of long-term capture effects in ursids: implications for wildlife welfare and research. Journal of Mammalogy (In review).

Cattet M., K. Christison, N.A. Caulkett, and G.B. Stenhouse. 2003. Physiologic responses of grizzly bears to different methods of capture. Journal of Wildlife Diseases 39:649–654.

Dhabhar, F.S. 2000. Acute stress enhances while chronic stress suppresses skin immunity. The role of stress hormones and leukocyte trafficking. Annals of the New York Academy of Sciences 917:876-893.

Findlay, C.S., and J. Bourdages. 2000. Response time of wetland biodiversity to road construction on adjacent lands. Conservation Biology 14: 86-94.

Moberg, G.P., and J.A. Mench. 2000. The biology of animal stress. CABI Publishing, New York, USA. 396 pp.

Reeder, D.M., and K.M. Kramer. 2005. Stress in free-ranging mammals: integrating physiology, ecology, and natural history. Journal of Mammalogy 86: 225-235.

Schwab, C.L., R. Fan, Q. Zheng, L.P. Myers, P. Herbert, and S.B. Pruett. 2005. Modeling and predicting stress-induced immunosuppression in mice using blood parameters. Toxicological Sciences 83:101-113.

Stevenson, R.D., S.R. Tuberty, P.L. DeFur, and J.C. Wingfield. 2005. Ecophysiology and conservation: the contribution of endocrinology and immunology – introduction to the symposium. Integrative and Comparative Biology 45: 1-3.

Wikelski, M., and S.J. Cooke. 2006. Conservation physiology. Trends in Ecology and Evolution 21: 38-46.

WWF International Arctic Programme and WWF-DetoX. 2006. Killing them softly – health effects in arctic wildlife linked to chemical exposures. World Wildlife Fund. 29 pp.

156

Chapter 7: GIS Progress Report

CHAPTER 7: GEOGRAPHIC INFORMATION SYSTEM (GIS)

PROGRESS REPORT

Jerome Cranston3

3 Arctos Ecological Services, Hinton, AB Introduction The objective of the GIS component of this study was to summarize the environmental conditions encountered by each grizzly bear prior to its capture, including measures of anthropogenic landscape change. Understanding the relationships between grizzly bear health and environmental conditions requires not only quantifying landscape attributes over the specific areas experienced by a grizzly bear, but modeling them as they existed at a particular time. Given the rapid changes in grizzly bear habitat caused by industrial activity, and the long period of time over which health data has been collected (1999-2006), it was essential to match environmental conditions to a particular period of interest. For this analysis we modeled landscape conditions for a 55,000 km2 portion of grizzly bear range over an eight-year period to correspond to grizzly bear health and activity profiles from radio collared grizzly bears (Figure 1). We selected two Grizzly Bear Population Units where there is genetic evidence to indicate that these units (BMA’s) represent two genetically distinct population units. These BMA’s were selected as they represented two populations where we had the most comprehensive and long term data set on health and movement characteristics. Four sets of environmental condition summaries have been performed.

157


Figure 1: Area and time frame for the annual landscape conditions maps created for the health analysis. Methods and Results Analysis 1: BMA 3 and 4 The first summary, in September 2006, focused on 58 bears living in Bear Management Areas (BMAs) 3 and 4, between Highway 1 to the south and Highway 16 to the north (Figure 2), a sample representing over half the total grizzly bear population in this area, as estimated by DNA mark-recapture surveys in 2004 and 2005. These 58 bears had been subject to a total of 155 capture events over 8 years (1999-2006), from which health data had been collected. The statistical analysis was based on a subset of the available health data. However, the focus was not toward the results per se, but instead toward developing the appropriate methodologies to merge and analyze this complex data set. This preliminary exploration allowed us to identify and address potential obstacles (e.g., repeated measures, confounding variables, data redundancy), and to plan an efficient strategy to conduct data mergers and analyze the complete data set.

158


Figure 2: Bear Management Areas (BMA) 3 and 4 and the bear location data available for this area. The health metrics from these samples were matched to environmental conditions experienced by the animal prior to its capture. The portion of the landscape occupied by a bear is called its home range, defined as “…that area traversed by the individual in its normal activities of food gathering, mating, and caring for young.” (Burt, 1943). Using ArcView 3.2 Animal Movement extension, we generated 128 annual 95% kernel home ranges from over 55,000 GPS locations collected from these bears. The 95% kernel is a contour on a point density surface within which a point has a 95% probability of occurrence (Figure 3).

159


Figure 3: The 95% kernel home range for G020 in 2002 using a single smoothing factor. Of the 155 captures, over half (81) were the bear’s second or subsequent capture and had GPS data preceding the capture event. The home range polygons generated from this data defined the precise portion of the landscape experienced by the bear prior to capture. The balance of captures (74) were initial captures for which the bear did not have GPS data prior to capture, so a home range polygon generated from GPS points collected after the capture was used to represent its approximate home range. Given the high degree of home range fidelity shown by grizzlies from one year to the next, this was considered a valid assumption. Each capture event was associated with a home range polygon based on GPS data preceding the capture if available, or subsequent to capture if not. Each polygon was then assigned a condition year. The condition year is the point in time at which Landsat imagery was acquired and the time in which landscape conditions were considered to have the greatest influence on a bear’s health parameters. For spring captures, condition year was the year prior to capture; for fall captures (after mid-July), the condition year was the year of capture. For each condition year a GIS layer was created to represent the corresponding configuration of forest and land cover classes, and major anthropogenic features such as roads, well sites, and cut blocks, as it existed at the

160


time. Since Landsat imagery was acquired in late summer or early fall, the landscape conditions would be about six months out of date for spring captures, but most of the change occurring during that time would have occurred during denning season and would have had less influence on the bear’s level of stress. Amount of annual anthropogenic change within a home range was measured between the time of image acquisition in the year prior to capture, to the time of image acquisition in the year during the capture. For spring captures, some of this change may have occurred in the summer months after the capture. Of the 128 annual kernel home ranges generated for all bears and all years of GPS data, summaries were performed on 89. (Some polygons were used for multiple captures). For each of the following three sets of environmental variables, summary statistics (mean, range, standard deviation, etc) were calculated for every home range polygon using ArcView 3.2 Zonal Statistics (Spatial Analyst extension). 1) terrain-based or static variables: (These variables are unchanging and independent of time of capture)

• Elevation: From 25m-resolution Digital Elevation Model (DEM) • Terrain ruggedness: raster surface (30m resolution) derived from DEM • Slope position: raster surface (30m resolution) derived from DEM • Solar radiation: raster surface (30m resolution) derived from DEM • Protected status: the proportion of a bear’s home range within a park or

protected area. 2) vegetation-based/habitat variables:

• Forest age: raster surface (30m resolution), derived from provincial AVI (Alberta Vegetation Inventory) polygons

• Crown closure: density of forest canopy; raster surface (30m resolution), a Remote Sensing product developed by the Foothills Facility for Remote Sensing and GIScience.

• Resource selection Function (RSF), an indicator of grizzly bear resource availability: raster surface (30m resolution), derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience.

• Mortality Risk, an indicator of grizzly bear habitat security: raster surface (30m resolution), derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience.

(Note: these variables were based on 2003 conditions, therefore did not always match the condition year).

3) Anthropogenic features:

• Density and number of roads (line features converted to 30m pixels; total pixel count, and number of pixels per unit area of home range)

161


• Density and number of well sites (total number, and number per unit area of home range)

• Density and number of cut blocks (total area, and per unit area of home range) • number and density of roads built in previous year • number and density of well sites built in previous year • number and density of cut blocks built in previous year

These anthropogenic features were created for each condition year (1998 – 2005) using attribute data for cut block and well site features. Roads and other linear features were classified using imagery and were assumed to precede the well site or cut block feature to which it led. Analysis 2: Phase 6 Mapping Extent In June 2007, a second GIS summary was performed on 95 kernel home ranges that represented the balance of the 184 kernel home ranges that had been generated for every bear and every year since 1999 (Figure 4). This updated dataset included health data from 2007 captures but GPS home ranges corresponding to these captures was not available as a full season of GPS data had not yet been collected from these bears.

Figure 4: Kernel home ranges created for the second analysis.

162


Variables: 1) Static variables:

• Elevation: from 25m-resolution Digital Elevation Model (DEM). • Terrain Ruggedness Index: raster surface (30m resolution) derived from DEM. • Compound Topographic Index: raster surface (30m resolution) derived from DEM ;

a measure of soil wetness. • Solar radiation: raster surface (30m resolution) derived from DEM. • Percent Protected: the proportion of a bear’s home range within a park or protected

area.

2) Anthropogenic variables: • Well site density: raster surface (30m resolution), representing the density of well

sites. • Road density: raster surface (30m resolution), negative exponential decay distance to

roads (value = 1 next to road, dropping toward 0 further away). • Seismic/trail density raster surface (30m resolution), negative exponential decay

distance to trails (value = 1 next to trail, dropping toward 0 further away). 3) Vegetation/habitat variables:

• Crown closure: density of forest canopy; raster surface (30m resolution), a Remote Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; 2005 conditions.

• Species composition: (percent conifer in forest stands; raster surface (30m resolution), a Remote Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; 2005 conditions.

• Forest age: raster surface (30m resolution), derived from provincial AVI (Alberta Vegetation Inventory) polygons; 2005 conditions.

• Resource selection Function (RSF), an indicator of grizzly bear resource availability: raster surface (30m resolution); average of three seasonal models; derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience.(2005 conditions).

• Mortality Risk, an indicator of grizzly bear habitat security: raster surface (30m resolution), derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience. (2005 conditions).

• Attractive sink: raster surface (30m resolution); combination of RSF surface and mortality risk surface, values proportional to probability of bear occurrence and risk of human-caused mortality; (2005 conditions).

• Safe harbor: raster surface (30m resolution); inverse of attractive sink surface; combination of RSF surface and mortality risk surface, values proportional to probability of bear occurrence and inversely proportional to risk of human-caused mortality; 2005 conditions.

One challenge encountered in merging this set of GIS summaries with the first was that many of the GIS variables from the first summary did not cover the larger area; and some of the variables for the complete area had not been available at the time of the initial analysis.

163


Analysis 3: All captures For the third analysis, in November 2007, a more precise correspondence between condition year and anthropogenic disturbance was required. Using Landsat TM7 imagery acquired in the late summer or early fall of each year, a manual classification of 23,350 cut block polygons, 7615 well sites, and 62,576 linear access features was performed for an area of 38,705 km2 in BMA 3 (1998-2005) and 17,182 km2 in BMA 4 (2002-2005). Each point (well site), line (linear access structure, such as road, pipeline, power line, or railway) and polygon (cut block) feature was assigned a year based on the year of imagery in which it first became visible. Figure 5 shows how the anthropogenic features have changed from 2000 to 2004.

2000 2001 2002 2003 2004

Figure 5: An example of the change in anthropogenic features for a small area of BMA 3 from 2000 to 2004. Of the total of 272 capture events, 225 had associated home ranges and environmental summary statistics. 142 unique home range polygons were used for summaries. The following variables were summarized: 1) Static variables:


a measure of soil wetness. • Percent Protected within 10km: raster surface (30m resolution), proportion of park or

protected area within 10km radius kernel. 2) Anthropogenic variables:

• Well site density: raster surface (30m resolution), kernel density of oil and gas well sites within 10 km radius.

• LAF density: raster surface (30m resolution), kernel density of linear access features within 1 km radius.

• MeanRegen: Proportion of home range in regeneration (or agriculture; agriculture only applies to 2005 polygons), calculated using Focal Statistics (Spatial Analyst) and a binary regeneration raster.

• NewFeatures_Mean: Proportion of home range occupied by new cut blocks, roads, and well sites, built between fall of year prior to capture, and fall of capture year.

164


3) Vegetation/habitat variables: • Crown closure: density of forest canopy; raster surface (30m resolution), a Remote

Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; 2005 conditions.

• Species composition (percent conifer) : (percent conifer in forest stands; raster surface (30m resolution), a Remote Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; 2005 conditions.


• Resource selection Function (RSF), an indicator of grizzly bear resource availability: raster surface (30m resolution); average of three seasonal models; derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience.(2005 conditions).

• Mortality Risk, an indicator of grizzly bear habitat security: raster surface (30m resolution), derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience. (2005 conditions).

• Attractive sink: raster surface (30m resolution); combination of RSF surface and mortality risk surface, values proportional to probability of bear occurrence and risk of human-caused mortality; 2005 conditions.

• Safe harbor: raster surface (30m resolution); inverse of attractive sink surface; combination of RSF surface and mortality risk surface, values proportional to probability of bear occurrence and inversely proportional to risk of human-caused mortality; 2005 conditions.

Analysis 4: All captures, annual habitat conditions For this extraction, a complete set of land cover, forest canopy, species composition, and grizzly bear habitat models were developed for each condition year (1998-2005) within BMA 3 and 4. For the first time, the condition year for all captures in the dataset could be matched to landscape and habitat conditions for the same year, with the exception of captures from fall 2006 and 2007, which were matched to conditions for 2005. Also, two new environmental variables were summarized: the footprint of anthropogenic change from one condition year to the next, and average movement rates from GPS data. The remote sensing rasters developed by the Foothills Facility for Remote Sensing and GIScience (land cover type, crown closure, and percent conifer) within BMA 3 and 4 were backcast to conditions for 1998 and later, with vector features for anthropogenic disturbance (roads, well sites, cut blocks, and fires) “burned in” to the land cover to capture features which may have been too small for the Landsat sensor to classify. These layers, along with terrain layers formed the basis of the grizzly bear habitat models describing grizzly bear occurrence (RSF), risk of human-caused mortality (Risk), and a combination of the two (Safe Harbor). Geoprocessing scripts in the Python language were written to derive other predictor variables, such as proximity to edges, from the terrain and remote sensing layers, and to regenerate the Mortality Risk, seasonal RSF, and Safe Harbor models for each year. The home range kernel type was changed from 95.0% to 99.999% as it was determined that patch-level metric calculations were less sensitive to edge effects with this configuration.

165


There were 228 capture events for which associated home ranges were summarized, using 146 home ranges. 1) Static variables:


a measure of soil wetness. • Percent Protected within 10km: raster surface (30m resolution), proportion of park or

protected area within 10km radius kernel. 2) Anthropogenic variables:

• Well site density: raster surface (30m resolution), kernel density of oil and gas well sites within 10 km radius.

• LAF density: raster surface (30m resolution), kernel density of linear access features within 1 km radius.

• MeanRegen: Proportion of home range in regeneration (or agriculture; agriculture only applies to 2005 polygons), calculated using Focal Statistics (Spatial Analyst) and a binary regeneration raster.

• MeanChange: Proportion of home range within the footprint of anthropogenic disturbance, calculated using Focal Statistics (Spatial Analyst) and a binary annual change raster.

3) Vegetation/habitat variables: • Crown closure: density of forest canopy; raster surface (30m resolution), a Remote

Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; matched to Condition year.

• Species composition (percent conifer) : (percent conifer in forest stands; raster surface (30m resolution), a Remote Sensing product developed by the Foothills Facility for Remote Sensing and GIScience; matched to Condition year.


• Resource Selection Function (RSF), an indicator of grizzly bear resource availability: raster surface (30m resolution); average of three seasonal models; derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience; matched to Condition year.

• Mortality Risk, an indicator of grizzly bear habitat security: raster surface (30m resolution), derived from Remote Sensing products developed by the Foothills Facility for Remote Sensing and GIScience. matched to Condition year.

• Safe harbor: raster surface (30m resolution); inverse of attractive sink surface; combination of RSF surface and mortality risk surface, values proportional to probability of bear occurrence and inversely proportional to risk of human-caused mortality; matched to Condition year.

References Burt, W.H. 1943. Territoriality and home range concepts as applied to mammals. Journal of

Mammalogy 24:346-352.

166

Chapter 8: Analysis of GIS, Landscape Matrices and Health Variables

CHAPTER 8: PROGRESS REPORT ON ANALYSIS OF GIS,

LANDSCAPE MATRICES AND HEALTH VARIABLES

John Boulanger12

12Integrated Ecological Research, Nelson, British Columbia

Introduction The main objective of this section is to provide a preliminary finding of some of the current analyses on the relationship between grizzly bear environmental variables and health scores. In this report I use constrained ordination methods, which are robust to many of the issues with health-environmental data, to further explore this data set. This report should be considered in unison with previous analyses (Boulanger 2006) that explored variable inter-relationships, sample size issues, and mixed-model based analysis. Methods The analysis of the relationship between landscape and health data is challenging for many reasons. First, the health data is based on repeated live capture of grizzly bears. It is known that live capture affects many of the health variable measurements and therefore the effect of live capture has to be modeled if the health measurement is not robust to capture effects. At the time of the analysis, micro-array, or CBG stress measures that are robust to live capture effects were not available. Second, it is likely that repeated measures of bears are not statistically independent and therefore first captures can only be used for analyses or statistical models robust to repeated measures need to be used. Third, it is likely that many health indicators vary by age and sex of bear. Therefore, the data set contains both categorical (sex, capture type) and continuous data. As a result data may need to be stratified by categorical variables, or statistical methods that allow both categorical and continuous variables need to be used for data analysis. All the above issues are further compounded by small sample sizes which also reflect the population status (N) within the various management units. Variables Yearly home ranges were estimated for each of the GPS collared bears using the kernel home range estimator (Worton 1989). GIS and landscape variables were summarized for each home range area. Health records from capture were matched to home range areas as detailed in Chapter 4 above. At the time of report preparation, yearly remote-sensing based GIS maps were not available for all the areas sampled (areas outside BMA’s 3 and 4) and therefore in some cases landscape conditions for bears were based upon the 2003 base map. Future analyses (final program report) will utilize the newly completed yearly base maps for all bears in the analysis. As an initial step, all data were assessed in terms of normality and potential outliers were identified. Percentage data was converted into a proportion and then transformed with an arcsine square root transformation to help meet the assumption of normality.

167


GIS and Landscape variables Table 1 describes the landscape variables used in the analyses. The majority of variables were based upon a 2003 base map of the study area. Therefore, change in these variables would be due to home range area shifts of bears as opposed to landscape change. Only the main anthropogenic variables (i.e. percent roads) were updated on a yearly basis. Table 1: Landscape variables considered in analyses. Variables marked with an * were updated for each year so yearly change in these variables reflected both landscape and bear home range shift for a portion of the study area. Other variable values were based upon a 2003 base map. Any yearly change in these variables would be due to home range shift of bears.

GIS Description solar_mean solar radiation Dem_mean Mean elevation topcls_mean Topographic class tri_mean terrain Ruggedness Index canopy10x_mean Crown Closure (0-10) pctcon_mean Percent conifer (0-1) age05adj10x Forest age (10-year increments) Pct_Protected Percent home range protected PctCut* Percent of each home range that was cut block (new or old). Numwells* Number of well sites PctRds* Percentage of home range covered in roads RSF_mean Mean RSF score Risk_mean Risk score for home range

Landscape level metrics were also considered (Table 2). All of these were based upon the 2003 base map and therefore changes in variables were due to changes in bear home range rather than landscape change. Table 2: Landscape level metrics considered in the analysis.

Metric Full name L1_Land_AREA_MN Mean patch size L2_Land_ECON_MN Contrast Edge density L3_Land_SHAPE_MN Shape index L4_Land_AREA_CV Variation in mean patch size L5_Land_PROX_AM Area-weighted proximity L6_Land_ENN_CV Variation in interpatch distances L7extra_Land_ENN_MN Mean interpatch distance

Health variables Health variables were originally grouped by Marc Cattet and David Janz in terms of their relationship to various health components (Table 3). I further grouped health variables into general components. Variables that did not directly correspond to a health component were not considered in the analysis.

168


Table 3: Health variables considered in the analysis. Italics variables were suggested after the analysis was conducted and will be considered in future analyses.

No Health component Variables Covary with 1 Growth/Condition BCI age/sex

alk_phos_UpL straight-line length age/sex

2 Reproduction progesterone_ngml age/sex estrogen_pgml age/sex prolactin_ngml age/sex luteinizing_ngml age/sex testosterone_ngml age/sex 3 Immunity wbc_e9pL neu_percent lym_percent lym_neu_ratio mono_percent capture eos_percent capture

baso_percent total protein globulin AG_ratio

capture capture capture capture

4 Stress total_cortisol_UW capture ggt_UpL capture glucose_mmolpL capture hsp_60_uw hsp are sample size hsp_70_uw limited

Most of the health variables were reduced in sample size when only the first capture of a bear was used for the analysis (Table 4). This was a key limiting factor in analyses given that the overall number of observations used in any type of analysis is based upon the variable with the lowest sample size. Sample sizes were higher when multiple capture events were used. Analysis strategies Previous analyses (Boulanger 2006) focused on eliminating redundant variables from each of the respective health and environmental variables using principal component analysis (McGarigal et al 2000). Mixed models were then used to test for potential relationships between single health variables and suites of environmental variables. For the next phase of this analysis I focused on constrained ordination methods that simultaneously tested the relationship between suites of health variables and environmental variables. I used program CANOCO for constrained ordination analyses (Jongman et al 1995, Ter Braak and Smilauer 2002, Leps and Smilauer 2003). In particular, redundancy analysis was used to assess the relationship between suites of health response variables and environmental predictors. The significance of environmental predictors was assessed using Monte Carlo tests for parameter significance (Ter Braak and Smilauer 2002). When applicable, covariates were used to condition out nuisance variables in the analysis such as the differential effect of snare or helidarting on some of the stress variables. Program CANOCO

169


also allowed the modeling of repeated measures of bear captures through the use of randomization tests that were robust to correlations between repeated measures (Ter Braak and Smilauer 2002). The main analysis strategy can be conceptualized by Figure 1 where each box forms an analysis node with respective indicator variables. Also included are capture effects which is a nuisance variable. I initially focused the analysis on the relationship between bear biology and demography and environmental variables. In particular I was interested if bears were segregated across the landscape as indicated by environmental variables. Using this information I further explored the relationship between bear health, biology/demography and the environment. I focused on the spatial comparison of bears as opposed to temporal change aspects of the data set. The next phase of the analysis will consider temporal change once the landscape and GIS data maps are temporally linked to yearly estimated bear home ranges.

Bear health andphysiologystress, immunity etc

Bear biology anddemographyAge-sex

Environmental changetemporal/spatial (GIS)

Capture effects

Figure 1: Conceptual diagram of health-environmental analysis. Results Sample sizes Table 4 summarizes sample sizes of captured bears. The actual sample size used for analyses was lower (about 20-30%) due to missing values for some of the health and/or environmental variables.

170


Table 4: Sample sizes of health variables.

Data set Sexclass Capture type helidart snare Total Individual Female 9 43 52 Female/cubs 9 12 21 Male 15 68 83 Total 33 123 156Repeated Female 27 51 78 Female/cubs 24 20 44 Male 30 78 108 Total 81 149 230

Exploration of the relationship between bear demography and GIS/Landscape conditions Constrained ordination (redundancy analysis) was conducted to determine if age and sex class could be predicted by the GIS and landscape conditions in a bear’s home range area. The GIS variables (Table 1) were used as predictor variables for this analysis. The amount of total variation in age-sex class explained by the ordination was 11.2% with 78% of the variation explained by the first 2 canonical axes; however, Monte Carlo tests suggested that most of the GIS variables were significant predictors of bear age-sex class (Figure 2).

-1.0 1.0

-1.0

1.0

AM

AF

AFC

SM

SF

Linear Access Features

WELLS

Canopy closure

Elevation

RISKRSFSUM

SAFEHARBPercentConifer

Percent protected

Health variables

Environmental variables

Bears (Sex-age class)AM AF AFC SM SF

Figure 2: Results of constrained ordination of bear age-sex class with GIS variables as predictor variables. The first 2 canonical axes are shown. All GIS variables that were underlined were significant predictors of bear age-sex class.

171


The ordination diagram reveals potential gradients and relationships in the data set. In review, the constrained ordination creates component axes that maximize the separation between age-sex classes. For each bear a score is produced that is a linear function of the axes. A procedure that is similar to linear regression is then conducted where the score is regressed with each GIS variable as a predictor. The resulting standardized scores from the regression of the GIS variables form the canonical axes that are displayed in Figure 2. The length of arrows for the age-sex and GIS variables is proportional to the strength of the relationship of these variables in the ordination. A value close to the origin implies minimal strength. The angle of arrows within predictor variables, and between predictor GIS variable and age-sex class reveals the correlation of these variables. Variables whose arrows are at a 90 degree angle have no correlation. Variables that are opposed (180 degree separation) have negative correlation and variables that are in the same direction are positively correlated. Finally, predictor variables that are underlined were considered significant predictors (at α=0.1). Using these rules, it is suggested that a gradient of linear access features (roads) and percent conifer is evident in the data set. Subadult males are positively correlated with linear access features (roads) and negatively correlated with percent conifer. Adult females are positively correlated with percent conifer and negatively correlated with linear access features. Another gradient in the data set is canopy closure and percent protected area within a home range. Adult males are positively correlated with canopy closure and negatively correlated with percent protected whereas adult females with cubs (AFC) are positively correlated with percent protected and negatively correlated with canopy closure. Standardized scores for each bear are also shown as subdivided by age-sex class. General clumping of these scores around each of the age-sex arrows can be shown further supporting the general direction of the age-sex arrows in the analysis. To further explore results I plotted the standardized means for each of the age-sex classes as a function of the most significant predictor variables (Figure 3). From this, it can be seen that adult females with cubs have the highest percent of protected area in their home range areas but lowest canopy closure. Subadult males have lowest percent conifer and highest linear access features. Adult males have highest degree of canopy closure. These results suggest that subadult classes are found in more open roaded areas whereas adult classes are found in higher protected areas (females with cubs) or higher percent conifer areas (adult females). Canopy closure is more difficult to interpret since high canopy closure areas are found in both foothills areas as well as mountainous valleys (Figure 4).

172


Percent protected

Stan

dard

ized

mea

n-0.3-0.2-0.10.00.10.20.30.40.50.60.70.8

sexclassAF AFC AM SF SM

Percent conifer

Stan

dard

ized

mea

n

-0.7-0.6-0.5-0.4-0.3-0.2-0.10.00.10.20.30.4


Linear access features (roads)

Stan

dard

ized

mea

n

-0.4

-0.3

-0.2

-0.1

0.0

0.1

0.2

0.3

0.4

0.5


Canopy closure

Stan

dard

ized

mea

n

-0.7-0.6

-0.5-0.4

-0.3-0.2

-0.10.0

0.10.2

0.3


Figure 3: Standardized mean scores for significant GIS predictor variables from ordination analysis (Figure 2).

173


Figure 4: The distribution of canopy closure values across the study area. It can be seen that high canopy closure areas are found both in foothills area and mountain valleys whereas low canopy closure is primarily in alpine areas. Another ordination analysis was conducted with landscape metric variables and GIS variables as predictors of age and sex class (Figure 5). The ordination diagram from this analysis suggested that all 5 landscape metrics were associated with age-sex classes. It also suggested that some metrics and GIS variables were correlated, such as elevation and mean patch size (L1), percent protected and contrast edge density (L2). Relationships between GIS and landscape variables were further explored by a constrained ordination to determine if landscape variables can predict GIS variables (Figure 6).

174


-1.0 0.8

-0.8

1.0

AM

AF

AFC

SM

SF

Linear access features

CC

DEMRISK

RSF

L1_area

L2_Econ

L3_shape

L4_CVL5_Proximity

%conifer

protected

Figure 5: Constrained ordination biplot showing the relationships between age-sex class and predictor GIS and landscape variables. Underlined GIS/Landscape variables were significant predictors of age-sex class.

-2.0 3.0

-2.0

2.0

CUTSMEAN

LAF WELLS

AGE

ATRSINK

CC

DEM

RISK

RSFSUM

SAFEHARB

TRI_MEAN

pctcon_t

protected

L1_area

L2_Econ

L3_shape

L4_CV

L5_Proxi

SAMPLES

AM AF AFC SM SF

Figure 6: Constrained ordination biplot showing the relationships between landscape and GIS variables. Underlined landscape variables were significant predictors of age-sex class.

175


Results of this analysis suggest that mean patch size (L1), contrast edge density (L2) and shape index (L3) are most related to various GIS variables. These results suggest that in some cases landscape variables are related to GIS variables. The question then becomes whether bears are responding to changes in landscape structure or GIS based habitat types. This question will be explored further in future analyses that consider temporal change in bear landscape and GIS variables. Relationship between environmental variables and health scores Condition scores Constrained ordination analysis was conducted using the GIS/Landscape variables as predictors of the condition scores. For this analysis, age, sex class, age*sex class and year were used as covariates. In doing this, the principal effects of these variables were conditioned out of the analysis to allow a clearer analysis of the relationship between GIS/Landscape variables and condition indices.

-0.8 0.6

-0.6

0.6

BCI

SLL

alkphos

risk

rsf

cc

L1area

L2econ

L3shapeL4cv L5prox

pctconpctpro

laf

dem

Figure 7: Constrained ordination biplot showing the relationships GIS/landscape variables and condition indices. Underlined GIS/landscape variables were significant predictors of condition indices. Results revealed linear access features, canopy closure, edge contrast (L2) and elevation as significant predictors. A gradient composed of canopy closure/linear access features as opposed by edge contrast/elevation was revealed by the biplot (Figure 7). Straight line length was aligned along this gradient. A weaker non-significant gradient of RSF versus area-weighted proximity (L5) and shape index (L3) was also suggested with BCI and alk-phos aligned along this gradient. Plots of the data suggested that straight-line length was influenced by elevation even when age-sex class was accounted for (Figure 8).

176


sexclass AF AFC AM SF SM

SLL

120

130

140

150

160

170

180

190

200

Elevation

800 1000 1200 1400 1600 1800 2000 2200 2400

Figure 8: Plot of straight line length (SLL) versus elevation as a function of age and sex class. A general decreasing trend in SLL with increasing elevation can be seen for each age-sex class. A plot of elevation, linear access features and straight line length reveals the relative distribution of these variables that form the predicted SLL gradient (from the ordination analysis) (Figure 9). It can be seen that areas of high road density are in lower elevations and vice versa. Furthermore, a trend for larger SLL bears is suggested with higher road density for males. For females this is also suggested, however, there are some larger adult females clustered in areas of 0 LAF. Further review of the data shows that these are older females. This general relationship could be explained by lower mortality pressure in areas of 0 LAF leading to a larger proportion of older females (with high SLL values).

177


0.00000.0007

0.00140.0021

LAF (Roads) 841

1350

1859

2368

Elevation126.00

142.67

159.33

176.00

SLL

Females

0.00000.0008

0.00160.0024

LAF (Roads) 730

1217

1705

2192

Elevation139

160

181

202

SLL

Males

Figure 9: Plots of straight line length versus LAF (linear access features/roads) and elevation as a function of sex. Blue points represent subadult bears, red points represent adult bears and a star represents females with cubs. Plot of body condition index and RSF score also suggest a weak relationship between these two variables with BCI score increasing with increasing RSF (Figure 10).

sexclass AF AFC AM SF SM

BCI

-3

-2

-1

0

1

2

3

Mean RSF score

7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Figure 10: Body condition index versus mean RSF score as a function of age and sex class.

178


Stress scores (from data sets currently available) I conducted a constrained ordination to explore the effect of GIS/landscape variables on stress indices. For this analysis, capture type (heli-dart or snare) was used as a covariate to allow bear stress levels to be different dependent on type of capture. I did not include age and sex as covariates but did include them in triplots to see if any clustering of age-sex class around environmental predictors was apparent. Results suggested that canopy closure (cc) and contrast edge density (L2) were significant predictors of stress levels. Examination of triplots suggested a gradient with canopy closure opposing contrast edge density. GGT and hsp60 had the closest orientation to this gradient (Figure 11).

AF AFC AM SF SM-0.8 0.8

-0.6

1.0

glucose

ggt

cortisol

hsp70

hsp60

LAFCC

DEM

RISK

RSFSUM

L1_Area

L2_Econ

L3_Shape

L4_CV

pctcon

protected

Figure 11: Constrained ordination triplot with canonical scores for bears shown as a function of age and sex class. Significant GIS/landscape variables are underlined.

Plots of the raw data suggest high GGT and high HSP60 scores in areas of highest canopy closure and highest contrast edge density (Figure 12). The trend is most noteworthy for adult and subadult males for canopy closure. For contrast edge density the trend is most apparent for females with cubs as well as subadult males.

179


37.0445.86

54.6863.50

Contrast_edge 10.00

26.33

42.67

59.00

Canopy closure4.00

35.33

66.67

98.00

GGT

37.8646.41

54.9563.50

Contrast_edge 10.00

26.33

42.67

59.00

Canopy closure0.20

11.47

22.75

34.03

hsp60

Figure 12: Plots of observed GGT and hsp60 levels as a function of contrast edge density (L2) and canopy closure. Females are depicted by circles and males by squares. Red symbols are adults and blue symbols are subadults. Light blue stars are females with cubs. Reproduction Constrained ordination procedures were conducted for the reproduction variables. This analysis was restricted to females captured before July which corresponds in general with the end of the spring breeding time period. Ordination results suggested that variation in mean patch size was significantly related to reproductive variables. Observation of the biplot suggested that variation in mean patch size was most directly related to levels of testosterone (Figure 13). Observation of data plots suggests this relationship is strongest for subadult females and weakest for adult females.

-0.4 1.4

-1.2

0.4

progeste

estrogen

luteiniz

testoste

risk

rsf

cc

L1area

L2econ

L3shapeL4cv

L5prox

pctconpctpro

laf

dem

sexclass AF AFC SF

test

oste

rone

_ngm

l

0.00.10.20.30.40.50.60.70.80.91.01.11.21.3

L4) Variation in mean patch size0 1000 2000 3000 4000

Figure 125: Constrained ordination biplots for adult female reproductive hormone analysis. Significant environmental variables are underline. A plot of observed testosterone versus variation in mean patch size is also shown.

180


Immunity Constrained ordination analysis suggested that elevation influenced levels of immunity. However, inspection of biplots did not suggest that elevation was directly related to any of the individual immunity variables (Figure 14). As discussed later, individual biological variation in immunity may limit the ability to detect spatial differences in immunity levels based on environmental variables. Future analysis for the final program report will focus on temporal change in immunity of individual bears.

-1.0 0.4

-0.4

1.2

wbc

total_prglobulinAG_ratio

neu

lym

mono

baso

eos

LAFMEANCC

DEM

RISKRSF

L1_area

L2_Econ

L3_shape

L4_CV

pctcon

pctct10k

Figure 14: Constrained ordination biplots for immunity analysis. Significant environmental variables are underlined. Discussion The results of these analyses demonstrate the utility of constrained ordination methods to explore the relationship between bear demography and health and the environment. There are several advantages to this method. First, it allows the simultaneous evaluation of all the predictor GIS/landscape and demographic/health variables. This approach considers the correlation within GIS/landscape and health variables when evaluating the relationship between these two variables therefore allowing a more thorough investigation of relationships than univariate analyses. By doing this the number of overall statistical tests is reduced therefore reducing type 1 error rates. Program CANOCO adds the use of Monte Carlo significance tests to help interpretation of complex canonical relationships as well as inclusion of repeated measures data. Finally, biplots and triplots condense analysis results into a figure that shows likely gradients as well as statistically strong predictors and response variables. The main disadvantage of constrained ordination is that the choice of health response variables as well as predictor variables can potentially influence the overall fit of the ordination model to the data. For example, within a data set there may be a few environment-health variables that are strongly related as well as some variables that are completely unrelated. Unlike principal components analysis, the ordination is specifically

181


constrained so that the health variables can only be explained by the predictor variables. This reduces the overall fit of constrained ordination models compared to unconstrained ordination methods. Some of the most interesting results pertain to potential relationships between bear demography, condition and GIS/environmental variables. The demographic relationship identified environmental gradients in the data set that partially segregated age and sex classes. Some of the gradients could potentially be explained by decreased survival rates in roaded areas shifting the distribution of ages to younger bears. For example, subadult males were more likely to have higher road densities within their home range areas since decreased survival rates in these areas potentially eliminated many of the adult males. However, some of the relationships, such as adult females with cubs displaying home range areas, higher amounts of protected status are less easily explained by mortality history. Regardless, this result suggests that different age and sex groups may have different environmental conditions and stressors within their home range areas. The condition analysis suggest that road density and elevation also affect bear size (straight line length). Higher road density is correlated with larger body size which may seem counterintuitive. This is potentially due to better habitat resources such as game and other meat sources whose densities are higher in areas of disturbance and younger forests. In contrast bear body size is lower in higher elevation areas, which is potentially due to lower habitat quality at higher elevations. This general result suggests that while mortality pressure may be higher in anthropogenic areas, the availability of food resources may actually reduce stress of bears living in these areas. Our next planned analysis of temporal trend in individual bears, will include structural equation modeling which will hopefully elucidate further details about this relationship. Although we are now only working with a portion of the animal stress data our preliminary analyses with available stress variables suggest potential relationships between canopy closure, contrast edge density and GGT levels. GGT is a potential indicator of oxidative stress (Lee et al 2004). Our finding of higher GGT levels with increasing canopy closure is difficult to explain. Areas of high canopy closure exist in both roaded anthropogenic foothills habitat as well as protected mountainous areas. In general, it is negatively correlated with elevation and terrain ruggedness. However, principal components analysis of GIS variable suggests that it still forms a more unique indicator of the environment (in comparison with other GIS variables). Future analyses will consider this variable in more detail given that it occurs as a significant predictor in many of the health analyses. In addition, future analyses will explore how landscape metrics may influence GGT and other levels. The results of ordination analyses suggest that landscape metrics are potentially correlated with some of the health indicators. Direct interpretation of metrics is more difficult than GIS variables. Further collaboration with Julia Linke (PhD student – U of C) on interpretation of metrics will help further interpret results. In addition, future analyses will more directly assess the link between energy expenditure and health through the inclusion of movement rates and home range size as predictor variables of health scores. Many of the landscape metrics are descriptors of habitat interspersion. It is possible that bears have to use different strategies and levels of energy expenditure as anthropogenic change causes changes in

182


landscape metrics and patch interspersion. This will be best examined through the assessment of temporal change in individual bear habitat and health parameters. These results identify potential linkages between bear demography, health, and the environment. However, the results are statistical correlations, and the next step is to use these results, and the actual biology of bears to derive more ecologically meaningful models. This type of analysis will be conducted using structural equation models (Mitchell 1992, Shipley 1997) in the next phase of the project. Future analyses for final program report Future analysis will focus on temporal change in individual bear health and environment now that temporally matched data sets are available. This may potentially have more power than the spatial-only analyses detailed in this report. One interesting approach is the tracking of individual bears’ ordination scores over time. For example some ordination analyses suggest health and environmental gradients within data sets. The question becomes, if the composition of GIS indicator variables change in a bear home range what is the corresponding change of the bears score relative to health variables? A technique to interpret multivariate time series called principal response curves will be used to explore these type of relationships (Van den Brink and Ter Braak 1999).

Further analysis of reproductive rates will be conducted using multi-strata models mark-recapture models. The methods of Schwartz et al (In press) will be used to estimate age of first reproduction and reproductive rate. Both the effect of health (i.e. condition, movement rate) and environment will be explored using covariates. If successful, this analysis will provide a link between environmental/health and population demography (productivity). The results of the ordination and mixed model analyses will influence the choice of covariates for this analysis.

Finally, Structural Equation model analysis (in collaboration with Dr. Tak Fung and Dr. Scott Nielsen) will be used to further explore hypothesized linkages between latent health and environmental variables. This approach provides a useful way to directly test biologically based models of the relationship between health and the environment. Literature cited Boulanger, J. 2006. Preliminary analysis of GIS, Landscape metric and health variables

Nelson, BC: Integrated Ecological Research. Jongman, R. H. J., C. J. F. Ter Braak, and O. F. R. V. Tongeren, editors. 1995. Data analysis

in community and landscape ecology. Cambridge, England: Cambridge University Press. Lee, D.-H., R. Blomhoff, and D. R. Jacobs. 2004. Is serum gamma glutymyltransferase a

marker of oxidative stress? Free Radical Research 38:535-539. Leps, J., and P. Smilauer. 2003. Multivariate analysis of ecological data using CANOCO.

Cambridge, UK: Cambridge University Press. McGarigal, K., S. Cushman, and S. Stafford. 2000. Multivariate statistics for wildlife and

ecology research. New York: Springer. Mitchell, R. J. 1992. Testing Evolutionary and Ecological Hypotheses Using Path-Analysis

and Structural Equation Modeling. Functional Ecology 6(2):123-129.

183


184

Shipley, B. 1997. Exploratory path analysis with applications in ecology and evolution. The American Naturalist 149(6):1113-1138.

Ter Braak, C. J. F., and P. Smilauer. 2002. CANOCO: Reference manual and CanoDraw for windows user's guide: Software for canonical community ordination (Version 4.5). Ithasca, New York: Microcomputer Power.

Van den Brink, P. J., and C. J. F. Ter Braak. 1999. Principal response curves: Analysis of time-dependent multivariate responses of a biological community to stress. Environmental toxicology and chemistry 18:138-148.

Worton, B. J. 1989. Kernal methods for estimating the utilization distribution in home-range studies. Ecology 70:164-168.

Chapter 9: Training Program

CHAPTER 9: DELIVERY OF NEW PRODUCTS AND DEVELOPMENT

OF TRAINING PROGRAMS

Gordon Stenhouse1

1 Foothills Research Institute

In June 2007 we delivered new program products to all research sponsors. These new products represent the use of landscape and habitat conditions current as of 2005 and the amalgamation of 9 years of intensive data from GPS radio collared grizzly bears. These products include:

• RSF maps for each Bear Management Unit (BMU) • Mortality risk maps for each BMU • Safe Harbour maps for each BMU • New Grizzly Bear Movement Corridor Maps for each BMU • Habitat maps which are comprised of: leading species, crown closure , NDVI • New GIS applications to allow evaluation of forest management actions within

grizzly bear habitat • A listing of all research publications from our program (1999-2007)

Our research team is now completing teaching materials (chapters) for an integrated training program which will be delivered through the ENFORM training program. We anticipate that all materials from our scientists will be delivered to ENFORM in February 2008 and then we will begin the assembly of training manuals, with a goal of being able to offer the delivery of this course as scheduled in the spring of 2008. An outline of this course is presented in Table 1.

185


Table 1: Training course outline.

GGRRIIZZZZLLYY BBEEAARR HHAABBIITTAATT MMAANNAAGGEEMMEENNTT TTOOOOLLSS:: DDEESSIIGGNN DDOOCCUUMMEENNTT –– DDRRAAFFTT

COURSE OUTCOME: AN UNDERSTANDING OF FRI GRIZZLY BEAR RESEARCH RESULTS AND HOW TO APPLY THESE TO LAND MANAGEMENT DECISIONS.

Module 1: What is the data we use?

Outcome: An understanding on how the key data sets used for product development are gathered. (Strengthens and weaknesses)

Includes Subjects/Topics:

Objectives: (30minutes)

1. Explain how grizzly bear location data is obtained and what this represents. (Gordon Stenhouse)

Capture, collaring, GPS evolution and issues

(30minutes) 2. Other data and samples collected and what they tell us. (Gordon Stenhouse)

Animal health, vegetation and scats

( 1-1.5 hr) 3. Habitat mapping-evolution and current status. (Dr. Greg McDermid)

RS datasets, GIS layers and Data layers

Note: These provide the foundation for the development of products

Module 2: How are models created to identify grizzly bear habitat and bear movements across the landscape?

Outcome: A basic understanding of the techniques and procedures used to create some of the grizzly bear research products.

Objectives:

Dr. Scott Nielsen (1.5 hours) 1. Resource Selection Function

Models.

-RSF occurrence -Seasonal use -Mortality risk -Safe harbours -Strengths and limitations

⇐K

NO

WL

ED

GE

1 D

AY⇒

Barb Schwab (1 hour ) 2. Graph Theory Models.

-Methods -Testing -Input layers -Interpreting outputs

Module 3: New Data Sets – Population Inventory and Putting Model Outputs Together

Outcome: Understanding grizzly bear inventory and key conservation principles

⇐A

PPL

ICA

TIO

N1

½D

AY

S

Objectives:

1. Review of DNA based grizzly bear population inventory.

-Methods -Population units

186


(30minutes) (Gordon Stenhouse) -Results (30minutes)

2. Key concepts for the Conservation of Grizzly Bears in Alberta.

(Gordon Stenhouse)

-Use of research products -Human caused mortality -Grizzly bear priority areas -Grizzly bear response to landscape change

Module 4: Using Research Products to Support Land Use Planning Decisions

Outcome: This module will allow participants to understand the steps required in conducting an assessment of 2 land use activities (forestry and oil and gas) using the research program models and tools

Objectives: (1.5 hours)

1. Review of GIS applications developed

(Jerome Cranston)

-RSF -Risk -Safe harbour

⇐A

PPL

ICA

TIO

N ⇒

2. Forestry Analysis Example - 2 pass harvest - Future (10 years) - Mitigation

3. Oil and Gas Analysis example - Comparing options GROUP

DISCUSSION 4. Interpreting Output files

- End of DAY 1 - Exam Required

DAY 2

This day will be a GIS technical day for those students who will be running script and conducting assessments using the tools and products received from this research program. (Prerequisites will be identified for participation). Observers are welcome to attend to gain a further understanding of processes and output files. We are planning to hold our first training course for program partners with an ENFORM training coordinator and the individual research scientists as the delivery team. As we move forward with additional training courses we may decide to train other course lecturers to deliver this course on our behalf. Regular program updates are expected to occur from the research team as new work is completed. In addition to the preparation of our training course the program GIS specialist and program leader also delivered a number of group training efforts in Calgary, Hinton and Grande Prairie with a focus on assisting program sponsors in using the research products and GIS applications. These sponsors have provided the industry matching funding within our current Innovation and Science grant. These partner specific training sessions will continue as our larger training program is completed. We also recognize that there will be an ongoing need to assist program partners with the use and application of the products that our research team has provided.

187


188

In 2007 we have been working with a group of SRD staff to assist them in using and understanding the new products that we have provided. This work is the next logical step in putting the research results into practice and fulfills the “Service Excellence” goal of our current grant with Alberta Advanced Education and Technology. As part of this ongoing work our team has been working with SRD staff on the development of an instruction or guidance document which the government will provide to industrial planners and developers (forestry/oil&gas sectors). This document will form a crucial link for research results and government policy direction in relation to grizzly bear and landscape management in Alberta. The focus of this document is to explain what is required by an applicant who is planning industrial activities within occupied grizzly bear range in the province. The proposed analysis and applications are deliverables from our research team. The Draft document is presented below:


Guidelines on Conducting an Analysis of Land Use Activities on Grizzly Bear Habitat in Alberta

February 2008

This document describes how the models and tools developed by the Foothills Research Institute (FRI) Grizzly Bear Program (GBP) can be used to evaluate proposed developments (roads, pipelines, cut blocks, well sites, etc) in terms of their impact on grizzly bear habitat. This supports the long-term goal of the GBP, which is:

To provide resource managers with the necessary knowledge and planning tools to ensure the long-term conservation of grizzly bears in Alberta.

Background All program partners who have contributed to the GBP during the past nine years (1999-2007) have received a number of program deliverables from our research team. These deliverables are provided to program partners on a regular basis and are regularly improved as data quality and geographic extent are enhanced with additional work. We are now able to utilize the program deliverables (maps, models and GIS applications) to conduct assessments of proposed land use activities on grizzly bear habitat. The intent of this document is to provide guidance and direction for those conducting these analyses. We have provided a step-by-step description of how to conduct these analyses, and recommended specific grizzly bear habitat objectives, as well as listing some possible mitigation options. As more users gain experience with these analyses and application of results we expect that new insights and mitigation techniques will evolve. The objective of using our models and tools is to assist in managing for landscape conditions necessary for the long-term health and persistence of grizzly bears in the province of Alberta. It is recommended that land-use plans and applications be examined in terms of their current and projected impacts on grizzly bear habitat quality, in order that the effect of planned development/activities, and mitigation efforts, can be quantified. It is important to keep in mind that the primary objective of these analyses is to ensure the maintenance of adequate grizzly bear habitat to ensure the long-term survival of the species. Before You Begin

189

The tools and models that you are using are complex and highly technical. There is no requirement that you have a complete understanding of all aspects of model construction or GIS application development. However there is an expectation that the product user understand the research findings and the key principles of grizzly bear conservation.


To aid with this understanding we have previously provided all program partners with a PowerPoint slide show, which pertains directly to the suggested analysis. (File :GBRP Analysis 2007.ppt) The GBRP_2006Deliverables_31mar07.ppt PowerPoint is divided into 4 sections. It is recommended that users familiarize themselves with the following slides before performing any analysis: Slides 1 – 49:

History and objectives of GBRP Introduction to remote-sensing based habitat mapping Summary of research results

Slides 50 – 165:

Current program areas: 2.3 Introduction to Graph Theory corridor modeling 2.5 DNA –based population inventory 2.9 GIS applications 2.10 Delineation of habitat zones

Slides 166 – 188: Description of models and tools

Slides 189 – 249:

How to use the models and tools (sample analysis).

THE PRODUCTS

1. Remote Sensing - These datasets are raster-based GIS layers, in ESRI grid format, identifying vegetation and forest structure and are derived from remote sensing imagery (Landsat) at 30m resolution (pixel size). The 10-class landcover was generated by object-oriented classification of 22 Landsat Thematic Mapper 5 images and topographic derivatives from a digital elevation model, followed by a change-detection update to 2005 over the Phase 6 area. Associated datasets include species composition (percent conifer/broadleaf), crown closure, and NDVI (Normalized Difference Vegetation Index), a time-series of 12 biweekly images tracking changes in plant phenology. A 15-class landcover was also created by combining the 10-class landcover with the crown closure and species composition models. Also included is a forest age layer, based on AVI and stand history data.

190

2. Resource Selection Function (RSF) - These are raster-based datasets, at 30m resolution, showing the relative probability of grizzly bear occurrence on the landscape. They are derived from landcover and other GIS layers, and have been tested and validated with at least 2 years of grizzly bear location data collected by GPS radio collars. Population-level models for each population unit have been developed for three seasons (spring, summer


and fall). Grizzly bear occurrence, as represented by the RSF model, is highly correlated with the amount and spatial distribution of grizzly bear habitat resources such as food, water, and thermal cover and thus the RSF model can be used as a surrogate of habitat quality. For more information on the RSF model, please refer to the document RSF_faq_v3.doc.

3. Mortality Risk - Using spatial and temporal data on grizzly bear mortalities, we have produced a raster-based grizzly bear mortality risk model, which predicts the probability of human-caused grizzly bear mortality over the landscape. This dataset is based on the most current data for open, motorized linear access structures, including roads and rights-of-way, for the entire Phase 6 area. This model describes habitat security, which is a critical aspect of grizzly bear habitat quality. For more information on the Mortality Risk model, please refer to the document risk_faq_v3.pdf

4. Grizzly Bear Movement Corridors - Graph Theory has been applied to the RSF datasets to predict the configuration of grizzly bear travel corridors on the landscape. These datasets also provide a ranking of the relative importance of movement corridors and have been constructed using GPS grizzly bear location data. Only data from the Phase 3 area and Phase 4 areas are currently available (covers the Waterton, Livingstone, Clearwater and FRI Core population units). For more information on the Movement Corridors model, please refer to the document GT_faq.doc.

5. Watersheds - Watershed analysis units for the Phase 6 study area were created to provide an appropriate mesoscale landscape unit for generating summary statistics for grizzly bear habitat. Major watersheds were subdivided (generally along heights-of-land, occasionally along watercourses) to approximate the size of an adult female grizzly bear home range (~700 sq km).

6. Grizzly Bear Population Units - Alberta Grizzly Bear range has been subdivided into population units based on genetic distinctions within the Alberta grizzly bear population (Fig. 1). The FRI Core population unit, between Highway 16 and Highway 11, corresponds to BMA (Bear Management Area) 3 and was the subject of a DNA-based mark-recapture survey in 2004. Bear density was concentrated in the west half of the study area, and averaged 4.79 bears per 1000 sq km. The Clearwater population unit, between Highway 11 and Highway 1, corresponds to BMA 4 and was DNA-surveyed in 2005. Bear density was concentrated in the west half of the study area, and averaged 5.25 bears per 1000 sq km. The Livingstone population unit, between Highway 1 and Highway 3, corresponds to BMA 5 and was DNA-surveyed in 2006. Bear density was concentrated in the west half of the study area, and averaged 11.77 bears per 1000 sq km. The Waterton population unit, between Highway 3 and the US border, corresponds to BMA 6 and was DNA-surveyed in summer 2007. Results will be released following ministerial review.

191

7. Safe Harbours and Attractive Sinks - From the work of Dr. Scott Nielsen (research team member) comes the concepts of attractive sinks and safe harbours. A safe harbour is an area of good habitat (as indicated by high RSF values) and low risk of human-caused mortality. An attractive sink, the inverse of a safe harbor, is an area of good habitat, to which bears are attracted by an abundance of resources, but where bears face a high risk of mortality. The safe harbor model combines the two critical aspects of habitat


quality – resource availability, and security – into a single model, and has been developed for the entire Phase 6 area.

8. Habitat States – This raster-based dataset is a categorical classification of the safe harbor model, which identifies primary habitat areas (serving as a population source), and attractive sinks (acting as a population sink).

Class Description 0 non-critical habitats (very low habitat value to non-habitat) 1 secondary sinks (moderate/secondary habitat value & high mortality

risk) 2 primary sink (high/primary habitat value & high mortality risk) 3 secondary habitats (moderate habitat value & low mortality risk) 4 primary habitats (high habitat value & low mortality risk)

9. Priority Areas and Dispersal Zones – The Alberta Grizzly Bear Recovery Plan (#15) identifies two important landscape units to aid in recovery efforts in Alberta. The first are Grizzly Bear Priority Areas which contain the highest quality grizzly bear habitat combined with the lowest level of grizzly bear mortality risk. These areas also have road density thresholds of .6km/km2. The second areas have been termed Grizzly Bear Dispersal Zones which are adjacent to the Priority Areas. These Dispersal Zones contain good quality grizzly bear habitat and have a higher level of mortality risk and a road density threshold of 1.2 km/km2. Within Priority Areas, long-term access plans will be required. Draft Priority Areas and Dispersal zones have been submitted to ASRD for review.

OTHER PRODUCT DESCRIPTIONS

10. GIS applications - Geoprocessing scripts, written in the Python language, and associated GIS input layers allow the user to regenerate the habitat models with different scenarios and thereby forecast changes to grizzly bear habitat caused by industrial development. Planned features such as roads, trails, cutblocks, wellsites, or pipelines can be incorporated into the model base layers, and the RSF and mortality risk models regenerated. These scripts require ESRI ArcGIS 9x with ArcInfo functionality and Spatial Analyst extension.

11. Documents: There are a number of documents included in the deliverables including:

2006 annual report.pdf GBP_Deliverables2_2006_srd.pdf 100k_risk_metadata.txt risk_faq.doc GT_faq.doc RSF_faq_v3.doc Ursus_habitat_modeling.pdf Nielsen_RSF_report.pdf habitat_states_definitions.pdf

192

mortality risk Nielsen et al 2004.pdf


ReadMe_Risk_p6.pdf

CONDUCTING THE ANALYSIS The analysis is a 6-step process:

1. Collect datasets from the files provided from the GBRP, and user development plans. 2. Select analysis extent and determine management objectives. 3. Examine current conditions at regional (FMA/FMU) and Grizzly Bear Watershed

Units (WU)-levels. i. RSF

ii. Mortality Risk iii. Open Route Density iv. Safe Harbour Index

4. Incorporate development scenarios into habitat models (using GB GIS tools) and examine future conditions at Grizzly Bear Watershed Units (WU)-level.

i. RSF ii. Mortality Risk

iii. Open Route Density iv. Safe Harbour Index

5. Compare current and future conditions. 6. Investigate and analyze options for mitigation.

This process is explained in more detail below. For a step-by-step description of the analysis process, including a case study analyzing a harvest plan for FMU E8, refer to the Appendix. 1.0 Collect datasets Using GIS software (ArcMap, ArcView, etc), determine where the planned development lies in relation to the GBP Phase 6 study area boundary (Figure 1, outlined in yellow), habitat model extents (blue), and Habitat Zones (brown). This will determine whether analysis is required, and what datasets are available (see list of products above).

193


2.0 Select Analysis extent, determine management objectives Grizzly Bear watershed units (approx. 700 km2) are provided as an appropriate unit for analysis. Other extents, such as operating compartments or other administrative units of similar scale, may also be suitable. The analysis should be conducted for lands identified as grizzly bear habitat. However the focus should be on projected impacts on Grizzly Bear Priority Areas and Dispersal Zones. The management objective for these areas is to improve or maintain grizzly bear habitat quality as indicated by an increase in Safe Harbour Index. The Safe Harbour Index is the mean Safe Harbour pixel value in the analysis unit. If the analysis unit lies within a Priority area or Dispersal area (Figure 2), an additional management objective is to maintain open route density below 0.6 km/km2 (Priority areas) or 1.2 km/km2 (Dispersal zones). 3.0 Examine Current Conditions Examine the current state of habitat quality in the vicinity of the planned development, at the regional (BMA) and analysis unit levels, using the RSF, Risk, and Graph Theory models.

3.1. RSF*

194

Fig. 1 Fig. 2


• Use RSF script to regenerate seasonal maximum RSF model over analysis unit. • Calculate mean RSF over analysis unit.

3.2. Mortality Risk** • Use Risk script to regenerate Risk model over analysis unit • Calculate mean Risk value over analysis unit.

3.3. Open road density (applicable within Priority and Dispersal zones) • Clip existing open route network by analysis unit. • Sum length of all arcs, divide by area of analysis Unit to calculate open route

density. 3.4. Safe harbour

• Combine Risk and seasonal maximum RSF to generate Safe Harbour model for analysis unit.

• Calculate Safe Harbour Index (Mean Safe Harbour). * note: there are 3 seasonal RSF map layers (spring, summer and fall) along with a mean annual RSF product. The user should select the mean annual RSF for long term activities and a seasonal RSF map for short term activities (i.e. drilling a gas well over a short time frame. **note: 4.0 Generate Future Landscapes Use the RSF and mortality risk scripts to regenerate the RSF and Risk models with planned roads, trails, and openings (cut blocks, rights-of-way, well sites, etc).

4.1. RSF • Use RSF script to regenerate seasonal maximum RSF model over analysis unit. • Add shapefiles or feature classes of planned roads (line) and openings (polygon)

in the second and third boxes. • Calculate mean RSF over analysis unit.

4.2. Mortality Risk • Use Risk script to regenerate Risk model over analysis unit. • Add shapefiles or feature classes of planned roads (line), trails (line), and

openings (polygon) in the second, third, and forth boxes. • Calculate mean Risk over analysis unit.

4.3. Open road density • Clip future open route network by analysis unit. • Sum length of all arcs, divide by area of analysis Unit to calculate open route

density. Mean safe harbour

• Combine Risk and seasonal maximum RSF to generate Safe Harbour model for analysis unit.

• Calculate Safe Harbour Index (Mean Safe Harbour). 5.0 Compare Current and Future Conditions

195

Compare current and forecast conditions in the table:


196

Grizzly Bear Habitat Assessment Form Date of Analysis: Analysis performed by:

1 Watershed Unit # 2 GB Population Unit 3 Priority/Dispersal Area:

Analysis Variables: Current Future % Change Increase/Decrease Mean RSF score Mean Mortality Risk Open road density Safe Harbour Index

Typically, the creation of new openings in forested areas will result in an increase in RSF scores, due to the formation of edges and, in the case of forest harvesting, the replacement of mature or overmature stands with young seral stands. However, the construction of new access features that accompanies such development also leads to an increase in mean mortality risk. The Safe Harbour Index incorporates both these changes into a single value; if the Safe Harbour Index is forecast to decrease as a result of the planned development, the management objectives have not been achieved and alternative plans or mitigation measures (such as road deactivation) should be provided and an additional analysis conducted. 6.0 Investigate and Re-Analyze Options for mitigation The GBP has provided a several options for mitigating the effect of new developments on GB habitat. Some of these include:

• In the case of forest harvesting – redesign harvest to increase edges (irregular) • Leave visual buffers around key habitats • Change timing of operations to seasons of lower occupancy • Control public access – locked gates, remove access structures. • Investigate disturbance reclamation possibilities to enhance grizzly bear foods (note :

research currently underway to evaluate well site reclamation at abandoned wellsites.

APPENDIX 1: Program Partners

APPENDIX 1: LIST OF PROGRAM PARTNERS

Ainsworth Lumber Co. Ltd. Alberta Advanced Education and Technology (formerly Innovation and Science) Alberta Conservation Association Alberta Environment Alberta Fish & Game Association Alberta Newsprint Company Alberta Sustainable Resource Development (formerly Natural Resources Service) Anadarko Canada Corporation Anderson Exploration Ltd. AVID Canada B P Canada Energy Company Banff National Park BC Oil and Gas Commission Blue Ridge Lumber Inc. Buchanan Lumber – Tolko OSB Canada Centre for Remote Sensing Canadian Association of Petroleum Producers (CAPP) Environmental Research Advisory Council (ERAC) Fund Canadian Forest Service Canadian Natural Resources Ltd. Canadian Wildlife Service Canfor Corporation Center for Wildlife Conservation ConocoPhillips Canada (formerly Burlington Resources Canada Ltd.) (formerly Canadian Hunter Exploration Ltd.) Conservation Biology Institute Daishowa Marubeni International Ltd. Devon Canada Corp

Elk Valley Coal Cardinal River Operations

Enbridge Inc. Encana Corporation Environment Canada –HSP Foothills Research Institute Forest Resources Improvement Association of Alberta (FRIAA) G&A Petroleum Services GeoAnalytic Inc. Government of Canada Canadian Forest Service, Natural Resources Canada Canadian Wildlife Service Environment Canada – HSP Human Resources and Skills Development Canada Natural Sciences and Engineering Research Council of Canada (NSERC) Parks Canada Banff National Park Jasper National Park Hinton Fish and Game Association Hinton Training Centre Husky Energy Inc. Komex International Ltd. Lehigh Inland Cement Limited Luscar Ltd. Coal Valley Resources Inc

Gregg River Resources Ltd. Manning Diversified Forest Products Ltd. Manning Forestry Research Fund Millar Western Forest Products Ltd. Mountain Equipment Co-op Nexen Inc. Northrock Resources Ltd.

Peregrine Helicopters Persta Petro Canada Ltd. Petroleum Technology Alliance Canada (PTAC) Peyto Energy Trust Precision Drilling Corporation Rocky Mountain Elk Foundation - Canada Shell Canada Limited Slave Lake Division – Alberta Plywood Spray Lake Sawmills Ltd. Suncor Energy Inc. Sundance Forest Industries Ltd. Talisman Energy Inc. Telemetry Solutions TransCanada Pipelines Ltd. University of Alberta University of Calgary University of Lethbridge University of Saskatchewan Western College of Veterinary Medicine University of Washington University of Waterloo Veritas DGC Inc. West Fraser Mills Ltd. Alberta Plywood

Blue Ridge Lumber Inc. Hinton Wood Products Slave Lake Pulp Sundre Forest Products Weyerhaeuser Company Limited Wilfred Laurier University World Wildlife Fund Canada

197

APPENDIX 2: Published Papers

APPENDIX 2: LIST OF PUBLISHED PAPERS

Boulanger, J., G.C. White, M. Proctor, G. Stenhouse, G. Machutchon, S. Himmer. 2008. Use

of occupancy models to estimate the influence of previous live captures on DNA-based detection probabilities on grizzly bears. Journal of Wildlife Management 72:000:000.

Boulanger, J., M. Proctor, S. Himmer, G. Stenhouse, D. Paetkau, J. Cranston. 2006. An

empirical test of DNA mark-recapture sampling strategies for grizzly bears. Ursus 17:149-158.

Boulanger, J., G. Stenhouse, R. Munro. 2004. Sources of heterogeneity bias when DNA

mark-recapture sampling methods are applied to grizzly bear (Ursus arctos) populations. Journal of Mammalogy 85:618-624.

Cattet, M., G. Stenhouse, and T. Bollinger. 2008. Exertional myopathy in a grizzly bear

(Ursus arctos) captured by leg-hold snare. Journal of Wildlife Diseases 00:000-000. Cattet, M.R., A. Bourque, B.T. Elkin, K.D. Powley, D.B. Dahlstrom, N.A. Caulkett. 2006.

Evaluation of the potential for injury with remote drug-delivery systems. Wildlife Society Bulletin 34:741-749.

Cattet, M.R.L., K. Christison, N.A. Caulkett and G.B. Stenhouse. 2003. Physiologic

responses of grizzly bears to different methods of capture. Journal of Wildlife Diseases 39(3):649-654.

Cattet, M.R.L., N.A. Caulkett, and G.B. Stenhouse. 2003. Anesthesia of grizzly bears using

xylazine-zolazepam-tiletamine or zolazepam-tiletamine. Ursus 14(1):88-93. Cattet, M.R.L., N.A. Caulkett, M.E. Obbard and G.B. Stenhouse. 2002. A body-condition

index for ursids. Canadian Journal of Zoology 80:1156-1161.

Frair, J.L., S.E. Nielsen, E.H. Merrill, S. Lele, M.S. Boyce, R.H.M. Munro, G.B. Stenhouse, and H.L Beyer. 2004. Removing GPS-collar bias in habitat-selection studies. Journal of Applied Ecology 41, 201-212.

Franklin, S. E., P. K. Montgomery, and G. B. Stenhouse. 2005. Interpretation of land cover

using aerial photography and satellite imagery in the Foothills Research Institute of Alberta. Canadian Journal of Remote Sensing 31:304-313.

Franklin, S.E., D.R. Peddle, J.A. Dechka, G.B. Stenhouse. 2002. Evidential reasoning with

Landsat TM, DEM and GIS data for landcover classification in support of grizzly bear habitat mapping. International Journal of Remote Sensing 23(21):4633-4652.

198


Franklin, S.E., M.B. Lavigne, M.A. Wulder, and G.B. Stenhouse. 2002. Change detection

and landscape structure mapping using remote sensing. The Forestry Chronicle 78(5):618-625.

Franklin, S.E., M.J. Hansen, G.B. Stenhouse. 2002. Quantifying landscape structure with

vegetation inventory maps and remote sensing. The Forestry Chronicle 78(6):866-875.

Franklin, S.E., G.B. Stenhouse, M.J. Hansen, C.C. Popplewell, J.A. Dechka, D.R. Peddle.

2001. An integrated decision tree approach (IDTA) to mapping land cover using satellite remote sensing in support of grizzly bear habitat analysis in the Alberta Yellowhead ecosystem. Canadian Journal of Remote Sensing 27(6):579-592.

Gau, R.J., R. Mulders, L.M. Ciarniello, D.C. Heard, C.B. Chetkiewicz, M. Boyce, R. Munro,

G. Stenhouse, B. Chruszcz, M.L. Gibeau, B. Milakovic, K. Parker. 2004. Uncontrolled field performance of Televilt GPS-Simplex collars on grizzly bears in western and northern Canada. Wildlife Society Bulletin 32:693-701.

Huettmann, F., S.E. Franklin, G.B. Stenhouse. 2005. Predictive spatial modelling of

landscape change in the Foothills Research Institute. Forestry Chronicle 81:525-537. Hunter, A., N. El-Sheimy, G. Stenhouse. 2005. GPS/Camera Collar Captures Bear Doings.

http://www.gpsworld.com/gpsworld/article/articleDetail.jsp?id=146689 (this will only be here until September 2006 or so).

Hunter, A. 2007. Sensory Based Animal Tracking. PhD Thesis. Department of Geomatics

Engineering, University of Calgary, Calgary, Alberta, Canada. Linke, J. 2003. Using Landsat TM and IRS imagery to Detect Seismic Cutlines: Assessing

their Effects on Landscape Structure and on Grizzly Bear (Ursus arctos) Landscape Use in Alberta. MSc. Thesis. Department of Geography, University of Calgary, Calgary, Alberta, Canada.

Linke, J., S.E. Franklin, F. Huettmann and G.B. Stenhouse. 2005. Seismic cutlines, changing

landscape metrics and grizzly bear landscape use in Alberta. Landscape Ecology 20:811-826.

McDermid, G. J., S. E. Franklin, and E. F. LeDrew (in press). Radiometric normalization and

continuous-variable model extension for operational mapping of large areas with Landsat imagery. International Journal of Remote Sensing 00:000-000..

McDermid, G. J., S.E. Franklin and E.F. LeDrew. 2005. Remote sensing for large-area

habitat mapping. Progress in Physical Geography 29:449-474.

199

http://www.gpsworld.com/gpsworld/article/articleDetail.jsp?id=146689


McDermid, G.J. 2005. Remote Sensing for Large-Area, Mult-Jurisdictional Habitat Mapping. PhD. Thesis. Department of Geography, University of Waterloo, Waterloo, Ontario, Canada.

Mowat, G., D.C. Heard, D.R. Seip, K.G. Poole, G.Stenhouse, D.W. Paetkau. 2005. Grizzly

Ursus arctos and black bear U. americanus densities in the interior mountains of North America. Wildlife Biology 11: 31-48.

Munro, R.H.M, S.E. Nielsen, M.H. Price, G.B. Stenhouse, M.S. Boyce. 2006. Seasonal and

diel patterns of grizzly bear diet and activity in west-central Alberta. Journal of Mammology 87:1112-1121.

Nielsen, S.E., G.B. Stenhouse, M.S. Boyce. 2006. A habitat-based framework for grizzly

bear conservation in Alberta. Biological Conservation 130:217-229. Nielsen, S.E. 2004. Habitat ecology, Conservation, and Projected Population Viability of

Grizzly Bears (Ursus arctos L.) in West-Central Alberta, Canada. PhD Thesis. Department of Biological Sciences, University of Alberta, Edmonton, Alberta, Canada.

Nielsen, S.E., M.S. Boyce, and G.B. Stenhouse. 2004. Grizzly bears and forestry I: selection

of clearcuts by grizzly bears in west-central Alberta. Canada.Forest Ecology and Management 199:51–65.

Nielsen, S.E., R.H.M. Munro, E. Bainbridge, M.S. Boyce, and G.B. Stenhouse. 2004.

Grizzly bears and forestry II: distribution of grizzly bear foods in clearcuts of west-central Alberta, Canada. Forest Ecology and Management 199:67–82.

Nielson, S.E., M.S. Boyce, G.B. Stenhouse, R.H.M. Munro. 2003. Development and testing

of phenologically driven grizzly bear habitat models. Ecoscience 10(1):1-10. Nielson, S.E., M.S. Boyce, G.B. Stenhouse, and R.H.M. Munro. 2002. Modeling grizzly

bear habitats in the Yellowhead ecosystem of Alberta: taking autocorrelation seriously. Ursus 13:45-56.

Pape, A. 2006. Multiple Spatial Resolution Image Change Detection for Environmental

Management Applications. MSc. Thesis. Department of Geography, University of Saskatchewan, Saskatoon, Saskatchewan, Canada.

Pereverzoff, J.L. 2003. Development of a Rapid Assessment Technique to Identify Human

Disturbance Features in a Forested Landscape. MSc. Thesis, Department of Geography, University of Calgary, Calgary, Alberta, Canada.

Pape, A. D. and S. E. Franklin (in press). MODIS-based change detection for grizzly bear

habitat mapping in Alberta. Photogrammetric Engineering and Remote Sensing 00:000-000.

200


Popplewell, C., S.E. Franklin, G.B. Stenhouse, and M. Hall-Beyer. 2003. Using landscape

structure to classify grizzly bear density in Alberta Yellowhead Ecosystem bear management units. Ursus 14: 27-34.

Popplewell, C. 2001. Habitat Structure and Fragmentation of Grizzly Bear Management

Units and Home Ranges in the Alberta Yellowhead Ecosystem. MSc. Thesis. Department of Biological Sciences, University of Calgary, Calgary, Alberta, Canada.

Ritson-Bennett, R.A. 2003. Assessing the Effects of a Heli-portable 3D Seismic Survey on

Grizzly Bear (Ursus arctos horribilis) Distribution. MSc. Thesis, Department of Geography, University of Calgary, Calgary, Alberta, Canada.

Roever, C.L. 2006. Grizzly Bear (Ursus arctos L.) Selection of Roaded Habitats in a Multi-

use Landscape. M.Sc. Thesis, University of Alberta, Edmonton, Alberta, Canada. Schwab, B.L. 2003. Graph Theoretic Methods for Examining Landscape Connectivity and

Spatial Movement Patterns: Application to the FRI Grizzly Bear Research. MSc Thesis. Department of Geography, University of Calgary, Calgary AB.

Stenhouse, G.B., J. Boulanger, J. Lee, K. Graham, J. Duval, J. Cranston. 2004. Grizzly bear

associations along the eastern slopes of Alberta. Ursus 16:31-40. Stenhouse, G.B. and K.Graham. (Eds.). 2005. Foothills Research Institute Grizzly Bear

Research Program 1999-2003 Final Report. 289 pp. Stenhouse, G.B., R. Munro and K.Graham. (Eds.). 2003. Foothills Research Institute

Grizzly Bear Research Program 2002 Annual Report. 162 pp. Stenhouse, G.B., R. Munro. (Eds.). 2002. Foothills Research Institute Grizzly Bear

Research Program 2001 Annual Report. 126 pp. Stenhouse, G. and R. Munro. (Eds.). 2001. Foothills Research Institute Grizzly Bear

Research Program 2000 Annual Report. 87 pp. Stenhouse, G. and R. Munro. (Eds.). 2000. Foothills Research Institute Grizzly Bear

Research Program 1999 Annual Report. 98 pp. Wasser, S.K., B. Davenport, E.R. Ramage, K.E. Hunt, M. Parker, C. Clarke, and G.B.

Stenhouse. 2004. Scat detection dogs in wildlife research and management: Application to grizzly and black bears in the Yellowhead Ecosystem, Alberta, Canada. Canadian Journal of Zoology 82:475-492.

Wulder, M. A., and S. E. Franklin, eds., 2003, Remote Sensing of Forest Environments: Concepts and

Case Studies, Kluwer Academic Publishers, Boston, MA, 519p.

201


202

Wulder, M.A., S.E. Franklin, J.C. White, J.Linke and S. Magnussen. 2005. An accuracy assessment framework for large-are land cover classification products derived from medium-resolution satellite data. International Journal of Remote Sensing: in press.

Wunderle, A.L. 2006. Sensitivity of multi-resolution satellite sensor imagery to regenerating forest age and site preparation for wildlife habitat analysis. M.Sc. thesis, Department of Geography, University of Saskatchewan, 89p.

Wunderle, A.L., S.E. Franklin, X.G. Guo. 2006. Regenerating boreal forest structure estimation using SPOT-5 pansharpened imagery. International Journal of Remote Sensing: in press.

gbp_2008_03_annrpt_2007

Documents

research program

ongoing support

program field activities

logistical support

ongoing gis support

annual report

valuable support

interim report