snow and ice cover changes in the … degrassi...chapter 8. spectral analysis and...
TRANSCRIPT
SNOW AND ICE COVER CHANGES IN THE GLACIERS OF THE CORDILLERA REAL (BOLIVIA) 1987-2001:
A REMOTE SENSING STUDY
by
CHRISTIAN DEGRASSI
A thesis submitted in partial fulfillment of the requirements for the degree of
Masters of Arts in Geography
Appalachian State University
May 2004
ii
SNOW AND ICE COVER CHANGES IN THE GLACIERS OF THE CORDILLERA REAL (BOLIVIA) 1987-2001:
A REMOTE SENSING STUDY
A Thesis
By
CHRISTIAN DEGRASSI
May 2004
APPROVED BY: ______________________________ Michael Mayfield Chairperson, Thesis Committee ______________________________ James Young Member, Thesis Committee ______________________________ Jeffrey Colby Member, Thesis Committee ______________________________ James Young Chairperson, Department of Geography and Planning ______________________________ Judith E. Domer Dean, Graduate Studies and Research
iv
Abstract
“SNOW AND ICE COVER CHANGES IN THE GLACIERS OF THE
CORDILLERA REAL (BOLIVIA) 1987-2001:
A REMOTE SENSING STUDY.” (May 2004)
Christian Degrassi, B.S., University of Trieste, Italy
M.A., Appalachian State University
Thesis Chairperson: Dr. Michael Mayfield
Fluctuations of mountain glaciers and ice caps are key variables for the
observation of climate-related changes and the detection of enhanced greenhouse effects
(IPCC 2001). Due to the narrow range of climate conditions under which glaciers and ice
caps develop, monitoring the distribution, intensity, and rate of change of ice masses is of
utmost importance in delineating spatial and temporal variations of the Earth’s climatic
system.
The purpose of this research was to analyze snow and ice cover changes of the
glaciers of the Cordillera Real (Bolivia) during the period 1987-2001. The study was
based on satellite remote-sensing technology, and sought to enhance the regional
understanding of trends and dynamics of the glacier retreat observed on benchmark
glaciers. As a result of the research produced in this thesis an estimate was made of the
snow and ice cover changes affecting the glaciers of the Cordillera Real. The results are
in line with field observations of a gradual but quantifiable retreat from 1987. The snow
v
and ice cover of the Cordillera Real were found to be highly sensitive to changes in
regional weather patterns such as those induced by El Niño events, but may also be
related to long-term global change processes in the troposphere such as the increasing
planetary temperature. This study confirmed that, while not developed for glaciological
applications, Landsat platforms are a valuable resource to study snow and ice cover, but
present serious limitations in snow-covered tropical alpine environments.
vi
Acknowledgments
I would like to thank my thesis chair, Dr. Michael Mayfield, from whom I have
learned so many great things. I would like to thank him for helping and supporting this
research project, and for his endless patience. Also, I would like to thank my committee
members, Dr. James Young and Dr. Jeffrey Colby for their insight, guidance and support
throughout my graduate career and the thesis process; I must thank Mr. Baker Perry for
his assistance during the data-mining process.
Within the wonderful faculty of the Department of Geography and Planning, I
have particular thanks for Dr. Richard Crepeau, for being always available for discussion
and for helping solving many problems, for Dr. Kathleen Schroeder, without whom I
would not have been able to make it to the end, and Mr. Arthur Rex, for being always
available.
I must thank Dr. Edelma Huntley, Dr. Judith Domer, and the Graduate School for
supporting my research and providing a thesis research grant. A thank you also to the
GSAS for the research award.
I will never been able to thank enough my wife Allyson for being always close to
me, and for always letting me pursue all my dreams.
viii
Table of contents List of tables ..................................................................................................................... xi
List of figures ................................................................................................................. xiii
Chapter 1. Introduction ....................................................................................................... 1
Chapter 2. The Study of Glaciers: Methodology Overview ............................................... 3
On-site Measurements .................................................................................................... 5
Indirect Measurements.................................................................................................... 7
Coring ........................................................................................................................... 11
Chapter 3. Study Area....................................................................................................... 14
Exploration and Historical Data.................................................................................... 14
Area of Interest ............................................................................................................. 16
Chapter 4. Climatic Conditions Affecting Tropical Glaciers of Bolivia .......................... 21
Chapter 5. Remotely Sensed Change Detection of Snow and Ice .................................... 26
Supervised and Unsupervised Classifications .............................................................. 29
Band Ratioing ............................................................................................................... 30
Spectral Un-mixing....................................................................................................... 32
Chapter 6. Spectral environment....................................................................................... 34
Spectral Characteristics of Snow and Ice...................................................................... 34
Landsat platforms.......................................................................................................... 42
ix
TM and ETM+ spectrometers and sensors calibration ................................................. 43
Use of Landsat TM and ETM+ for Snow and Ice Spectrometry.................................. 52
Chapter 7. Preprocessing .................................................................................................. 59
Data Characteristics ...................................................................................................... 59
Geometric Correction and Co-registration.................................................................... 63
Study Area Subset......................................................................................................... 66
Atmospheric Correction................................................................................................ 69
Image Standardization .................................................................................................. 81
Chapter 8. Spectral Analysis and Classification ............................................................... 85
Spectral Characteristics of the Subscenes..................................................................... 85
Spectral Endmembers Definition.................................................................................. 88
Binary classification...................................................................................................... 90
Supervised Classification: Minimum Distance............................................................. 92
Spectral unmixing ......................................................................................................... 94
Chapter 9. Results ............................................................................................................. 97
Binary Classification: Zonal Statistics.......................................................................... 97
Minimum Distance: Zonal Statistics........................................................................... 100
Binary vs. Minimum Distance Classifier.................................................................... 109
Regional vs. Local Observations ................................................................................ 110
Chapter 10. Discussion and Conclusion ......................................................................... 114
Discussion................................................................................................................... 114
Technical, and Physical Limitations of the Results .................................................... 118
Conclusions................................................................................................................. 121
x
Bibliography ................................................................................................................... 124
References Cited ......................................................................................................... 124
References Consulted.................................................................................................. 131
Appendix A. Anisotropic Reflectance Correction and Topographic Normalization of
rugged terrains and snow covered surfaces. ................................................................... 132
Concepts of Anisotropic Reflectance and Topographic normalization ...................... 132
Topographic Normalization: Methods and Techniques.............................................. 133
Case Study: Image Processing .................................................................................... 136
Case Study: Results and Discussion ........................................................................... 139
Limitations and Final Consideration........................................................................... 144
Biographical information .................................................................................................146
xi
List of tables Table 3.1. Glaciers of the Bolivian Andes as November, 1st 1984 (modified after Jordan
1991). .................................................................................................................... 18
Table 6.1. Spectral and spatial specifications for TM and ETM+ spectrometers (NASA 2003). .................................................................................................................... 43
Table 6.2. LTAP gain rules for Landsat ETM+ (modified after NASA 2003). .............. 47
Table 6.3. ETM+ Radiance Dynamic Range (wm-2sr-1µm-1) (modified after NASA 2003). .................................................................................................................... 48
Table 6.4. ETM+ low-gain and TM Radiance Dynamic Range (w•m-2•sr-1•µm-1) (modified after NASA 2003). ............................................................................... 51
Table 7.1. Constraints applied to the selection of suitable Landsat scenes. ..................... 60
Table 7.2 Metadata summary for the selected Landsat scenes. ........................................ 61
Table 7.3. Solar energy path (modified after Schott 1997)............................................... 70
Table 7.4. Manually selected DOS values based on the analysis of dark targets present in the five scenes. The values are represented as digital numbers. ........................... 78
Table 7.5. Atmospheric scattering models (after Chavez 1988)....................................... 79
Table 7.6. Scattering function coefficients obtained for each band in different atmospheric conditions (methodology after Chavez 1988). ................................. 79
Table 7.7. Procedure to predict scattering from TM band 1 reference DOS (methodology after Chavez 1988)................................................................................................ 80
Table 7.8. Observed DOS and final predicted DOS values.............................................. 81
Table 7.9. Post-calibration parameters (after Irish 1998). ................................................ 83
Table 7.10. Mean Solar Exo-atmospheric irradiances (Wcm-2µm-1) (values after Irish 1998). .................................................................................................................... 84
Table 7.11. Earth-Sun distance (astronomic units AU) and Sun elevation (degrees) at the time of acquisition (distance d calculated after Irish 1998). ................................. 84
xii
Table 8.1. Maximum in-scene reflectance registered by the TM and ETM+ sensors for the five subscenes. Note: Reflectance values greater than 100% are possible due to path radiance and background effect illumination............................................ 86
Table 8.2. Percent saturated pixels for each band of the subscenes. ................................ 87
Table 9.1. Snow/ice area measured from the zonal statistics of the binary classification. ............................................................................................................................... 98
Table 9.2. Inter-scene zonal statistics illustrating gain, losses and total net balance. ...... 99
Table 9.3. Inter-scene snow/ice cover percent changes. ................................................... 99
Table 9.4. Snow/ice area measured from the zonal statistics of the minimum distance classification. .......................................................................................................100
Table 9.5. Inter-scene zonal statistics from minimum distance statistics illustrating gain, losses and total net balance. ................................................................................ 102
Table 9.6. Inter-scene snow/ice cover percent changes determined from minimum distance classification. ........................................................................................ 102
Table 9.7a. Change detection statistics from 1987-1989................................................ 103
Table 9.7b. Change detection statistics from 1989-1997................................................ 104
Table 9.7c. Change detection statistics from 1977-2000................................................ 105
Table 9.7d. Change detection statistics from 2000-2001................................................ 106
Table 9.8. Detail of class areal changes (sq. km.)........................................................... 107
Table 10.1. Inter-scene snow/ice cover percent loss....................................................... 116
Table 10.2. Inter-scene snow and ice cover change estimates for Chacaltaya Glacier. . 117
Table 10.3. Inter-scene snow and ice cover change estimates for Zongo Glacier.......... 118
Table 10.4. Inter-scene snow and ice cover change estimates for Laguna Glaciar. ....... 118
Table A.1. Rugged terrain subset.................................................................................... 132
Table A.2. Ancohuma-Illampu glaciers.......................................................................... 132
xiii
List of figures Figure 2.1. Schematic representation of the profile of a glacier. Net balance is positive in
the accumulation zone and negative in the ablation zone. Ice movements transfer mass from the accumulation to the ablation zone constantly reshaping the profile of the glacier (modified after Bennett and Glasser 1996)....................................... 4
Figure 2.2. Ice sample extracted from the coring tool by scientist of Thompson’s research group studying the retreating glaciers of the Kilimanjaro (Tanzania) (NOAA 2002). .................................................................................................................... 12
Figure 3.1. Spatial profile of the Andes in proximity of the study area (elevation data from EDC 2003). (Vertical scale is exaggerated)................................................. 17
Figure 3.2. Bolivia and relative location of the area of study area of study (datasets from EDC 2003). ........................................................................................................... 19
Figure 3.3. Area of study. Landsat image of the Cordillera Real. .................................... 20
Figure 4.1. During La Niña years, the wind pattern is dominated by easterlies rich in moisture collected over the Amazon basin. Easterlies favor snow precipitation and accumulation during the austral summer months. During El Niño events, low moisture westerlies are predominant; the rain shadow effect of the Andean coastal ranges reduces precipitation (Francou, Vuille, Wagnon, Mendoza and Sicart 2003). .................................................................................................................... 25
Figure 6.1. Physical states of H2O (after Martin 2004).................................................... 35
Figure 6.2. Spectral reflectance of snow. Each curve represents the spectrum for a different grain size. The dashed line indicates the center (1.03 µm) of a characteristic absorption band extending from 0.96 µm to 1.08 µm (modified after Nolin and Dozier 2000). ....................................................................................... 38
Figure 6.3. Modeled spectral reflectance of glacier facies (modified after Hall et al. 1988, Zheng et al. 1984). ................................................................................................ 39
Figure 6.4. Mie Scattering (modified after Nave 2000).................................................... 41
Figure 6.5. Relative spectral responses and spectral differences between TM (grey) and ETM+ (black) spectrometers (modified after NASA 2003). ................................ 45
xiv
Figure 6.6. LTAP gain rules for the month of august. The rules depend on expected landcover, seasonality, and illumination. (after NASA 2003).............................. 47
Figure 6.7. Optimization of radiometric resolution based on ETM+ Radiance Dynamic Range adjustments (after NASA 2003). ............................................................... 48
Figure 6.8. Modeled maximum reflectance detected by Landsat 7 ETM+ P01R71 06-26-00 for low- gain mode and high-gain mode.......................................................... 50
Figure 6.9. ETM+ low-gain and TM Radiance Dynamic Range (w•m-2•sr-1•µm-1) (modified after NASA 2003). ............................................................................... 51
Figure 6.10. Maximum reflectance detected by Landsat 5 TM P01R71 06-10-97 (L5) and Landsat 7 ETM+ P1R71 06-26-00 (L7). The ETM+ is preset by LTAP on high-gain mode (HHHHHH) (modified after NASA 2003). ........................................ 52
Figure 6.11. Modeled Directional-Hemispherical reflectance of deep snow (modified after Nolin and Dozier 2000). TM bandpasses are represented by light gray boxes overlapping the spectra. “r” represents grain size................................................. 53
Figure 6.12. Modeled Directional Hemispherical reflectance of glacier facies in TM bandpasses for the visible and NIR domains (modified after Hall et al. 1988, Zheng et al. 1984). ................................................................................................ 54
Figure 6.13. Integrated spectral directional-hemispherical reflectance of deep snow and glacial facies in TM bandpasses (modified after Zheng et al. 1984, Dozier and Mark 1987, Hall et al. 1988, Rosenthal et al. 1996). TM bandpasses are represented by light gray boxes overlapping the spectra. “r” represents snow grain size. ........................................................................................................................55
Figure 6.14. TM and ETM+ reflectance cut-offs (left) and resulting snow and ice signatures from modeled directional hemispherical reflectance (right). Refer to Figure 6.13 for snow and ice classes legend......................................................... 57
Figure 7.1 Relative position of the Cordillera Real within a Path01-Row71 Landsat scene. Approximate coordinates of the scene center are latitude 15.89S, longitude 67.99W...................................................................................................................59
Figure 7.2. ERDAS Image 8.6 Geometric Correction tool. Example of Ground Control Points (GCP) selected in proximity to the study area. .......................................... 64
Figure 7.3. Geometric correction and co-registration of Landsat TM Path 01 – Row 71 8/2/1987. Before (left) and after (right). The original image was a L1R product. ............................................................................................................................... 65
xv
Figure 7.4. Landsat ETM+ Path 01 – Row 71 6/26/2000 and 7/31/2001. Spatial offset at La Paz Airport (approx. 200 m). Both scenes were originally projected in UTM 19S WGS84. ......................................................................................................... 65
Figure 7.5. Glaciers of the Cordillera Real and detail of the Ancohuma-Illampu Massif. Area is calculated as planimetric surface. Glacier classification by Jordan (1999). ................................................................................................................................67
Figure 7.6. Details of the subset procedure based on Jordan’s glacier inventory (1999). ............................................................................................................................... 68
Figure 7.7. Solar energy path (modified after Schott 1997). ............................................ 69
Figure 7.8. Effects of atmospheric transmission and path radiance on resulting total at-sensor reflectance (modified after Schott 1997). .................................................. 71
Figure 7.9. MODTRAN 3. Example of longwave atmospheric transfer model output from standardized input (http://geosci.uchicago.edu/~archer/cgimodels/radiation.html). ............ 75
Figure 8.1. Maximum in-scene reflectance registered by TM and ETM+ sensors in the five subscenes. ...................................................................................................... 86
Figure 8.2. Percent saturated pixels for each band of the subscenes. .............................. 87
Figure 8.3. Accumulation (Zone III and Zone II) and ablation (Zone I) zones of Zongo Glacier as identified by Klein and Isacks (1999).................................................. 89
Figure 8.4. Sample of training sites used to obtain the spectral signatures for the information classes................................................................................................ 89
Figure 8.5. Spectral signatures obtained from the training sites on Zongo Glacier.......... 90
Figure 8.6. Binary classification of Zongo Glacier using the snow/ice mask after Dozier (1989) applied to four Landsat scenes. ................................................................. 92
Figure 8.7. Minimum distance supervised classification of Zongo Glacier from Landsat TM 1987................................................................................................................ 93
Figure 8.8. Minimum distance supervised classification of Zongo Glacier from Landsat ETM+ 2000........................................................................................................... 93
Figure 8.9. Accumulation and ablation endmembers of Zongo Glacier, and AB transect (figure 8.10). The fraction is represented in grey scale with white as the highest and black as lowest . ............................................................................................. 95
Figure 8.10. Accumulation and ablation endmember fractions along the AB transect (figure 8.9). ........................................................................................................... 96
xvi
Figure 9.1. Total snow/ice cover of the Cordillera Real measured by binary classification. ............................................................................................................................... 98
Figure 9.2. Total snow/ice cover of the Cordillera Real measured by minimum distance classifications...................................................................................................... 101
Figure 9.3. Surface changes in km2 for each scenes pair obtained from the minimum distance classification. ........................................................................................ 107
Figure 9.4. Percent change from ablation class to other classes. .................................... 108
Figure 9.5. Comparison of the total snow/ice cover of the Cordillera Real measured by binary and minimum distance classifications. .................................................... 109
Figure 9.6. Laguna Glaciar: Comparison of the total snow/ice measured by binary and minimum distance classifications. ...................................................................... 111
Figure 9.7. Zongo Glacier: Comparison of the total snow/ice measured by binary and minimum distance classifications. ...................................................................... 112
Figure 9.8. Chacaltaya Glacier: Comparison of the total snow/ice measured by binary and minimum distance classifications. ...................................................................... 113
Figure 10.1. ASTER scene (left) acquired on 06/29/2001 and Landsat ETM+ scene (right) acquired on 07/31/2001 representing the central portion of the Cordillera Real after an extended snow fall event (USGS Global Visualization Viewer, http://glovis.usgs.gov). ............................................................................................. 115
Figure A.1. Snow and ice cover sample of approximately 59 km2................................ 129
Figure A.2. Soils and vegetation sample of approximately 43 km2............................... 130
Figure A.3. Soils and vegetation sample before and after topographic normalization... 132
Figure A.4. Snow and ice cover sample before and after topographic normalization.... 133
Figure A.5. Regression analysis for the soil and low vegetation subset (band TM 3) before normalization. .......................................................................................... 134
Figure A.6. Regression analysis for the soil and low vegetation subset (band TM 3) after normalization. ..................................................................................................... 135
Figure A.7. Regression analysis for the snow and ice cover subset (band TM 3) before normalization. ..................................................................................................... 136
Figure A.8. Regression analysis for the snow and ice cover subset (band TM 3) after normalization. ..................................................................................................... 136
xvii
Figure A.9. Regression analysis for the snow and ice cover subset (band TM 5) before normalization.... 137
1
Chapter 1
Introduction
Fluctuations of mountain glaciers and ice caps are key variables for the
observation of climate-related changes and the detection of enhanced greenhouse effects
(IPCC 2001). The World Glacier Monitoring Service (WGMS) and the Global Land Ice
Measurements from Space (GLIMS) projects are currently coordinating a world-wide
research program to continue monitoring glaciers in areas already represented by more
than a century of observations and to obtain direct and inventory data from a number of
under-represented locations, such as the Andes. These areas represent a critical gap in the
existing research network. The analysis of spatially and temporally distributed
glaciological data will contribute to a better understanding of the driving forces and
processes involved in global climate changes.
Glaciers and ice caps develop under a narrow range of climate conditions (defined
temperature range, humidity and precipitation), showing strong links with atmospheric
circulation changes and climate variables. Monitoring the distribution, intensity, and rate
of change of ice masses is of utmost importance in delineating spatial and temporal
variations of the Earth’s climatic system.
The purpose of this research is to analyze snow and ice cover changes of the
glaciers of the Cordillera Real (Bolivia) during the period 1987-2001. The study is based
2
on satellite remote sensing technology (Landsat TM and ETM+), which is able to provide
a regional synoptic view of the study area and discrete data coverage during the defined
period of time. This research seeks to enhance the regional understanding of trends and
dynamics of the glacier retreat observed on benchmark glaciers such as Zongo and
Chacaltaya. This research will highlight the advantages and disadvantages of remote
sensing as a tool to generate relatively fast and inexpensive information about glacial
advance or retreat. Another potential outcome of this research (but not further
investigated) is the development of water resource management policies and planning for
the study area. In fact, glacial melt-water from the Cordillera Real provides vital fresh
water resources for local agriculture, the hydroelectric industry, and the large human
population present in the area. The rate of retreat and glacial mass loss are of vital
concern for the planning and provision of future alternative water resources. Thus,
detecting and monitoring snow and ice mass variations in the Bolivian glaciers are not
only important for delineating spatial and temporal variations of the Earth’s climatic
system, but also for fresh water resource management necessary for sustainable human
living and power generation (Bindschadler, Dowdeswell, Hall, and Winther 2001).
This research seeks to: (1) provide an overview of the climatic variables that
affect tropical glaciers in the Bolivian Andes; (2) describe the most valuable research
methodologies to study glaciers and ice-caps; (3) investigate the application of remote-
sensing based measurements for the development of glacier inventories and the
estimation of regional mass balance variations; and (4) apply a set of remote-sensing
based techniques to Landsat TM and ETM+ imagery to estimate regional snow and ice
cover variations of the Cordillera Real from 1987 to 2001.
3
Chapter 2
The Study of Glaciers: Methodology Overview
By definition, a glacier (or ice-cap) will form from the snow accumulated and
compacted to ice where the rate of accumulation exceeds the rate of ablation.
Accumulation occurs most commonly as a result of snow, hail, and frost precipitation;
avalanches and secondary processes may increase the accumulation rate locally. Ablation
can occur by melting and/or sublimation. The balance between accumulation and ablation
defines the mass balance of a glacier; mass balance is largely dependent on climate
(Bennett and Glasser 1996). Accumulation dominates the upper region of a glacier where
temperatures are constantly (or mostly) below freezing, while ablation dominates the
glacier terminus where temperatures are generally at, or above, the freezing point.
The spatial imbalance between accumulation and ablation determines the surface
morphology and drives ice gravitational flow (Bennett and Glasser 1996). The net
balance gradient is defined as the increase in net balance (accumulation minus ablation)
with altitude; moving from lower to higher elevation, the accumulation of ice exceeds
loss of ice at a greater rate (figure 2.1). The equilibrium of the surface slope of a glacier is
a function of the net balance gradient (Bennett and Glasser 1996). As a result, net balance
gradient is steepest on glaciers that experience warm damp maritime climates, and lowest
in cold dry continental areas.
4
Mass balance measurements are based on the study of glacier inputs and outputs.
Measurements can be achieved with direct and indirect methodologies, but even with the
most accurate procedure, the results are based on interpolations and estimations.
Braithwaite (2002) explained that glacier mass balance forms an important link between
the atmospheric environment and glacier dynamics and hydrology. Glacier mass balance
changes influence oceanic mass and, thus, are of global impact. Locally, changes in ice
mass affect water resources and hydrology (such as surging floods or irreversible drought
in the case of complete melting).
Haeberli, Frauenfelder, Hoelzle and Maisch (1999) emphasized that the final
purpose of mass balance studies is to gain enough understanding to be able to
parameterize unmeasured glaciers and describe ongoing global changes. The following
mass balance measurement techniques are the most frequently used in research programs.
Accumulation wedge(positive net balance)
Ablation wedge (negative net balance)
Equilibrium line
Ice movement
Figure 2.1. Schematic representation of the profile of a glacier. Net balance is positive in the accumulation zone and negative in the ablation zone. Ice movements transfer mass from the accumulation to the ablation zone constantly reshaping the profile of the glacier (modified after Bennett and Glasser 1996).
5
On-site Measurements
On-site measurements produce the most accurate and significant data. Accurate
measurements of mass variation over time and space are essential for the identification of
those processes that modify glacier dynamics. Through direct measurement, it is possible
to collect reliable information to be used as benchmarks to calibrate mass-balance and
energy-balance models. The disadvantages of on-site measurements are mainly the high
demand of resources (technical and human) and the high cost of long-term operations in
extreme environments (Haeberli et al. 1999).
Mass balance measurements can be achieved by on-site surveys of mass density,
volumetric change (geometric), and flow rates. This methodology relies on the
installation of spatially distributed survey stations over the glacier surface to sample snow
and ice thickness variation and ice flow rates throughout a specific period of time; the
survey stations use graduated borehole stakes and snow pits. Densimetric data are
obtained from ice and snow cover samples. Using on-site surveys, it is possible to obtain
direct measurements of the accumulation and ablation zones (especially with regard to
the surface extent), the equilibrium line (in particular its elevation shifts in time), the
thickness of the ice to bedrock, and the velocity vectors of flow (directions and
intensities).
Data collected from the survey stations are spatially interpolated to model the
glacial surface; the resulting surface provides the base to determine mass losses and gains
and the estimation of mass variation per unit of time and space. Nearby meteorological
stations provide data to correlate mass balance with atmospheric parameters (such as
temperature, pressure, moisture, and wind) and the opportunity to monitor glacier
6
response to changes in climatic patterns. Examples of direct mass balance measurements
are extensive, because this methodology was the first applied to the study of glacier
dynamics (Braithwaite 2002).
Another form of direct measurement is to record data relative to the energy fluxes
that control melting and freezing processes. Sensors measuring incoming and outgoing
long- and short-wave radiation, albedo, and latent/sensible heat fluxes (to mention a few
variables) are installed as part of meteorological stations. Mass variation is estimated
from energy fluxes over time and space, and requires complex glacier-climate models.
The models take into consideration variation of mass balance in response to temperature
changes. An important parameter considered in these models is the climate sensitivity of
mass balance (∆balance/∆T), which is the basis for static and dynamic models. Static
models state that temperature changes are reflected by instantaneous balance adjustment
(such as the models used by IPCC’s Sea Level Panel); dynamic models consider the
delayed responses in thickness and area adjustment induced by temperature variations.
One of the most commonly used models of sensitivity is the degree-day model.
As Braithwaite and Zhang (1999) explained, the +1 degree temperature increase
represents a suitable standard for comparing sensitivities from different glaciers and
models applied to the same glacier. The degree-day model takes into account the spatial
distribution of mass balance over the entire glacier area, accounting for the mass balance
variations in elevation bands.
Some examples of degree-day melting factors (Braithwaite and Zhang 1999) are:
8.0 mm day-1 deg-1 for ice and 4.5 mm day-1 deg-1 for snow (millimeters of snow and ice
melting each day for +1°C increase in temperature). The degree day melting factors are
7
highly dependent upon local climatic and morphologic settings. Once the model is
calibrated, it is possible to associate climate changes with mass balance variations and
vice versa. While requiring fewer human resources than direct mass balance
measurements, energy balance and sensitivity models require costly instrumentation and
data processing. As is the case for the direct mass balance measurements, energy balance
methods require long-term continuous climatic and energy flux measurements over the
glacier. Energy balance methods provide a better description of the physical processes
involving climate in mass balance variations.
Indirect Measurements
Indirect mass balance estimation methods rely on remote sensing technology and
ancillary databases. Indirect measurements and estimates are highly cost-effective, but
due to the nature of the data are more prone to measurement error and biases, and are thus
best applied in studies of remote areas, for large area inventories, and for the
development of inventories adjusted and calibrated on benchmarks from direct mass
balance measurements (proxies).
One possible compromise to costly on-site measurements is mass balance
estimation from topographic surveys, which represents a way to estimate thickness
variations and infer flow vectors from superficial flow. The advantage of this method is
that it is possible to rely on ancillary data such as topographic maps generated from
airborne stereo-pairs, digital elevation models (DEM) produced from satellite/airborne
sensors, and GPS/geodetic on-site control points. The method compares topographic
variations over a period of time, which leads to the estimation of volumetric changes and
8
mass balance variations. The method relies on the accuracy and precision of topographic
elevation measurements, which cannot always be estimated for historical topographic
data.
The topographic measurement methodology is of value for the estimation of
regional mass balance variations where direct methods are not cost effective, due to the
extent of the study area, the lack of research resources, or the lack of historical on-site
mass measurements. On the other hand, the area of study must be well covered by
topographic surveys (such as from military agencies, or private and public
airborne/satellite sensors).
Modern technologies (especially SAR-SRTM, LIDAR, and more recently,
ASTER) are improving the accuracy of elevation measurements, thus increasing the
reliability of this methodology. Topographic methods have been successfully applied on
single glaciers (Rivera and Casassa 1999; Lange, Araos and Rivera 2003; Rivera,
Benham, Casassa, Bamber and Dowdeswell 2003) and for the estimation of extensive
regional mass balance (Rignot, Rivera and Casassa 2003; Schnirch, Schneider, Casassa
and Kilian 2003; Rivera et al. 2003; Bamber 2003; Keller, Casassa, Rivera, Forsberg and
Gundestrup 2003).
When topographic data are not available or have limited availability, mass
balance and mass variation can be estimated from surface area changes (indicated
primarily by the retreat or advance of the terminus). Airborne photography and satellite
imagery can be used in conjunction with Geographic Information System (GIS) data to
map the spatial variation of glaciers and ice caps over time. While providing indications
of change, surface area variations do not always represent mass variation in a linear
9
fashion; while easy to detect, surface changes present considerable difficulties of
interpretation (Haeberli 2003).
The accumulation area ratio (AAR = Accumulation area / Total area; an AAR=0.7
corresponds to mass equilibrium at the end of the melting season) and the equilibrium
line altitude (ELA) are two identified parameters that can provide researchers a better
understanding of the interrelation of area and balance variation. The two parameters are
directly connected: the boundary between the accumulation and the ablation zone is
defined by the equilibrium line. The ELA is a function of temperature, and thus may be
considered covariant to mass balance changes (also a function of temperature). Shifts of
the ELA are reflected in changes of surface extent of the accumulation and the ablation
zones.
Bennett and Glasser (1996) discussed how ablation and accumulation zone
geometry is the result of dynamic processes that work to maintain a constant slope; these
processes are functions of climatic conditions. The AAR is thus a property of a glacier in
specific climatic conditions and slope equilibrium: any variation of the AAR through
time reflects rearrangements in the net balance gradient, and thus mass balance.
Numerical estimates of mass balance from the AAR are based on proxy data and
regression from neighboring glaciers. The AAR and the ELA are relatively easy to
determine through remote sensing and allow regional estimates of mass balance variation
once calibrated with measurements on benchmark glaciers.
The error of estimate determined through remote sensing is a complex function of
proximity, geomorphology, and local climatic conditions between the study area and the
benchmark glaciers. A considerable number of assumptions must be taken into
10
consideration, but the net advantage of the method is that it provides an acceptable
estimate of mass balance variations relatively quickly, inexpensively, and on a regional
scale. Remote sensing based estimates reduce human resource requirements, and costs.
These factors compensate for the loss of accuracy. Also, it is the only methodology that
can reasonably be applied for preliminary estimates on remote or unexplored areas such
as alpine glaciers.
Remotely sensed imagery can provide other important glaciological information.
Skvarca, Stuefer and Rott (1999), for instance, successfully determined flow velocity
vectors from the observation of change of shape of surface debris from rock falls. These
measurements are of utmost importance for verifying the amount of mass displaced by
flow. The rate of change of flow velocity is a function of dynamic processes that maintain
slope equilibrium within the glacier, and thus is associated with mass balance variation
(in particular, mass displacement) and climatic variability.
A more sophisticated indirect estimate of mass balance relies on the integration of
topographic measurements (such as thickness variation from DEM and topographic
maps, as previously discussed) and surface change detection. Haeberli and Hoelzle
(1994) presented a model based on a series of easily calculated glacier parameters to
estimate mass balance. The estimate is based on four simple geometric parameters that
are commonly found in detailed inventory data: length of the glacier, maximum and
minimum altitude, and total surface area of the investigated glacier. From these
parameters, mean altitude, vertical extent, and average surface slope are derived.
Analysis involves estimation of total volume, glacier thickness, and the mean flow
velocity, which are obtained from empirical models. Glacier length changes for given
11
changes in mass balance are also related to a characteristic dynamic response time. The
time interval between first reaction and full response is called relaxation time (Haeberli
and Hoelzle 1994).
Haeberli and Hoelzle (1994) argued that the simplicity of the model is reflected in
a number of uncertainties. While the methodology presents inherited error of calculation
of derived flow velocities, response time, and other factors, the results are acceptable to
determine mass balance for medium to large glaciers. The authors suggested that
empirical and regional approaches must be investigated to tune the parameters of the
model for the different sites and conditions of mountain ranges around the world.
Haeberli and Hoelzle (1994, 11) stated that the most important potential benefit of
the methodology (and in general of all indirect methodologies) “is given with the
possibility of quantitatively inferring average decadal mass balance for unmeasured
glaciers by analyzing cumulative length changes from field evidence;” the field evidence
includes moraines, satellite imagery, aerial photography, and long-term observation. In
addition, Haeberli and Hoelzle (1994, 11) stated that regional inventories open new
opportunities “to provide the basis for evaluating scenarios of global warming.”
Coring
Borehole surveys and ice coring represent other important methodologies for
studying glaciers and ice caps. Coring focuses on the analysis and reconstruction of
climatological and paleo-ecological information preserved in glaciers and ice caps.
Thompson and his research team at Ohio State University (Thompson 2000; Thompson
and Davis 1998) are probably the most accomplished scientific group using ice coring.
12
Since the beginning of the 1980s, Thompson has undertaken coring survey expeditions
on a considerable number of glaciers and ice caps worldwide (figure 2.2). The purpose of
his research is to develop a global climate change chronological map suitable for
analyzing the changes in time and the pattern of changes in space of the Earth’s climate
(Thompson 2003).
Climatological and atmospheric information are retrieved from ice cores,
primarily through the analysis of particulate matter and isotopes trapped in the snow
precipitation and subsequently diagenized in ice. Best results are obtained from relatively
stable glaciers, where melting processes do not obliterate the climatic records.
Ice coring is suitable for the reconstruction of time series of temperature,
precipitation, particulate composition, and atmospheric composition; time series might
help determine if the global climatic change measured over the past 200 years is a
Figure 2.2. Ice sample extracted from the coring tool by scientist of Thompson’s research group studying the retreating glaciers of the Kilimanjaro (Tanzania) (NOAA 2002).
13
consequence of natural variability or the result of anthropic activity. Glaciological time
series derived from ice cores present a unique opportunity to understand glacier processes
as a function of global and regional climatic variability. Time series make it possible to
study glacier response to climate change and the relation between these changes at a
global scale. El Niño/Southern Oscillation (ENSO) events, for instance, present a
characteristic signature that has been identified world wide in ice cores. Time series have
shown that the magnitude and frequency of ENSO events have a strong influence on
glacier dynamics at a global scale.
Global climatic maps based on time series from ice cores eventually might help to
identify those mechanisms behind events such as ENSO. In addition, the development of
models to predict future variability as a function of natural patterns and anthropic
perturbation on climate might become possible.
14
Chapter 3
Study Area
Exploration and Historical Data
The glaciers of Bolivia were first described by d'Orbigny (1835-1847); no
previous record of the physiography of the glaciers is available from the native
populations, mostly due to the lack of a native writing system (Jordan 1999). During the
early 20th century, scientific expeditions aimed at “eliminating blank places on maps of
the Earth” (Jordan 1999) produced early maps of the Bolivia’s glaciated mountains.
These early research projects were directed toward the identification and mapping of
snowline and glacier terminus locations, the understanding of glacier morphology, and
the distribution of glaciers. Jordan (1999) listed some of the important researchers of the
time including Conway in 1900, Hauthal in 1911, and Herzog between 1913 and 1915. In
1927-28, a successful climbing expedition led by Carl Troll of the German-Austrian
Alpine Association added important cartographic details to the available maps of the area
using terrestrial photogrammetric methods. Troll specialized in photographs and
triangulation measurements; he was able to produce a number of general and detailed
topographic maps of the northern Cordillera Real (Illampu area) at a scale of 1:50,000
(Troll and Finsterwalder 1935). Later, geological studies provided general maps of the
distribution of glaciers.
15
The most recent topographic map of the Illampu area was developed from aerial
photos taken in 1963 and 1975 (Finsterwalder 1987). The accuracy of the topographic
maps from Troll and Finsterwalder (1935) and Finsterwalder (1987) allowed an initial
quantification of mass balance. At the beginning of the 1990s, Jordan and Finsterwalder
produced another updated cartographic map of the southern Cordillera Real (Illimani
area) at a scale of 1:50,000 (Finsterwalder 1990; Jordan and Finsterwalder 1992).
Only part of the La Paz region and the Cordillera Occidental has been covered by
published maps as of early 2004 (Jordan 1999). In the 1940s, the U.S. Army Map
Service, and later the Instituto Geográfico Militar (IGM) in La Paz, started official
mapping of Bolivia based on vertical aerial photographs; the maps have scales of
1:50,000 and 1:250,000. These maps are unsuitable for glacier studies, because no
distinction is made between snow patches and glaciers (Jordan 1999). However, the
original vertical aerial photographs are suitable for analyzing distribution of glaciers as
they existed at the time of the survey.
In 1967, Mercer conducted surveys of the glaciers of Bolivia using mass-balance
data and energy records, including measurements of ice movement, precipitation,
temperature, evaporation, ablation, and glacier runoff. Mercer charted his observations
from direct measurements and terrestrial photogrammetric photographs. In 1980, Mercer
participated in the compilation of a glacier inventory for Bolivia (Jordan, Brockman
Fernandez, Alvarez and Jacobsen 1980); the inventory was based on aero-triangulations
of Bolivia's glaciated regions carried out from the northern Cordillera Real across the
Cordillera Apolobamba to the Peruvian border (Jordan 1991; Herrmann 1993). The data
are now part of the "World Glacier Inventory" of the United Nations Environment
16
Program/United Nations Educational, Scientific, and Cultural Organization/International
Commission on Snow and Ice (UNEP/UNESCO/ICSI), now the World Glacier
Monitoring Service in Zürich, Switzerland (Jordan 1999).
The first valuable regional image of Bolivian glaciers was taken by the Gemini 9
space flight in June, 1966. The 1972 launch of Landsat 1 carrying the Multi Spectral
Scanner (MSS) marked the beginning of a new era for remote sensing studies. The
glaciers of Bolivia have been regularly observed from space since that time.
Area of Interest
Glaciers in Bolivia are found in two main ranges of the Andes, the Cordillera
Occidental (latitude 18°03' S to 18°25' S, longitude 68°53' W to 69°09' W) along the
western border with Chile, and the southern extension of the Cordillera Oriental (figure
3.1) extended from latitude 14°37' S to 17°04' S and longitude 67°13' W to 69°14' W.
Bolivia lies completely within the tropics, so glaciers are confined to the highest peaks of
the Andes Mountains (Francou 1993) as ice caps, valley glaciers, and mountain glaciers.
Most of the glaciers in Bolivia are found on the Cordillera Oriental, with 1,826 glaciers
covering 591.6 km2 (Jordan 1999); this region includes the Cordillera Apolobamba,
Cordillera Real, Cordillera Tres Cruces, Cordillera de Muñecaus, and the Nevado Santa
Vera Cruz (table 3.1). Annual precipitation within Bolivia is variable, but generally
decreases toward the southwest, where it reaches its seasonal minimum (Jordan 1999);
south of latitude 18°30' S even the highest peaks (6,000 m and above) cannot sustain
glaciers due to the lack of precipitation. The southern limit of glaciers roughly matches
17
the northern line of the large salars (salt pans), such as Salar de Coipasa and Salar de
Uyuni in the central plateau (Altiplano) of the Andes (Jordan 1999).
The Altiplano (extending about 200 km wide between the Cordillera Occidental
and the Cordillera Oriental) does not have sufficient elevation or precipitation to sustain
glaciers. This research focuses on the glaciers of the Cordillera Real (latitude 15°45' S to
16°40' S, longitude 67°40' W to 68°34' W), a section of the Cordillera Oriental (figure 3.2
and figure 3.3). The total glaciated area surveyed on November 1, 1984 (Jordan 1991)
amounted to 323.603 km2 (964 glaciers), corresponding to 54% of the glaciated surface
area of the Cordillera Oriental (53% of the number of glaciers). At the time of the survey,
the glaciers extended from a maximum elevation of 6,436 m to a lowest terminus
elevation of 4,420 m. The entire area is included in a single Landsat scene (MSS, TM and
ETM+ Path 01-Row 71). The study of the glaciers of the Cordillera Real is particularly
important due to the presence of a large population concentration (especially in the La
Paz district) that relies on the melt-waters of the glacier as a fresh water resource. It is
important to understand local glacier variations in relation to global change, but the
analysis of the historical evolution, trends, and pattern of the glacier mass balance of the
Cordillera Real also might help in the development of planning policies and mitigation
projects to optimize administration of water resources in the area.
Ele
vatio
n (m
)
Pacific Ocean Amazon Basin
Cordillera Occidental
Cordillera OrientalAltiplano
Ele
vatio
n (m
)
Pacific Ocean Amazon Basin
Cordillera Occidental
Cordillera OrientalAltiplano
Figure 3.1. Spatial profile of the Andes in proximity of the study area (elevation data from EDC 2003). (Vertical scale is exaggerated).
18
Area
(sq.
km
)Pe
rcen
tN
umbe
r of
glac
iers
Perc
ent
CO
RD
ILLE
RA
OR
IEN
TAL
14°3
7'-1
7°04
'67
°13'
-69°
14'
591.
600
100.
0018
26.0
0010
0.00
6436
.043
11.0
CO
RD
ILLE
RA
AP
OLO
BA
MBA
14°3
7'-1
5°04
'68
°58'
-69°
14'
219.
804
37.2
065
2.00
036
.00
6059
.043
11.0
Cha
upi O
rko
14°4
0'69
°10'
129.
357
21.9
034
6.00
019
.00
6059
.043
65.0
Col
olo
14°5
0'69
°06'
43.0
727.
3013
5.00
07.
5057
74.0
4311
.0U
lla K
haya
15°0
0'69
°03'
47.3
758.
0017
1.00
09.
5056
69.0
4390
.0C
OR
DIL
LER
A D
E M
UÑ
EC
AS
15°2
0'-1
5°38
'68
°33'
-68°
55'
0.68
40.
1016
.000
1.00
5237
.048
28.0
Mor
ocol
lu15
°20'
68°5
5'0.
148
0.03
8.00
00.
5051
56.0
4828
.0C
uchu
15°3
8'68
°33'
0.53
60.
108.
000
0.50
5237
.048
86.0
CO
RD
ILLE
RA
REA
L15
°45'
-16°
40'
67°4
0'-6
8°34
'32
3.60
354
.70
964.
000
53.0
064
36.0
4420
.0
NO
RTH
ER
N C
OR
DIL
LER
A R
EA
L15
°45'
-16°
20'
68°0
1'-6
8°34
'26
2.76
644
.40
784.
000
43.0
064
36.0
4420
.0Ill
ampu
-Anc
ohum
a15
°50'
68°3
0'10
3.09
917
.40
147.
000
8.00
6436
.044
38.0
Cal
zada
-Chi
aroc
o- C
hach
acom
ani
16°0
0'68
°20'
94.0
7215
.90
251.
000
14.0
061
27.0
4676
.0N
igru
ni-C
ondo
riri
16°0
8'68
°15'
40.8
686.
9024
1.00
013
.00
5752
.044
20.0
Saltu
ni-H
uayn
a Po
tosí
16°1
5'68
°08'
14.5
042.
5050
.000
3.00
6088
.048
04.0
Zong
o-C
umbr
e- C
haca
ltaya
16°1
8'68
°05'
10.2
231.
7095
.000
5.00
5519
.045
78.0
SO
UTH
ER
N C
OR
DIL
LER
A R
EA
L16
°20'
-16°
40'
67°4
0'-6
7°58
'60
.837
10.3
018
0.00
010
.00
6414
.044
99.0
Ham
patu
ri-Ta
ques
i16
°26'
67°5
2'11
.685
2.00
70.0
004.
0055
48.0
4723
.0M
urur
ata
16°3
0'67
°47'
17.2
072.
9075
.000
4.00
5836
.045
92.0
Illim
ani
16°3
8'67
°44'
31.9
455.
4035
.000
2.00
6414
.044
99.0
CO
RD
ILLE
RA
TR
ES C
RU
CE
S (Q
UIM
SA C
RU
Z)16
°47'
-16°
09'
67°2
2'-6
7°32
'45
.276
7.70
177.
000
9.50
5760
.047
08.0
Cho
quet
anga
16°5
4'67
°22'
6.99
21.
2021
.000
1.00
5541
.048
12.0
Hig
h re
gion
of T
res
Cru
ces
16°5
6'67
°24'
38.2
846.
5015
6.00
08.
5057
60.0
4708
.0N
EV
AD
O S
AN
TA V
ER
A C
RU
Z17
°03'
-17°
04'
67°1
3'-6
7°14
'2.
233
0.40
17.0
001.
0055
60.0
4853
.0
CO
RD
ILLE
RA
OC
CID
ENTA
L18
°03'
-18°
25'
68°5
3'-6
9°09
'10
.000
100.
00--
--65
42.0
5100
.0
Nev
ado
Saja
ma
18°0
6'68
°53'
4.00
040
.00
----
6542
.051
00.0
Nev
ados
Pay
acha
ta18
°09'
69°0
9'4.
000
40.0
0--
--62
22.0
5500
.0C
erro
s Q
uim
sach
ata
18°2
3'69
°03'
2.00
020
.00
----
6032
.055
00.0
Low
est
term
inus
(m
eter
s)
Area of Interest
CO
RD
ILLE
RA
OC
CID
ENTA
L
CO
RD
ILLE
RA
OR
IEN
TAL
Gla
cier
Mou
ntai
n gr
oup
Latit
ude
(sou
th)
Long
itude
(w
est)
Hig
hest
el
evat
ion
(met
ers)
Tabl
e 3.
1. G
laci
ers o
f the
Bol
ivia
n A
ndes
as N
ovem
ber,
1st 1
984
(mod
ified
afte
r Jor
dan
1991
).
19
Figure 3.2. Bolivia and relative location of the area of study area of study (datasets from EDC 2003).
La Paz
Sucre
Santa Cruz de la Sierra
BoliviaLake Titicaca
70 W
65 W
10 S
15 S
Tropic of Capricorn
15 S
20 S
Area of Study
La Paz
Sucre
Santa Cruz de la Sierra
BoliviaLake Titicaca
70 W
65 W
10 S
15 S
Tropic of Capricorn
15 S
20 S
Area of Study
20
Cordillera Real
La Paz
LakeTiticaca
Landsat Path 01 – Row 71
Cordillera Real
La Paz
LakeTiticaca
Landsat Path 01 – Row 71
Figure 3.3. Area of study. Landsat image of the Cordillera Real.
21
Chapter 4
Climatic Conditions Affecting Tropical Glaciers of Bolivia
Tropical climates are characterized by a homogenous atmosphere, infrequent
frontal activity, and one or two distinct and pronounced precipitation seasons, thus
presenting a characteristic impact on tropical glaciers (Kaser 1999). By definition (Kaser
1999), a tropical glacier must be located within the tropics, have a daily temperature
variation that exceeds the annual temperature variation (thermal delimitation), and be
within the Inter-tropical Convergence Zone (ITCZ) oscillation area (hygric delimitation).
Another important boundary condition characteristic of tropical climates is the thermal
homogeneity which is reflected in a fairly stable elevation of the 0º C isotherm.
Kaser (1999) argued that, in the absence of seasonal thermal variability, the
vertical mass balance gradient on the tongues of tropical glaciers is weaker than that
observed on mid-latitude glaciers. As a result, accumulation area ratio (AAR) is higher
and the terminus is more sensitive to equilibrium line altitude (ELA) shifts; changes in
temperature affect ablation (due to sensible heat flux) and accumulation zone (due to
ELA shift).
Historically, tropical glaciers had the most recent maximum extent during the
Little Ice Age (LIA; estimated from 1350 to 1900), and have receded since the middle of
the 19th century. Starting in the 1930s, it was possible to observe a drastic glacier retreat
22
that slowed, but did not stop, during the 1950s to 1970s period. After the 1970s, the rate
of retreat has consistently increased; only a few locations have been marked by terminus
advance (Kaser 1999; Hooker and Fitzharris 1999; Haeberli and Hoelzle 1999; Rivera
and Casassa 1999; Thompson 1997; Thompson and Davis 1998; Thompson 2000). The
general trend of tropical glacier fluctuation agrees with the pattern of mid-latitude
glaciers, leading to the assumption that the processes behind observed change are of
global scale (Kaser 1999; Thompson 2000). This assumption becomes evident when
considering that air temperature and water vapor pressure, the two main factors affecting
tropical glacier fluctuations, are the result of the energetic state of the planetary climatic
system.
Kaser (1999) explained that in tropical glaciers temperature changes affect all
respective mass balance components uniformly, while air moisture changes have
different effects in dry and humid areas and seasons. As an example, an increase of
humidity during the dry season may reduce sublimation, but increase long wave radiation
with a general increase in negative mass balance; in humid areas (or during a wet season)
a reduction of humidity may cause mass loss due to increased insolation and decreased
precipitation. Due to the thermal stability of the tropical climate, it may be assumed that
fluctuations of tropical glaciers are associated with vapor pressure and air humidity
oscillation, thus making tropical glaciers a suitable benchmark for global change studies
(Kaser 1999).
In agreement with Kaser’s observations, Francou, Vuille, Wagnon, Mendoza and
Sicart (2003) highlighted the influence of precipitation variability and air humidity on the
mass balance of tropical glaciers. From a decade-long study of direct mass balance
23
measurements on Zongo and Chacaltaya glaciers in Bolivia, Francou et al. (2003)
observed that there was a close relationship between mass balance and atmospheric
condition during the austral summer (October to April), in particular from December to
February. During this period (the wet season), humidity and temperature are high and
melting prevails over sublimation. Net all-wave radiation is the dominant factor which
governs ablation. Albedo is a function of solid precipitation: a lack of precipitation (low
cloudiness) reduces the snow cover, exposing bare ice (characterized by lower albedo
properties), thus increasing energy available for melting processes. This process is
particularly important in the Bolivian tropical glaciers where the ablation and
accumulation season coincides with the austral summer. This is quite distinct from mid-
latitude glaciers, where there is a clear distinction between accumulation and ablation
seasons. The tropical austral summer presents the highest temperatures (thus energy
available for melting) and highest air humidity, resulting in the most intense snow
precipitation of the year.
Francou et al. (2003) argued that temperature is a key variable for long term
variations of tropical glaciers, because temperature governs all climatic fluxes on a global
scale. However, air humidity and precipitation patterns better predict mass balance
variation over short periods (annual/inter-annual) for the study area. Francou and his
colleagues observed that a lack of precipitation (reduced snow cover), reduced cloudiness
(low albedo), and increased net incoming solar radiation during the December-January-
February (DJF) months resulted in negative mass balance. During the dry winter months,
sublimation is the dominant ablation process, which is less effective than melting. If
abundant snow precipitation is available during the DJF months, the snow cover will
24
maintain high albedo during dry winters, reducing the energy available for melting.
Tropical dry winters represent a period of mass conservation. Francou et al. (2003)
showed that, in fact, the relationship between climatic/meteorological conditions and
mass balance variations is highly influenced by the summer season variability and, in
particular they found that mass balance and climatic/meteorological condition present a
more significant correlation when the dry season observations are removed from the
annual series.
Variability during the summer season has been associated with Sea Surface
Temperature (SST) oscillation (in this case of the western pacific coast of South
America) in conjunction with El Niño Southern Oscillation (ENSO) events. In particular,
glacial mass balance in the Bolivian Andes is highly associated with the increased
frequency of ENSO during recent decades (Thompson and Davis 1998; Kaser 1999;
Thompson 2000; Francou et al. 2003).
During La Niña years, the wind pattern is dominated by moisture rich easterlies
(moisture is collected over the Amazon basin), which favor snow precipitation and
accumulation during the austral summer months (figure 4.1). During El Niño events, low
moisture westerlies are predominant; the rain shadow effect of the Andean coastal ranges
reduces precipitation. Reduced precipitation and increased overall temperature negatively
affect the annual mass balance of Bolivian glaciers (Francou et al. 2003). The higher
frequency and changed spatial-temporal distribution of El Niño events since the 1970s,
and a general tropospheric warming, might explain the recent rapid glacial retreat in the
Bolivian Andes. Thompson and Davis (1998) showed similar results based on a
comparative analysis of ice cores (containing 25,000 years of paleo-climatologic data)
25
from Bolivian glaciers. Time series from ice cores have shown strong correlations
between drought in the Bolivian Andes and El Niño events; the effect is visible in the
reduced seasonal ice thickness and an overall increase in tropospheric temperature from
measurements of oxygen isotopes (18O).
El Niño drought episodes, corresponding to reduced or negative glacier mass
balance, are characterized by a high concentration of dust particulates, high temperatures,
and low precipitation. Ice cores also confirm the general historical trend of deglaciation
in the area, only temporarily delayed by the relatively colder episode of the LIA. Analysis
of the composition of particulate matter allowed Thompson and Davis (1998) to state that
during La Niña events the main source of moisture is localized over the Amazon basin, in
accordance with recent climatologic models.
Figure 4.1. During La Niña years, the wind pattern is dominated by easterlies rich in moisture collected over the Amazon basin. Easterlies favor snow precipitation and accumulation during the austral summer months. During El Niño events, low moisture westerlies are predominant; the rain shadow effect of the Andean coastal ranges reduces precipitation (Francou, Vuille, Wagnon, Mendoza and Sicart 2003).
Ele
vatio
n (m
)
Pacific Ocean Amazon Basin
Cordillera Occidental
Cordillera OrientalAltiplano
El Niño La Niña
Ele
vatio
n (m
)
Pacific Ocean Amazon Basin
Cordillera Occidental
Cordillera OrientalAltiplano
El Niño La Niña
26
Chapter 5
Remotely Sensed Change Detection of Snow and Ice
Remotely sensed imagery offers an unequaled possibility to study landforms from
a synoptic point of view, thus making remote sensing a suitable technology for placing
results from local research within a regional context. Four remote sensing technologies
can be applied to glaciological studies: visible (analog and digital), multi-spectral, radar,
and laser. All four sensors can be carried by two different vectors: airborne (low
elevation flying missions) and satellite (high elevation).
Airborne platforms (equipped with any of the four sensors) have the advantage of
providing very high spatial resolution, but the disadvantage of being relatively costly and
not suitable for extremely remote and unexplored areas. On the other hand, satellite
platforms offer a generally continuous coverage of the Earth’s surface. Satellite imagery
is a relatively low cost product, and the low resolution typical of older systems is being
replaced by modern high spatial resolution sensors. Within the four technologies, each
sensor is best suited for different research questions, but the digital format of modern
products allows the integration of heterogeneous sensors in order to increase the amount
of available information.
Imagery derived from visible wavelengths is mainly used for feature detection; in
glaciology, the imagery is used as base maps for change detection of landforms (such as
moraines), terminus position, areal distribution, and surface flow. Airborne photography
27
was the first remote sensing technology utilized in glaciology for local and regional
studies, and thus, there are numerous historical image databases that depict glaciers since
the invention of photography (Braithwaite 2002). These databases are of utmost
importance and, in some cases, allow researchers to extend change detection time series
back past the beginning of the 20th century.
Multispectral satellite imagery is the most direct descendent from airborne
photography. Originally used for military purposes, multispectral imagery is based on
optical sensors (mechanical and/or digital) that simultaneously record the visible and
infrared domains. Each sensor is sensitive to a particular electromagnetic frequency (or
range of frequencies) and thus provides specific information about the reflector. In fact,
depending on the physical properties of an object, electromagnetic energy is reflected,
absorbed, and transmitted in different ways for each wavelength of the spectrum.
Spaceborne multispectral sensors record the electromagnetic response (radiance) of an
object on the surface of the Earth irradiated by the incoming solar energy, assign to this
measurement a scaled digital value, and store the value in the corresponding pixel of the
image. In glaciology, multispectral imagery is essential for the detection of snow and ice
cover (at the local, regional, and global scales) and to distinguish between various types
of snow and ice facies; thus, multispectral imagery provides a suitable tool for change
detection studies.
Radar imagery is mainly applied as a geomorphologic tool; it provides selective
feature detection, especially based on relief and shape contrast. Most recently, multiband
radar sensors have opened new opportunities to approach classification and change
detection through non-optical methodologies. The most important application of radar
28
technology in glaciology is known as interferometry. This new technology is based on the
assumption that phase changes detected from multi-temporal radar scans of a specific
scene are a function of topographic or morphologic variations on the surface. As a result,
interferometry can be applied to detect velocity flow of the surface of glaciers (JPL 2003;
Rignot et al. 2003).
In recent years, especially in response to widespread use of GIS technology,
satellite imagery has been expanded and applied to develop a new class of remotely
sensed information: digital elevation models (DEMs) and terrain models (DTMs).
Historically, topographic maps and elevation models were created from stereo
photography sampled by airborne sensors. Most recently, new sampling techniques and
processing capabilities have allowed the development of DEMs and DTMs from virtually
any of the four sensor types (visible optical, multispectral, radar, and laser). Starting with
radar sensors, followed by laser altimetry (LIDAR), and currently by
optical/multispectral sensors such as the Advanced Spaceborne Thermal Emission and
Reflection (ASTER), it is possible to produce a DEM of virtually any place on the
Earth’s surface. The full extent of the new DEM potential will be known within a few
years, after the generation of time-based digital elevation series of the surface of the
Earth.
In the meantime, DEMs and DTMs allow geographers and glaciologists to detect
topographic changes by comparing digital information to historical topographic maps,
calculating volumetric or mass changes over time; this is particularly important (as
previously discussed) for the detection of mass balance variation over time. Differential
topographic measurements are the key to obtaining the best possible estimate of mass
29
balance from remote sensing imagery. While topographic changes and topographic
parameters (see Haeberli and Hoelzle 1994) are essential for the study of glacier
dynamics, it is also very important to be able to detect accumulation and ablation zone
extents and their variation over time (to calculate AAR), and to determine the shifts and
migration of ELA to quantify change detection of glaciers. This is the final purpose of
spectral classification of remotely sensed multispectral imagery.
Like every other substance on the Earth’s surface, snow and ice present a
characteristic response to electro-magnetic radiation. Unfortunately, apart from specific
global snow and ice satellite systems (with very low spatial resolution), the most common
multispectral sensors utilized for land cover change detection (Multi-Spectral Scanner
[MSS], Thematic Mapper [TM] and Enhanced Thematic Mapper [ETM+] on Landsat) do
not offer specific tools or observation sensors targeted for snow and ice. The best
possible classification results come from the integration of heterogeneous information
(different sensors) and ancillary data. Nevertheless, a set of theories and techniques have
been developed to classify the surface of glaciers using the available multispectral and
radar sensors. The main classification and change detection approaches found to be
effective for classifying snow and ice facies of glaciers include supervised and
unsupervised classifications based on spectral reflectance, band ratioing, and spectral un-
mixing.
Supervised and Unsupervised Classifications
Multispectral classification implements spectral pattern recognition algorithms to
identify groups of pixels that have common spectral characteristics (spectral signatures),
30
and then arranges the pixels in a finite number of information classes. Pattern recognition
algorithms require a set of training sites, or spectral classes; the training sites may be
selected manually (supervised methods), or automatically (unsupervised methods)
(ERDAS 1997).
Supervised classification is controlled by the analyst, who provides a set of
spectral classes from in-scene training sites or from spectral libraries. Unsupervised
classification is automated, but the user may control a set of general options, such as the
number of final information classes. This research aims to develop a classification based
on the recognition of physical properties of snow and ice facies of glaciers; for this
purpose, one of the most suitable methodologies is a supervised classification based on
spectral reflectance libraries. The spectral libraries for this research were derived from
the integration of previous research in the area (Klein and Isacks 1999), in-scene training
sites, and theoretical spectral signatures (Rosenthal and Dozier 1996).
Band Ratioing
Band rationing is a technique based on the development of indices from relative
reflectance ratios between different bands. Band ratioing removes (or reduces) spectral
differences resulting from variations of illumination and is thus useful for removing the
shadowing effect typical of high-relief glacial environments in alpine areas (Schott 1997).
Landsat TM and ETM+ imagery are used for land cover and land use change
detection, and have been widely implemented in glaciology to map temporal changes of
terminus position and glacial albedo factors. Within the seven TM (and ETM+)
wavelength bands, the visible bands (1, 2 and 3) tend to saturate on snow and ice fields.
31
These bands are used to detect visible features to isolate ice and snow from non-glaciated
areas using digital masking (Klein and Isacks 1999). Band 4 (near infrared) can be used
to distinguish between ablation and accumulation zones; however, to avoid illumination
variations, the best results can be obtained from TM4/TM5 band ratioing (Hall, Ormsby,
Bindschadler and Siddalingaiah 1987; Paul, Kaab, Maisch, Kellenberger, and Haeberli
2002). Gao and Liu (2001) reported that TM3/TM4 ratioing can be used to enhance blue
ice from snow, and TM3/TM5 ratioing is suitable to enhance snow grain size variations.
Normalized difference (ND) ratios represent a more sophisticated ratioing
technique and allow a better filtering of atmospheric and illumination effects (Schott
1997). Five main ND indices are currently used to classify glaciated areas.
1. 5252
TMTMTMTMNDSI
+−
= (equation 5.1)
known as Normalized Difference Snow Index (NDSI) (Dozier 1989; Hall, Riggs,
and Salomonson 1995; Hall, Foster, Verbyla, and Klein 1998);
2. 5353
TMTMTMTMNDSII
+−
= (equation 5.2)
known as Normalized Difference Snow/Ice Index (NDSII) (Dozier 1989; Xiao,
Shen and Qin 2001);
3. 5454
TMTMTMTM
+− (equation 5.3)
which is a grain size index for finer grains (higher values represent larger grains)
(Dozier 1989);
32
4. 4242
TMTMTMTM
+− (equation 5.4)
another grain size index but for all sizes (higher values represent large grains)
(Dozier 1989);
5. 2121
TMTMTMTM
+− (equation 5.5)
is used as contamination index (higher values represent cleaner snow) (Dozier
1989).
All these indices have the advantage of requiring little digital processing.
Spectral Un-mixing
One of the main problems in producing an accurate estimate of snow and ice
cover in alpine terrain is that snow and ice cover varies at spatial resolutions finer than
the instantaneous field of view (IFOV) of most satellite platforms. At the pixel level
(IFOV), snow and ice might mix with other landcover types, such as forests, exposed
rock, and bare soil, resulting in a mixed and complex spectral response; atmospheric
properties and anisotropic reflectance also might vary at sub-pixel level. The resulting
spectral response is a complex function of the reflective properties of landcover types that
can be solved through spectral un-mixing techniques. Spectral mixture analysis assumes
that the energy measured by satellite sensors is a linear combination of the radiances
reflected by the landcovers in the IFOV (Painter, Dozier, Roberts, Robert, and Green.
2003). Spectral un-mixing analysis uses least square modeling to describe the linear
combination of pure components known as endmembers (Mather 1999).
33
Rosenthal and Dozier (1996) found that satellite derived spectral un-mixing
classification estimated snow and ice fractions with accuracy comparable to that of aerial
photography, but in less time, and suggested that spectral un-mixing is suitable to study
rugged terrain environments where there is extreme variation of landcover types over a
short horizontal distance; spectral un-mixing also could normalize terrain effects and
shadowing when spectral signatures vary in amplitude, but not in shape.
Nolin, Dozier, and Mertes (1993) pointed out that spectral un-mixing models
assume a linear combination of the spectral endmembers and argued that this assumption,
while making for easier computation and interpretation, induces error of estimate due to
the non-linear nature of the spectral mixing phenomenon. The error of estimate is mostly
related to atmospheric effects and anisotropic reflectance, and must be taken into account
when evaluating landcover percentages from spectral un-mixing analysis.
Klein and Isacks (1998; 1999) proposed a method to detect AAR and ELA
change based on spectral mixture classification. This technique was found to be
extremely effective for glacier change detection, due to the transitional nature of snow
and ice properties; in fact, snow and ice present an almost unlimited gradient of water
content, grain size, density, particulate concentration and texture structure. Spectral
mixture analysis has been successful in the estimation of fractions of snow cover within
pixels. In estimating the transient snowline position (ELA), spectral mixture analysis was
superior to the use of single bands and band ratios, and in most cases, endmembers from
benchmark glaciers can be exported for regional inventory change detection as proxy data
(Klein and Isacks 1999).
34
Chapter 6
Spectral environment
Spectral Characteristics of Snow and Ice
The study of the spectral characteristics of snow and ice is necessary in order to
perform snow cover classification based on spectral signatures. Changes in snow
reflectance and optical properties observed through remotely sensed imagery have been
recognized as indicators of environmental changes associated with local and global
energy fluxes such as mass and energy balance, diagenetic and metamorphic processes,
and facies changes (Hall, Chang, and Siddalingaiah 1988; Hall, Chang, Foster, Benson,
and Kovalick 1989). Reflectance of snow cover is also an important parameter used in
the evaluation of albedo. As discussed earlier, glaciers are characterized by an area of
accumulation and one of ablation. The accumulation area can be further subdivided into
three distinct facies: a dry snow area, usually at higher elevation, where little melting
occurs; a percolation area where local or sporadic melt occurs; and a wet facies
characterized by intense melting during the warm season (Hall et al. 1989). The
identification of the three accumulation facies and the ice cover of the ablation area are
the base parameters for any estimate of mass balance, starting with the simplest
measurement of accumulation area ratio (AAR). To properly identify the snow facies
based on their spectral signatures, it is important to understand the spectral and optical
properties of snow and ice.
35
While of simple chemical composition (H20), snow presents a complex structure
due to the physical properties of water. Water presents a number of triple phase points at
a temperature approaching 273°K (0°C) as a function of the atmospheric pressure (figure
6.1). Triple points are characterized by the presence of coexisting multiple phases (liquid,
vapor and solid) of a substance. Snow is a form of solid precipitation composed of ice
crystals or grains developing from direct conversion of water vapor into ice (McKnight
2000).
The open crystalline structure of snow allows the inclusion of high air content
during the deposition process, leaving enough room for other water phases to exist. As a
consequence, depending on temperature and pressure, snow is characterized by the
presence of ice grains, ice crystals, air and water vapor, and a variable fraction of water.
During deposition, snow develops around nuclei of atmospheric particulates and
chemical impurities such as dust, pollen, soot and ions. The number and types of phases
Figure 6.1. Physical states of H2O (after Martin 2004).
36
present in the snow, the specific structure and density of the snow itself, and the influence
of impurities are important factors that affect the optical properties of snow (Dozier
1989). In fact, the optical properties of snow result from the integration of the optical
properties of all the components present in the snowpack. The complex optical properties
of snow are due to the fact that the resulting combination is not always a linear function
of the fractional components present in the snowpack, but a more complex interaction of
absorption, refractive, and reflective indices of each contributing component (Nolin and
Dozier 2000).
Study of the spectral properties of snow and ice through remote sensing platforms
must be addressed in term of reflectance (this subject will be thoroughly discussed in the
next chapter). Optical satellite imaging registers the amount of energy reflected by an
object on the Earth’s surface. This measurement, also defined as at-sensor-radiance, can
be used to calculate an object’s reflectance. Reflectance is a physical property of a
reflector and is defined by the bidirectional reflectance function (BRDF) (Schott 1997)
][),(),( 100 −= sr
EL
ii φθφθ
ρ (equation 6.1)
where ),( 00 φθL is the radiance from the object toward the sensor (wm-2sr-1µm-1),
and ),( iiE φθ is the incoming solar irradiance (wm-2µm-1) that illuminates the object; both
terms are dependent on the relative azimuth (φ ) and elevation (θ ) of illumination and the
observer’s location. The advantages of using reflectance, rather than recorded pixel
brightness values, are to remove illumination angle effects due to solar elevation and to
37
compensate for the exo-atmospheric irradiance variability registered by optical sensors
along the solar spectrum (Teillet, Barker, Markham, Irish, Fedosejevs, and Storey 2001).
Dozier (1989) observed that the optical properties of ice and water are similar in
the visible and near-infrared wavelengths (VNIR), and that the reflectance of the
snowpack in the VNIR domain is a function of the variations of the refractive index of
ice, the snow grain size, and the percent of impurities present with refractive indices
significantly different from ice. In particular, in the near-infrared wavelengths (NIR: 0.9
µm to 1.3 µm), where ice has significant absorption, snow reflectance depends on the
microstructure of the snow matrix, which is strictly a function of diagenetic processes of
compaction and metamorphic processes that take place during melting-refreezing cycles.
In the visible domain, ice is highly transparent; thus reflectance is inversely proportional
to the amount of impurities (Dozier 1989; Davis, Nolin, Jordan, and Dozier 1993; Nolin
and Dozier 2000). A low, but significant, reduction of reflectance in the visible
wavelengths can be observed when 10 part-per-million by weight (ppmw) of desert dust
or 0.1 ppmw of carbon soot are present as snow impurities (Warren, 1982; Hall et al.
1988).
The variability of the ice absorption index is responsible for the extreme
variability of snow reflectance over the solar irradiance spectrum. In fact, snow
reflectance may vary from nearly complete reflectance to nearly complete absorption
(Nolin and Dozier 2000). In general, snow reflectance decreases as grain size increases
(figure 6.2). Snow reflectance is high in the visible domain, but decreases after
metamorphic processes have occurred (Hall et al. 1988). The NIR domain is particularly
sensitive to grain size, offering a good basis from which to approach remote-sensing
38
based classification of firn and fresh snow over the accumulation zone of glaciers (Hall et
al. 1988). Figure 6.3 shows modeled spectral reflectance for the four glacial facies
defined by Hall et al. (1988; 1989); reflectance decreases as diagenesis and metamorphic
processes transform fresh snow to firn and eventually to glacial ice.
Painter et al. (2003) observed that fresh snowpack presents smaller grain size,
lower density, and lower water content than weathered snow. Firn is the result of
metamorphic and diagenetic processes and is characterized by larger ice grains, higher
density and higher water content. Grain size has a strong negative correlation with
elevation, and thus temperature mean and range. At higher elevation finer grains are
preserved due to the lower rate of melting-refreezing, while at lower elevation large
temperature oscillations allow snow matrix reorganization. Since ice and liquid water
present similar refraction indexes, the presence of liquid water itself does not directly
affect snow reflectance. However, liquid water induces reorganization of the snow matrix
Figure 6.2. Spectral reflectance of snow. Each curve represents the spectrum for a different grain size. The dashed line indicates the center (1.03 µm) of a characteristic absorption band extending from 0.96 µm to 1.08 µm (modified after Nolin and Dozier 2000).
39
structure, resulting in a general increase of snow grain size and lower reflectance (Dozier
1989). Another important observation related to grain size and density is that up to 650
kg/m3, density does not directly affect reflectance, because the center of reflection, the
distance between the grains, is usually greater than the wavelengths of the solar spectrum
(Dozier 1989). As snowpack density increases, grain size changes affect snow reflectance
variations until the packing reaches a threshold at which ice grains fuse together,
obliterating the snow matrix.
As expressed in equation 6.1, the reflectance ratio is a function of the incoming
solar irradiance and of the reflected radiance. Solar irradiance is controlled by the solar
elevation, which determines the amount of energy provided for illumination, while the
radiance is a function of the landcover, observer location, and surface orientation, and is
Ref
lect
ance
Wavelength (µm)
Fresh snow
Firn
Glacier ice
Dirty Glacier ice
Figure 6.3. Modeled spectral reflectance of glacier facies (modified after Hall et al. 1988, Zheng et al. 1984).
40
thus strongly affected by topography. Fresh snow is considered a Lambertian reflector
due to its fine granular structure that diffuses energy equally in all directions (Knap and
Reijmer 1998). As snow ages, diagenetic and metamorphic processes induce grain
growth and uneven layering of the snowpack, and the snow becomes a strong anisotropic
reflector with pronounced forward-scatter and forward-reflection (Knap and Reijmer
1998). Correction based on surface analysis, such as a topographic and anisotropic
normalization model, can be implemented to reduce the effect of forward scattering and
illumination angle (Hall et al. 1988; Dozier 1989). Knap and Reijmer (1998) presented an
alternative solution after observing that, for nadir sensors such as Landsat, the reflectance
error measured over snow and ice fields is consistent and directional and can be solved
by an empirical linear equation that describes the effective reflectance in the form:
),,(),,( φθθφθθρρ ssa c+= (equation 6.2)
where ),,( φθθρ sa represents the apparent at-sensor reflectance registered by the
sensor, and ),,( φθθ sc represents a correction constant obtained from regression (θs is the
solar zenith, and θ and Φ are the bidirectional components of the viewing angle). The
implementation of equation 6.2 is of great impact and benefit for remote sensing
applications for areas where terrain models are unavailable or inaccurate to correct
topographic and anisotropic reflectance of snow and ice cover.
Knap and Reijmer’s (1998) observations are in agreement with Dozier (1989) and
Nolin and Dozier (2000), who suggested that snow reflectance should be modeled as a
multiple scattering problem; the interrelationship between snow reflectance and grain size
is explainable by the Mie scattering theory. The Mie scattering theory and the non-
41
selective scattering model can be applied to calculate scattering and absorption of ice
crystals, assuming the shape of ice grains to be sphere-equivalent (Nolin and Dozier
2000). Mie scattering theory is extremely complex to implement; the main idea is that
Mie scattering occurs when the size of the particle is comparable to the incident
wavelength. The angle and intensity of the scattering is partly a function of wavelength λ,
and presents a characteristic forward-scatter pattern (Schott 1997) (figure 6.4) as
observed by Knap and Reijmer (1998). Non-selective scattering is wavelength
independent and it manifests itself when the size of the particle is considerably greater
than the incident wavelength, resulting in a more uniform scattering of energy in all
directions (Schott 1997).
To model snow reflectance for classification purposes, a number of determining
parameters must be taken in account. Those parameters include ice grain size and
structure (snow matrix), optical properties of the ice grains, chemical-physical
composition of the particulate fractions, illumination and observation geometry, and
energy fluxes.
Direction of incident light
Figure 6.4. Mie Scattering (modified after Nave 2000).
42
Landsat platforms
The first Landsat satellite, originally known as ERTS-1 (Earth Resources
Technology Satellite 1) was launched July 23, 1972. Landsat 1 carried the first MSS
spectrometer (Multi-Spectral Scanner) capable of acquiring three visible channels and
one near-infrared channel. The following missions, Landsat 2 (ERTS B) and Landsat 3,
were launched in January 22, 1975, and March 5, 1978. Landsat 3 was designed to
acquire data in a new thermal band (10.4-12.6 µm wavelength). The MSS spectrometer
provided digital data with a spatial resolution of 79m and a radiometric resolution of 6
bits (later expanded to 7 bits).
During the first half of the 1980s, new micro-technology breakthroughs allowed
the development of a generation of more sophisticated spectrometers. Landsat 4 (July 16,
1982) and Landsat 5 (March 1, 1984) were equipped with the old MSS, as well as the
new Thematic Mapper (TM) spectrometers. Thematic Mapper offered three visible
bands, one near-infrared band, two mid-infrared channels, and a thermal long-wave
infrared band. Some of the most important improvements were the finer spatial resolution
that allowed the TM to sample at 30m in the visible and near/mid-infrared bands, and to
return data with a radiometric resolution of 8 bits.
Landsat 6 was launched on October 5, 1993, but a launch failure resulted in
complete loss of the platform. Landsat 6 carried the Enhanced Thematic Mapper (ETM)
which was designed to provide a panchromatic band at a spatial resolution of 15m. It was
only in April 15, 1999, that the ETM spectrometer became operational with in its new
version (ETM+) on Landsat 7. In addition to the sensors aboard Landsat 5, the ETM+
provides two thermal bands with improved spatial resolution to 60m (with low and high
43
gain mode) and a high-resolution panchromatic band. Table 6.1 provides an overview of
the spatial and spectral specifications of TM and ETM+ spectrometers. In May 31, 2003,
Landsat 7 ETM+ was affected by an instrumentation failure involving the Scan Line
Corrector (SLC); the problem produces an anomaly that invalidates the sampled imagery.
The SLC has been recently remotely deactivated and the validity and quality of the
product are currently under investigation.
TM and ETM+ spectrometers and sensors calibration
Landsat TM and ETM+ do not present significant spectral resolution and spectral
bandwidth differences in the visible, near infrared (NIR), and mid-infrared (MIR)
domains (table 6.1). Although outstanding efforts have been made to maintain calibration
accuracy and over-time consistency between Landsat sensors, there are significant
differences in the radiometric response and detector calibrations that must be taken into
account when cross-correlating reflectance measurements from TM and ETM+ sensors
(Teillet et al. 2001; Chander and Markham 2003; Irish 1998; NASA 2003).
Table 6.1. Spectral and spatial specifications for TM and ETM+ spectrometers (NASA 2003).
Band TM ETM+ TM ETM+1 (Blue) 30 30 0.45-0.52 0.45-0.522 (Green) 30 30 0.52-0.60 0.53-0.613 (Red) 30 30 0.63-0.69 0.63-0.694 (Near IR) 30 30 0.76-0.90 0.78-0.905 (Middle IR) 30 30 1.55-1.75 1.55-1.756 (Thermal IR) 120 60 10.4-12.5 10.4-12.5 LG6 (Thermal IR) 60 10.4-12.5 HG7 (Middle IR) 30 30 2.08-2.35 2.09-2.358 (Panchromatic) 15 0.52-0.90
Spectral resolution (µm)Spatial resolution (m)
44
Figure 6.5 shows the spectral and radiometric response of the TM and ETM+
spectrometers. It can be observed that ETM+ presents a significant improvement on the
band-edge responses compared to TM. The spectral shifting that occurs on the TM sensor
is attributed to filter outgassing, which has been mostly corrected in the ETM+ sensors
(USGS 2003). TM and ETM+ cross-calibration studies that seek to obtain validation of
physical measurements (such as reflectance) have shown that the effects and the
magnitude of the radiometric and spectral differences on measured apparent-reflectance
are dependent on in-scene surface reflectance characteristic (i.e. land cover), and on
variations of exo-atmospheric solar irradiance and atmospheric conditions (Teillet et al.
2001).
Adjustments for spectral and radiometric differences might thus require
information about in-scene spectral content such as surface radiometric measurements.
Teillet et al. (2001) suggested a methodology to obtain calibration coefficients to correct
spectral differences of matching Landsat TM and ETM+ scenes, but emphasized that the
results are, in fact, scene-dependent. The error measured was within 3% for the VNIR
bands, but was less predictable for the MIR bands; these observations must be taken into
account in any attempt to develop standardized cross-platform spectral libraries to be
implemented in change detection and classification studies. Spectral and radiometric
differences, such as those discussed above, are dependent on technological and physical
variables and strictly associated with the sensors’ characteristics. It also is important to
consider those differences that are operator-controlled and are part of acquisition plan
strategies.
45
The Long-Term Acquisition Plan (LTAP) is a global strategy to optimize the use
of Landsat 7 ETM+ to collect and create an archive to document Earth’s land cover
processes (USGS 2003). The LTAP is based on three major sampling strategies that can
be overridden in case of major natural disasters, national security, or user requests. The
first strategy is based on seasonality and is designed to study vegetation distribution and
accomplish change detection. As a consequence, areas of the world that present
Figure 6.5. Relative spectral responses and spectral differences between TM (grey) and ETM+ (black) spectrometers (modified after NASA 2003).
46
interesting vegetation patterns are sampled at the maximum temporal resolution (16
days), while areas subject to low frequency change are sampled rarely. The second
strategy is based on cloud cover. Predicted cloud cover based on NOAA models
influence the acquisition priority of a specific scene. While not directly affecting the
radiometric settings of the platform, the seasonality and cloud cover scheme assumes that
LTAP targets specific land covers. In particular, the spectrometers must be optimized to
fit the reflective properties of vegetation.
The third strategy is the most important in term of defining the spectral and
radiometric settings of the sensors in relation to land cover and in-scene reflective
properties. LTAP defined calibration guidelines that adjust per-band sensitivity of the
ETM+ spectrometers (gain settings) to the expected reflective properties of the scenes;
adjustments are based on geography, seasonality, and land cover.
The gain setting strategy is based on fixed categories: Land (non-desert, non-ice),
Desert, Ice/Snow, Sea Ice, Water, and Volcano/Night. Table 6.2 summarizes the gain
settings based on land cover and geography (solar illumination). Bands 1 to 3 (visible
domain) change gain settings together, and a similar strategy if followed for band 5 and 7
(MIR domain). Each Landsat 7 scene belongs to one of the defined categories. Solar
elevation varies depending on geography (latitude), season, and acquisition time. The
Landsat 7 Science Data Users Handbook (Irish 1998) provides worldwide temporal maps
of the gain rules; for example, figure 6.6 represents the month of August.
47
Categories Exceptions Visible NIR MIR
Land (non-desert, non-ice) High < 45, High HighHigh > 45, Low High
Desert < 28, High < 45, High < 38, High> 28, Low > 45, Low > 38, Low
Ice/Snow/Sea Ice < 19, High < 31, High High> 19, Low > 31, Low High
Water/Coral Reef High High High
Volcano/Night High High Low
(including solar elevation in degrees)
Table 6.2. LTAP gain rules for Landsat ETM+ (modified after NASA 2003).
Figure 6.6. LTAP gain rules for the month of august. The rules depend on expected landcover, seasonality, and illumination. (after NASA 2003).
48
The purpose of the gain rules is to maximize the instrument's 8 bit radiometric
resolution without saturating the detectors (figure 6.7); the low-gain dynamic range is
about 1.5 times that of the high-gain, thus the low-gain setting is applied to regions with
expected high brightness, and the high-gain mode is applied when the in-scene expected
brightness is low (Irish 1998). Table 6.3 and figure 6.7 show that the maximum measured
radiance depends on the gain setting of the spectrometers.
Reflectance measurements depend on solar irradiance; therefore, the maximum
reflectance registered by ETM+ must be calculated on a scene-by-scene basis. Figure 6.8
represents the modeled maximum reflectance measured by ETM+ on the present area of
study (Path 01-Row 71 – 06-26-00). The Long-term Acquisition Program (LTAP)
classified Path 01- Row 71 as Land based on predicted landcover. The solar elevation at
acquisition time was 39.85 degrees. Following the gain rules from table 6.2, it can be
Figure 6.7. Optimization of radiometric resolution based on ETM+ Radiance Dynamic Range adjustments (after NASA 2003).
Low Gain High GainTM1 285.7 190.0TM2 291.3 193.7TM3 225.0 149.6TM4 225.0 149.6TM5 47.3 31.5TM7 16.7 11.1
Table 6.3. ETM+ Radiance Dynamic Range (wm-2sr-1µm-1) (modified after NASA 2003).
49
seen that the gain is set to “high” for all the bands. It can be argued that there is a
significant difference between the represented curves in figure 6.8. As a result, the gain
setting rules have a great impact in classification of features based on spectral signature,
especially when the preset gain mode overrides the spectral reflectance of target
landcovers, such as snow and ice (as will be discussed in the next chapter). Landsat TM
was not subject to any major gain adjustment over mid and low latitudes. The gain
settings for TM are comparable to the low gain settings of ETM+ (table 6.4 and figure
6.9). Figure 6.10 compares the modeled maximum reflectance for Path 01 - Row 71 from
a TM (06-10-97) and ETM+ (06-20-00) scene. Illumination conditions are similar, but
the high gain setting of ETM+ drastically affects the measurable spectral reflectance
range. This difference has a major impact on the definition of exportable snow and ice
spectral libraries between TM and ETM+.
50
0%
20%
40%
60%
80%
100%
120%
Band
Ref
lect
ance
Low Gain 74% 80% 73% 109% 106% 103%
High Gain 49% 53% 49% 73% 71% 69%
TM1 TM2 TM3 TM4 TM5 TM7
Figure 6.8. Modeled maximum reflectance detected by Landsat 7 ETM+ P01R71 06-26-00 for low- gain mode and high-gain mode.
51
Figure 6.9. ETM+ low-gain and TM Radiance Dynamic Range (w•m-2•sr-1•µm-1)(modified after NASA 2003).
0
50
100
150
200
250
300
350
Band
Rad
ianc
e D
ynam
ic R
ange
ETM+ Low Gain 285.7 291.3 225.0 225.0 47.3 16.7
TM 152.1 296.8 204.3 206.2 27.2 14.4
TM1 TM2 TM3 TM4 TM5 TM7
ETM+ TMLow Gain
TM1 285.7 152.1TM2 291.3 296.8TM3 225.0 204.3TM4 225.0 206.2TM5 47.3 27.2TM7 16.7 14.4
Table 6.4. ETM+ low-gain and TM Radiance Dynamic Range (w•m-2•sr-1•µm-1) (modified after NASA 2003).
52
Use of Landsat TM and ETM+ for Snow and Ice Spectrometry
Landsat imagery has been used for snow and ice studies since its first mission in
1972. Landsat TM and ETM+ have been used to study and monitor glaciers and to
measure snow and ice cover extent for global change, hydrology and climatic research.
Nevertheless, Landsat TM and ETM+ were not developed specifically to monitor snow
and ice spectral characteristics, but primarily to monitor vegetation (USGS 2003).
Landsat TM and ETM+ can be employed to monitor glaciers, but a number of important
issues arise when comparing the spectral characteristics of snow and ice to the spectral
0%
20%
40%
60%
80%
100%
Band
Ref
lect
ance
L5 39.7% 83.1% 67.2% 101.8% 64.7% 91.2%
L7 49.3% 54.1% 49.9% 76.4% 69.7% 66.7%
TM1 TM2 TM3 TM4 TM5 TM7
Figure 6.10. Maximum reflectance detected by Landsat 5 TM P01R71 06-10-97 (L5) and Landsat 7 ETM+ P1R71 06-26-00 (L7). The ETM+ is preset by LTAP on high-gain mode (HHHHHH) (modified after NASA 2003).
53
and radiometric capability of the sensors. Landsat ETM+, in particular, presents
important limitations in the discrimination of snow facies when gain settings override the
reflectance characteristic of snow and ice.
Figure 6.11 represents the spectral characteristics of modeled directional-
hemispherical reflectance of four classes of snow based on grain size (Nolin and Dozier
2000); TM and ETM+ bandpasses are represented by gray boxes. The best discriminator
for snow grain size and facies is the wavelength interval from 1.1 to 1.4 µm (MIR), where
the reflectance range is at a maximum (Davis et al. 1993; Nolin and Dozier 2000; Painter
et al. 2003); TM and ETM+ do not provide spectral bands in this interval. Band 4 (NIR)
is the closest option but the reflectance range is considerably reduced and more sensitive
to instrument noise and processing errors. Band 5, while not centered on the highest peak
Figure 6.11. Modeled Directional-Hemispherical reflectance of deep snow (modified after Nolin and Dozier 2000). TM bandpasses are represented by light gray boxes overlapping the spectra. “r” represents grain size.
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7
r = 50 µmr = 200 µm
r = 500 µm
r = 1000 µm
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4Wavelength (µm)
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7TM 1 2 3 TM 4 TM 5 TM 7
r = 50 µmr = 200 µm
r = 500 µm
r = 1000 µm
r = 50 µmr = 200 µm
r = 500 µm
r = 1000 µm
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.40.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4Wavelength (µm)
54
of the far MIR range, occupies a favorable interval and can be used for classification and
discrimination purposes, such as to discriminate clouds from snow due to the high
reflectance of clouds in the MIR. However, Nolin and Dozier (2000) observed that snow
reflectance in this range decreases quickly as grain size increases, approaching low
reflectance values proximate to instrument noise (low signal to noise ratio). Band 7 is
spectrally similar to Band 5. The visible domain, where snow has very high reflectance,
is not sensitive to grain size, and not particularly suitable for snow facies classification.
Figure 6.12 illustrates that TM bands 1 to 3 could potentially be used to distinguish
between snow and glacial ice (Hall et al. 1988) due to the low reflectance of ice and dirt
ice in the ablation zone.
Figure 6.12. Modeled Directional Hemispherical reflectance of glacier facies in TM bandpasses for the visible and NIR domains (modified after Hall et al. 1988, Zheng et al. 1984).
TM bandpass
Dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
0
20
40
60
80
100
TM 1 2 3 TM 4
Fresh snow
Firn
Glacier ice
Dirty Glacier ice
0.4 0.6 0.8 1.0 1.2Wavelength (µm)
TM bandpass
Dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
0
20
40
60
80
100
TM 1 2 3 TM 4
Fresh snow
Firn
Glacier ice
Dirty Glacier ice
0.4 0.6 0.8 1.0 1.2Wavelength (µm)
0.4 0.6 0.8 1.0 1.20.4 0.6 0.8 1.0 1.2Wavelength (µm)
55
Reflectance registered within each TM bandpass can be mathematically integrated
to obtain modeled directional-hemispherical spectral reflectance signatures for snow
(Rosenthal and Dozier 1996) and glacier facies (Hall et al. 1988) as shown in figure 6.13.
Modeled visible and NIR could be effectively used to discriminate glacier facies, but this
model does not consider the cut-off effect induced by the gain settings of TM and ETM+.
The spectral signatures need to be modified in relation to the spectral and radiometric
limitation of Landsat spectrometers.
TM and ETM+ spectral cut-off models for reflectance can be developed from
radiance dynamic ranges (table 6.3 and table 6.4) to produce scene dependent apparent
reflectance ranges (figure 6.8 and figure 6.10). The apparent reflectance obtained is the
Figure 6.13. Integrated spectral directional-hemispherical reflectance of deep snow and glacial facies in TM bandpasses (modified after Zheng et al. 1984, Dozier and Mark 1987, Hall et al. 1988, Rosenthal et al. 1996). TM bandpasses are represented by light gray boxes overlapping the spectra. “r” represents snow grain size.
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7
r = 50 µmr = 100 µmr = 200 µmr = 500 µmr = 1000 µm
FirnGlacial IceDirty Ice
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4Wavelength (µm)
TM in
tegr
ated
dire
ctio
nal-h
emis
pher
ical
Ref
lect
ance
(%)
TM bandpass
0
20
40
60
80
100
TM 1 2 3 TM 4 TM 5 TM 7
r = 50 µmr = 100 µmr = 200 µmr = 500 µmr = 1000 µm
FirnGlacial IceDirty Ice
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4Wavelength (µm)
0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.40.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4Wavelength (µm)
56
maximum reflectance detectable by the TM and ETM+ sensors for specific in-scene
illumination angle and solar irradiance, and can be defined as TM and ETM+ reflectance
cut-off curves. In figure 6.14 reflectance cut-off curves (left) based on TM P01-R71 06-
10-97 and ETM+ P01-R71 06-26-00 are intersected with the modeled curves from figure
6.13 to obtain platform dependent spectral signatures of glacier facies (right).
It can be argued that the spectral signature patterns of TM and ETM+ low-gain
are essentially identical except for band 1, leaving open the possibility for the
development of standardized spectral signatures for inter-platform classification
purposes. On the other hand, ETM+ high-gain presents a limited reflectance range that
substantially affects the modeled spectral signature of glacier facies. The resulting
spectral curves allow limited interpretation for detection of snow, firn, and ice in the
visible and NIR domain, leaving only the noise sensitive MIR range for facies
discrimination.
The effect of reflectance compression observed for TM and ETM+ low-gain band
1 and band 3, and for visible and NIR bands for ETM+ high-gain, is described in the
literature as band saturation (Hall et al. 1987, 1988, and 1989; Dozier 1989; Fily,
Bourdelles, Dedieu, and Sergent 1997; Knap and Reijmer 1998; Klein and Isacks 1999).
Band saturation is the physical result of surface reflectance going beyond the detector
capability, and, in case of TM and ETM+, confirms that Landsat platforms were not
developed or programmed to measure snow and ice landcovers over low and mid
latitudes. USGS (2003) reported that the main purpose of Landsat TM and ETM+
missions was to monitor vegetation. The spectrometers were developed to maximize
dynamic response over vegetated areas.
57
Figure 6.14. TM and ETM+ reflectance cut-offs (left) and resulting snow and ice signatures from modeled directional hemispherical reflectance (right). Refer to Figure 6.13 for snow and ice classes legend.
TM
ETM+ Low Gain
ETM+ High Gain
1 2 3 4 5 7TM bandpass
0.5 1.0 1.5 2.0 2.4Wavelength (µm)Wavelength (µm)
0 0.5 1 1.5 2 2.5
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
TM
ETM+ Low Gain
ETM+ High Gain
1 2 3 4 5 7TM bandpass
1 2 3 4 5 7TM bandpass
0.5 1.0 1.5 2.0 2.40.5 1.0 1.5 2.0 2.4Wavelength (µm)Wavelength (µm)
0 0.5 1 1.5 2 2.50 0.5 1 1.5 2 2.5
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
0
20
40
60
80
100
120
Ref
lect
ance
(%)
58
The most important outcomes of the analysis presented in this chapter are:
The effect of band saturation conceals important spectral information that would
allow a correct distinction between snow and ice facies;
Spectral reflectance signatures based on physical parameters obtained from
Landsat TM may not perform correctly when applied to Landsat ETM+ imagery (and
vice versa) depending on ETM+ gain mode;
For snow and ice classification purposes, Landsat ETM+ in high-gain mode
presents significant limitations for the interpretation of glacier facies due to the high
saturation of the visible and NIR bands and consequent compression of the spectral
signatures in these domains.
59
Chapter 7
Preprocessing
Data Characteristics
Figure 7.1 shows the relative position of the study area within a Path 01 – Row 71
Landsat scene. Five Landsat scenes were selected on the basis of temporal, seasonal,
cloud cover and cost criteria (table 7.1). The temporal criteria were based on TM satellite
image availability, and to allow the comparison of the results with previous field
observations in the area of study for the same time span. The seasonal constraint is linked
to the position of the equilibrium line, which, for this area, corresponds to the limit of the
firn facies at the end of the dry season (winter) (Klein and Isacks 1999).
+ 67.99 W - 15.89
Path 01 - Row 71 UTM WGS 84 Zone 19 S
Cordillera Real
Figure 7.1 Relative position of the Cordillera Real within a Path01-Row71 Landsat scene. Approximate coordinates of the scene center are latitude 15.89 S, longitude 67.99 W.
60
Table 7.2 summarizes the characteristics (metadata) for the five scenes. Three
Landsat TM images were acquired from the Earth Observatory System (EOS) Data
Gateway for $15 each. Of these, the Landsat 4 TM image belonged to the datasets
developed by the Landsat Pathfinder Humid Tropical Forest Project (HTFP), a project
oriented to establish long-term, medium- to high-resolution databases for global change
research (EDC 2003). The 1987 and 1997 Landsat 5 TM scenes were part of the NASA
Landsat Data Collection (NLDC), which contains MSS and TM data purchased by
NASA investigators. NLDC and HTFP make available only those scenes that were
previously used for the specific projects; thus there was a limited selection of scenes that
fit the requirements for this study. For year 1997, the only available image was acquired
in early June, slightly outside the seasonality requirements.
The two Landsat 7 ETM+ scenes were purchased from the USGS at $600 each.
The USGS Landsat ETM+ datasets where selected using the source criteria as the TM
scenes from NLDC and HTFP. Due to the Long-Term Acquisition Plan (LTAP)
priorities, Path 01 – Row 71 scenes are sampled at a lower frequency than the typical 16
day (Landsat 7 maximum temporal resolution).
Least possible due to limited research fundsCost
Less than 1%Cloud cover (on study area)
austral winter; late winter if possibleSeasonality
from mid 80s to todayTemporal Distribution
Landsat dataset query constraints
Least possible due to limited research fundsCost
Less than 1%Cloud cover (on study area)
austral winter; late winter if possibleSeasonality
from mid 80s to todayTemporal Distribution
Landsat dataset query constraints
Table 7.1. Constraints applied to the selection of suitable Landsat scenes
61
Table 7.2 Metadata summary for the selected Landsat scenes.
ID 5001071008721410Platform Landsat 5
Sensor TMAcquired 8/2/1987
Cloud Cover 10%(over study area) < 1%
Sun Elevation 39.23Sun Azimuth 49.68
ID 4001071008921110Platform Landsat 4
Sensor TMAcquired 7/30/1989
Cloud Cover 10%(over study area) < 1%
Sun Elevation 40.00Sun Azimuth 49.00
ID 5001071009716110Platform Landsat 5
Sensor TMAcquired 6/10/1997
Cloud Cover 20%(over study area) < 1%
Sun Elevation 37.79Sun Azimuth 42.90
ID 7001071000017850Platform Landsat 7
Sensor ETM+Acquired 6/26/2000
Cloud Cover 31%(over study area) < 1%
Sun Elevation 39.85Sun Azimuth 39.44
ID 7001071000121250Platform Landsat 7
Sensor ETM+Acquired 7/31/2001
Cloud Cover 32%(over study area) < 1%
Sun Elevation 42.59Sun Azimuth 45.58
62
As a consequence, there is a very limited amount of imagery available once
proper cloud cover and the narrow 3 month window necessary to study glacier
phenomena are considered. For the year 2000, for example, an early winter scene was the
only image available.
Landsat images are available to the public in three preprocessed versions. Level
0R (L0R) is a raw format provided along with ancillary acquisition information allowing
the user to perform first order processing. ETM+ imagery was initially meant to be
provided only in L0R format, but marketing and research demand forced NASA to
systematically correct the data to ensure availability and accuracy (Irish 1998). The
systematic correction applied to L0R is the Level 1R (L1R), which consists of
radiometrically corrected imagery. A further processing step is available through Level
1G (L1G), which provides systematically and geometrically corrected data. NLDC and
HTFP TM scenes are available in L1R or L1G format, depending on the original project;
the ETM+ scenes were purchased in L1G format to be used as a reference for geometric
correction and co-registration of all the other images.
Note:
For more information on HTFP and NLDC projects refer to respectively:
http://edcdaac.usgs.gov/pathfinder/pathpage.html
and
http://edcdaac.usgs.gov/pathfinder/nldc/nldc.html.
63
Geometric Correction and Co-registration
Geometric correction assigns scale and projection properties to remotely sensed
data, while co-registration is a process of matching the coordinate systems of two or more
images collected at different times or by different sensors (Mather 1999). Geometric
correction and co-registration of the five images was necessary in order to perform a
meaningful comparison of the distribution of the final snow and ice classes from each
scene (Mather 1999); the analysis of the dynamic variations of the glaciers occurred in
the specific period of time elapsed between the acquisition of the scenes, and is based on
quantified changes of snow and ice facies.
The first geometrically corrected image acquired for this project was the 2001
ETM+ image. This image was used as master image to co-register and geometrically
correct the remaining four scenes using the image-to-image geometric correction tool of
ERDAS Imagine 8.6. The correction was based on a final selection of 25 to 30 suitable
ground control points (GCPs) on the ETM+ image (figure 7.2); the operation was
rendered particularly difficult due to the lack of clearly defined anthropogenic features.
The final Total Root Means Square (TRMS) error obtained for each image was below
0.5, corresponding to less than one-half the cell size (14.25m) (ERDAS 1997). The best
results were obtained by distributing the control points in the vicinity of the study area.
While still providing a TRMS below 0.5, the initial attempt to select control points
distributed over the whole scene resulted in noticeable local error over the study area.
The final spatial adjustment was obtained using a first order polynomial rigorous
transformation and a nearest neighbor (NN) re-sampling algorithm; the resulting four co-
registered images had a spatial resolution of 28.5m to match the 2001 ETM+ cell size
64
(figure 7.2). Nearest Neighbor resampling ensured that the final images preserved
original digital values; other algorithms, such as bilinear interpolation and cubic
convolution, substitute the original data with interpolated or spatially averaged digital
values, resulting in the loss of the original information and introduction of digital artifacts
(Mather 1999).
Two important considerations must be taken into account: (1) the USGS ETM+
L1G scenes (geometrically corrected from the source) were expected to be spatially co-
registered, but in fact, appeared not to correspond geometrically (figure 7.4); the 2000
image had to be co-registered to the 2001 image as were the other images; (2) geometric
co-registration does not correct for relief effects and distortion induced by the angle of
view of the sensor, but assures that all the images represent the same geometric structure.
Figure 7.2. ERDAS Image 8.6 Geometric Correction tool. Example of Ground Control Points (GCP) selected in proximity to the study area.
65
Figure 7.3. Geometric correction and co-registration of Landsat TM Path 01 – Row 71 8/2/1987. Before (left) and after (right). The original image was a L1R product.
Figure 7.4. Landsat ETM+ Path 01 – Row 71 6/26/2000 and 7/31/2001. Spatial offset at La Paz Airport (approx. 200 m). Both scenes were originally projected in UTM 19S WGS84.
ETM+ 2001 ETM+ 2000
2001
2000
66
Study Area Subset
The glaciers of the Cordillera Real have been studied and catalogued since the
1970s. Jordan (1991 and 1999) and colleagues (Jordan at al. 1980) produced an inventory
of Bolivian glaciers available from the USGS publication Satellite Image Atlas of
Glaciers of the World - SOUTH AMERICA (USGS 1999). Jordan’s inventory is available
as a vector file in ArcInfo E00 format (figure 7.5). The polygons describing the glaciers
can be used to define spatial reference units for zonal statistics and to filter out ice and
snow classified covers that do not represent glaciers.
The vector polygons containing the spatial description of the glaciers were
rasterized and converted to a binary mask (value 1 = glacier, value 0 = non-glacier) and a
zonal layer (zones defined by attribute USGS_ID of the glacier inventory); a three pixel
buffer to better evaluate the transition from glacier to non-glacier classes was added to
the binary mask. The five Landsat scenes were then clipped with the binary mask to
obtain subset images of the glaciers of the Cordillera Real (figure 7.6). Removing
unnecessary data from the image also shortens the computational time of the
classification process.
67
Number of glaciers: 893
Smallest Glacier (sqm) 923.1
LargestGlacier (sqm) 8,298,608.0
Number of glaciers: 893
Smallest Glacier (sqm) 923.1
LargestGlacier (sqm) 8,298,608.0
Number of glaciers: 893
Smallest Glacier (sqm) 923.1
LargestGlacier (sqm) 8,298,608.0
Number of glaciers: 893
Smallest Glacier (sqm) 923.1
LargestGlacier (sqm) 8,298,608.0
Figure 7.5. Glaciers of the Cordillera Real and detail of the Ancohuma-Illampu Massif. Area is calculated as planimetric surface. Glacier classification by Jordan (1999).
68
Inventory
Binary mask
Landsat Source
Glaciers
Figure 7.6. Details of the subset procedure based on Jordan’s glacier inventory (1999).
69
Atmospheric Correction
Atmospheric correction of remotely sensed imagery is needed when performing
measurements seeking to determine a physical property of the surface of the Earth, such
as reflectance, and when it is necessary to standardize the at-sensor radiance to compare
imagery acquired with different atmospheric conditions (Mather 1999). To obtain a
measurement of the effective radiance leaving a surface and calculate its true reflective
properties, it is necessary to remove the optical effects induced by the atmosphere. Figure
7.7 shows the solar energy paths through the atmosphere and table 7.3 summarizes the
characteristics of each path. Thermal radiance terms will not be considered in this
research since they do not influence the energy measured in the visible, NIR, and MIR
domains of the Landsat spectrometers. There are two main types of radiance: (1) surface-
reflected solar energy that reaches the sensor
(LA); and (2) path radiance, which includes
radiances of type B, C, G, and I, and represents
the amount of solar energy that reaches the
sensor, but is not a function of the surface
properties (Schott 1997). Atmospheric
correction algorithms and models aim to
remove the effect of path radiance from
remotely sensed imagery. Figure 7.7. Solar energy path (modified after Schott 1997)
70
The relationship between effective radiance (LA), path radiance (LP = LB + LC +
LG +LI), and at-sensor (or apparent) radiance (LS) can be generalized as
LS = LP + LA; (equation 7.1)
This equation describes the common observation that during hazy days the
apparent radiance (LS) increases as scattering (haziness) increases (figure 7.8). The
equation also describes the effect of diffuse illumination in shadows, where, in the
absence of direct illumination, LA = 0 and LS = LP, radiance becomes a direct function of
scattering and background reflection. Using LOWTRAN models, Schott (1997) observed
that, in the visible and NIR domain, the G and I terms are significantly smaller than A, B,
and C. The magnitude of the path radiance versus effective radiance varies, depending on
the reflectance of the surface. Path radiance significantly increases the radiance of
surfaces with low reflectance values (reflectance lower than 5% in figure 7.8) and the
effect of path radiance is gradually reduced by high surface reflectors.
Table 7.3. Solar energy path (modified after Schott 1997).
Path Radiance Type Description
Object RadianceA L A Effective Radiance Measurement of solar energy directly reflected by an objectD L D Self emitted Radiance Radiance from the object, self emission, usually thermal
Downwelled RadianceB L B Skylight Radiance Reflected radiance from atmospheric scatteringE L E Sky thermal Radiance Reflected thermal radiance from atmospheric self emission
Upwelled RadianceC L C Uppwelled Radiance Solar energy scattered by the atmosphere reaching the sensorF L F Upwelled thermal Radiance Self emitted atmospheric thermal energy reaching the sensor
Background RadianceG L G Background Radiance Illumination provided by neighboring surfaces H L H Background thermal Radiance Thermal energy from neighboring surfaces I L I Adjacency Effect Scattered radiance from neighborsing surfaces
71
For dark surfaces, with reflectance from 1% to 3%, the path radiance may
represent 50% of the total at-sensor radiance; as a consequence, a small error in the
computation of upwelled radiance results in a large error of estimate of the effective
radiance (Schott 1997). For highly reflective surfaces such as fresh snow, path radiance is
a relatively minor component of total radiance. The magnitude of path radiance is also a
function of atmospheric conditions; Schott (1997) observed that in clear sky conditions,
the A term is up to seven times greater than the downwelled radiance, but the terms
become comparable in hazy atmospheric conditions. It is thus very important to account
for the reflective properties of the investigated surface and the atmospheric condition at
the time of acquisition when generalizing atmospheric correction models to be applied to
satellite imagery.
Figure 7.8. Effects of atmospheric transmission and path radiance on resulting total at-sensor reflectance (modified after Schott 1997).
Radiance leaving the ground (W*cm-2*sr-1)0.001 0.002 0.003 0.004 0.0050
Rad
ianc
e re
achi
ng th
e gr
ound
(W*c
m-2*s
r-1)
0.001
0.003
0.002
0.005
0.004
0
Ideal atmosphere
Visibility = 20 km
Visibility = 5 km
Visibility = 2.5 km
Visibility = 1 km
Radiance leaving the ground (W*cm-2*sr-1)0.001 0.002 0.003 0.004 0.0050
Rad
ianc
e re
achi
ng th
e gr
ound
(W*c
m-2*s
r-1)
0.001
0.003
0.002
0.005
0.004
0
Radiance leaving the ground (W*cm-2*sr-1)0.001 0.002 0.003 0.004 0.0050
Radiance leaving the ground (W*cm-2*sr-1)0.001 0.002 0.003 0.004 0.0050
Rad
ianc
e re
achi
ng th
e gr
ound
(W*c
m-2*s
r-1)
0.001
0.003
0.002
0.005
0.004
0
Rad
ianc
e re
achi
ng th
e gr
ound
(W*c
m-2*s
r-1)
0.001
0.003
0.002
0.005
0.004
0
Rad
ianc
e re
achi
ng th
e gr
ound
(W*c
m-2*s
r-1)
0.001
0.003
0.002
0.005
0.004
0
0.001
0.003
0.002
0.005
0.004
0
Ideal atmosphere
Visibility = 20 km
Visibility = 5 km
Visibility = 2.5 km
Visibility = 1 km
Ideal atmosphere
Visibility = 20 km
Visibility = 5 km
Visibility = 2.5 km
Visibility = 1 km
Reflectance = 5%
Reflectance = 25%
72
The effective radiance LA for a specific wavelength can be described as a function
of the total incoming solar energy (Htot), the atmospheric transmittance (T) in the specific
wavelength, and the reflective properties of the irradiated surface (Mather 1999):
LA = HtotρAT; (equation 7.2)
and rearranging
THL
tot
AA =ρ (equation 7.3)
where ρA describes the effective reflectance in absence of atmospheric path
radiance. By definition, for full atmospheric transmittance (T = 1), equation 7.3
represents the theoretical optical reflectance of a surface, described by the ratio of
outgoing exitance and incoming irradiance (equation 6.1). The value of this ratio is a
physical property of a surface, and the desired target measurement of remote sensing
studies. To better describe the contribution of the path radiance to at-sensor
measurements, equation 7.2 can be substituted to LA in equation 7.1
LS = LP + HtotρAT (equation 7.4)
and then solved for effective reflectance (ρA)
THLL
tot
PSA
)( −=ρ (equation 7.5)
Equation 7.5 can be further elaborated to obtain
TH
LTH
L
tot
P
tot
SA −=ρ . (equation 7.6)
73
Since in can be assumed that (Schott 1997; Mather 1999)
THL
tot
SS =ρ , (equation 7.7)
it possible to write equation 7.6 as
cSA −= ρρ , where TH
Lc
tot
P= . (equation 7.8)
Equation 7.5 and equation 7.8 are of utmost importance in the solution of
atmospheric correction models. Equation 7.5 is commonly applied to in-scene
atmospheric correction models (such as the Dark Object Subtraction method) and
illustrates that the effective reflectance (ρA) of an object can be obtained by subtracting
the estimated path radiance (LP) from the at-sensor radiance measurements (LS). Equation
7.8 shows that the effective reflectance (ρA) can also be obtained by subtracting a
correction value (c) from at-sensor apparent reflectance measurements (ρS). Three main
methodologies are available to estimate LP and c: ground station corrections, atmospheric
propagation models, and in-scene calibration.
Measurement of reflectance by using onsite spectrometers is the most accurate
procedure to calculate c, and is based on the comparison of ground reflectivity
measurements (ρG) and at-sensor reflectance. The difference between the measurements
is the path radiance:
c = ρS – ρG; (equation 7.9)
74
The obtained constant is a direct function of LP, because the ground
measurements are corrected for all the components of the path radiance. While extremely
accurate, this procedure requires field measurements using costly equipment and the
availability of spectrally pure target surfaces in the satellite image to avoid spectral
mixing. This method is also difficult to implement in remote and less accessible areas.
Atmospheric propagation models (APM) can be used to estimate atmospheric
radiative transfer properties from field measured parameters or from standardized
atmospheric conditions (Mather 1999). Atmospheric propagation models can be
computationally intensive and a number of software packages have been developed to
produce the most accurate atmospheric modeling, such as LOWTRAN, MODTRAN, and
5S/6S. Atmospheric propagation models are platform- and scene-independent and can be
applied to an area with little or no information about the atmospheric conditions present.
Processing parameters from field measurements that describe the atmospheric conditions
at acquisition time provides extremely accurate, but extremely costly results; therefore,
processing of standardized atmospheric conditions is more likely to be applied to the
study of remote areas, such as alpine environments (Schott 1997). Schott also observed
that APMs offer a range of acceptable results based on the quality of the input
parameters, but warned that the concept of “garbage in, garbage out” is of pertinent
application in APMs processing. Figure 7.9 shows an example of an online MODTRAN3
application offered by the Department of Geophysical Sciences at the University of
Chicago (http://geosci.uchicago.edu/~archer/cgimodels/radiation.html).
Atmospheric corrections that take advantage of the characteristics of the data
within multispectral imagery are known as in-scene calibrations (Schott 1997). In-scene
75
calibrations may be solved by comparing imagery from different altitudes, different
angles or different wavelengths to solve equation 7.8. Dark object subtraction (DOS)
(Chavez 1988) is an in-scene calibration methodology commonly used to correct satellite
imagery for path radiance. The theory behind DOS is based on the concept expressed in
equation 7.5 that the effective reflectance (ρA) can be obtained by subtracting the path
Model Output
Intensity (W/(m2-sr-wavenumber)) vs. wavenumber Atmospheric Profiles vs. Height, km
MODTRAN 3
Atmospheric Longwave Radiation Modeler
University of ChicagoDepartment of Geophysical Science
Model Input •pCO2 = 330 •CH4 = 1.7 •Tropospheric Ozone (ppb) = 28 •Stratospheric Ozone Scale Factor = 1 •T offset = 0 •Constant Water Vapor Pressure •H2O vapor twiddle factor = 1 •Tropical Atmosphere Model •No Clouds •Sensor Altitude = 750 Looking down •User Supplied Tag = Christian+Degrassi •Model Run Tag Number, 19133837
Figure 7.9. MODTRAN 3. Example of longwave atmospheric transfer model output from standardized input (http://geosci.uchicago.edu/~archer/cgimodels/radiation.html).
76
radiance (LP) from at-sensor apparent radiance measurements (LS), or adjusting the
apparent reflectance (ρS) by a correction factor c (equation 7.8). In the absence of
external references to solve equation 7.9, the only way to establish a radiance absolute
reference is to identify an in-scene feature that presents known spectral characteristics.
Such opportunities are offered by dark objects, features that, in absence of path radiance,
would present radiance equal or proximate to zero radiance (thus appearing black):
c(LP) = ρSDO – ρADO (equation 7.10)
where ρSDO is the apparent reflectance of a dark object and ρADO is the effective
reflectance of a dark object, which is equal to 0%; thus for ρADO = 0
c = ρSDO (equation 7.11)
Shadowy, clear lakes or recent basalt flows are examples of dark objects. The
correction factor (c) obtained equals the apparent reflectance of a dark object (Chavez
1989). This is in line with the practical observation that shadows and black objects appear
brighter in presence of haze (scattering). For a dark object, the apparent brightness is
produced by the path radiance.
While simple in theory, the Dark Object Subtraction (DOS) method requires an
accurate selection of the dark features to avoid under- and over-estimation of the
correction factor. Chavez (1988) explained that there is a high probability that at least a
few pixels within an image will present 0% reflectance, such as shadows from
topography or clouds. There are a number of techniques that can be implemented to
extract and model the correction factor for DOS calibration. The simplest of these
77
methods is based on the analysis of the band frequency histograms; a more complex
methodology takes into account the central wavelength of each multispectral band to
calculate atmospheric scattering effects (Chavez 1988). Other DOS models include a
more thorough correction for atmospheric absorption, transmittance and downwelling
radiance (Moran, Jackson, Slater and Teillet 1992; Chavez 1996). Three DOS techniques
were coordinated and implemented to remove path radiance from the Landsat TM and
ETM+ scenes of the Cordillera Real in this project.
A first order DOS correction was performed by identifying the lowest non-zero
frequency for each of the band histograms, assuming that it represented the shifted
radiance of the darkest object in the scene (Chavez 1988). However, very low non-zero
histogram values are sensitive to instrument noise and may provide a false dark object,
resulting in underestimation of the path radiance. Selecting the frequency at which the
histogram slope starts increasing more consistently may help to filter out background and
instrument noise, but the selection of the ideal frequency can be highly biased, subject to
analyst interpretation, and may result in inconsistent under- or over-estimation of the path
radiance. The method used in this research was based on the identification of in-scene
dark targets, such as lakes, deep topographic shadows and cloud shadows. A maximum of
200 of the darkest object pixels were sampled and analyzed through descriptive statistics.
Table 7.4 shows the resulting final set of DOS values for the five images. The
atmospheric conditions, and thus the path radiances, are very similar in the five images.
The main difference is observed between TM and ETM+; this is due differences in sensor
calibration and spectral response, as discussed in chapter 7.
78
Chavez (1988) argued that the selection of uncorrelated Dark object Subtraction
values from the spectral bands might not reflect the real relative scattering model; in fact,
the manual selection of independent DOS band values does not have a physical
foundation, resulting in a possible distortion of the natural atmospheric scattering
phenomenon. It is possible to increase DOS performance by introducing an atmospheric
scattering correction model to obtain a correction consistent with scattering optical and
physical phenomena. Relative atmospheric scattering models assume that scattering is
inversely proportional to the wavelength by a power factor dependent on the type of
scattering. Rayleigh scattering is inversely proportional to the fourth power of the
wavelength, while the Mie scattering power factor may vary between 4, representing
clear atmospheric conditions, and 0 for complete cloud cover or maximum scattering
(Chavez 1988). Chavez suggested that for observed real atmospheric conditions, the
power factor ranges within 2 and 0.7, representing a mixture of Rayleigh and Mie
scattering, and between 0.7 and 0.5 for increasing hazy conditions. Dark Object
Subtraction techniques cannot correct for heavy haze conditions (power factor lower than
0.5). The relative scattering model uses the central wavelength of each TM band, a
starting DOS value from one representative band, and a standardized model for
atmospheric conditions to estimate DOS correction for all the bands (Chavez 1988).
1987 1989 1997 2000 2001TM1 40.0 35.0 36.0 36.0 37.0TM2 11.0 11.0 11.0 22.0 23.0TM3 5.0 7.0 6.0 14.0 15.0TM4 3.0 3.0 3.0 9.0 10.0TM5 1.0 1.0 1.0 7.0 7.0TM7 1.0 1.0 1.0 6.0 6.0
TM ETM+
Table 7.4. Manually selected DOS values based on the analysis of dark targets present in the five scenes. The values are represented as digital numbers.
79
Table 7.5 shows the power factor corrections corresponding to five standardized
atmospheric scattering models. Table 7.6 illustrates the resulting coefficients estimated
for the atmospheric scattering model for each Landsat TM and ETM+ band obtained
substituting the central wavelength of the bands for each atmospheric scattering model.
TM and ETM+ band 1 DOS values from table 7.4 were used as the reference starting
values to estimate the relative scattering for each band.
Table 7.7 illustrates, as an example, the calculations and results for the 1987
Landsat TM image used in this research. It can be observed that the scattering function
coefficients for each atmospheric condition are normalized on band 1 (the selected
reference band). The predicted haze values for each band and atmospheric condition are
calculated by multiplying the observed reference DOS value from band 1 (40 in the
Table 7.5. Atmospheric scattering models (after Chavez 1988).
Atmospheric Models
Wavelength Power Factor
Very Clear λ-4
Clear λ-2
Moderate λ-1
Hazy λ-0.7
Very Hazy λ-0.5
Table 7.6. Scattering function coefficients obtained for each band in different atmospheric conditions (methodology after Chavez 1988).
Min Max Central Very Clear Clear Moderate Hazy Very HazyTM1 0.450 0.520 0.485 TM1 18.073 4.251 2.062 1.660 1.436TM2 0.520 0.600 0.560 TM2 10.168 3.189 1.786 1.501 1.336TM3 0.630 0.690 0.660 TM3 5.270 2.296 1.515 1.338 1.231TM4 0.760 0.900 0.830 TM4 2.107 1.452 1.205 1.139 1.098TM5 1.550 1.750 1.650 TM5 0.135 0.367 0.606 0.704 0.778TM7 2.080 2.350 2.215 TM7 0.042 0.204 0.451 0.573 0.672
ETM+1 0.450 0.520 0.485 ETM+1 18.073 4.251 2.062 1.660 1.436ETM+2 0.530 0.610 0.570 ETM+2 9.473 3.078 1.754 1.482 1.325ETM+3 0.630 0.690 0.660 ETM+3 5.270 2.296 1.515 1.338 1.231ETM+4 0.780 0.900 0.840 ETM+4 2.009 1.417 1.190 1.130 1.091ETM+5 1.550 1.750 1.650 ETM+5 0.135 0.367 0.606 0.704 0.778ETM+7 2.090 2.350 2.220 ETM+7 0.041 0.203 0.450 0.572 0.671
Scattering FunctionWavelength
80
example) by the normalized coefficient. The atmospheric conditions for the five Landsat
scenes used in this research study were estimated to be “very clear” under the conditions
proposed by Chavez’ models (Chavez 1988).
A third correction can improve the quantification of DOS values. Chavez (1989)
suggested that natural surfaces rarely present a perfect 0% reflectance (complete
absorption) and are more likely to show a minimum of 1% to 2% reflectance. For each
TM and ETM+ band it is possible to estimate the radiance value (or digital number)
corresponding to 1% reflectance. The number is then added to the previously estimated
DOS values. Table 7.8 summarizes the results for the five scenes after adjusting the
predicted values for sensor calibration, band specific spectral responses, and 1%
minimum reflectance.
Table 7.7. Procedure to predict scattering from TM band 1 reference DOS (methodology after Chavez 1988).
min max avg Very Clear Clear Moderate Hazy Very HazyTM1 0.450 0.520 0.485 TM1 18.073 4.251 2.062 1.660 1.436TM2 0.520 0.600 0.560 TM2 10.168 3.189 1.786 1.501 1.336TM3 0.630 0.690 0.660 TM3 5.270 2.296 1.515 1.338 1.231TM4 0.760 0.900 0.830 TM4 2.107 1.452 1.205 1.139 1.098TM5 1.550 1.750 1.650 TM5 0.135 0.367 0.606 0.704 0.778TM7 2.080 2.350 2.215 TM7 0.042 0.204 0.451 0.573 0.672
Very Clear Clear Moderate Hazy Very HazyTM1 1.000 1.000 1.000 1.000 1.000TM2 0.563 0.750 0.866 0.904 0.931TM3 0.292 0.540 0.735 0.806 0.857TM4 0.117 0.341 0.584 0.687 0.764TM5 0.007 0.086 0.294 0.424 0.542TM7 0.002 0.048 0.219 0.345 0.468
Reference Very Clear Clear Moderate Hazy Very HazyTM1 40 40 TM1 40.000 40.000 40.000 40.000 40.000TM2 11 TM2 22.505 30.003 34.643 36.170 37.225TM3 5 TM3 11.664 21.600 29.394 32.240 34.289TM4 3 TM4 4.664 13.658 23.373 27.462 30.577TM5 1 TM5 0.299 3.456 11.758 16.976 21.686TM7 1 TM7 0.092 1.918 8.758 13.814 18.717
Observed 1987
Haze Function
Normalized Haze Function Model
Predicted Value
Wavelength
81
Chavez (1989) argued that the predicted values for bands TM5 and TM7 are
usually too high due to sensor calibration parameters; a DOS equal to 1 should be used to
correct MIR TM bands. Unfortunately there is not sufficient literature discussing
predicted values for ETM+; therefore, in this project, the observed values were
substituted for those predicted for ETM+ 5 and ETM+ 7.
Image Standardization
Change detection and quantification of physical parameters of the Earth’s
environment depend on the ability of satellite sensors to provide calibrated and consistent
measurements (Markham and Chander 2003). Huang, Yang, Homer, Wylie, Vogelman
and DeFelice (2003) argued that one of the major challenges to obtaining satellite based
change detection is the removal of noise arising from sensor calibrations, instrument
errors, atmospheric effects, and illumination geometry; such noise may render unreliable
and inconsistent the information extracted from satellite imagery. Preprocessing
Table 7.8. Observed DOS and final predicted DOS values.
Observed Predicted Observed Predicted Observed PredictedTM 1 40 40 35 35 36 36TM 2 11 13 11 12 11 12TM 3 5 10 7 9 6 9TM 4 3 5 3 5 3 5TM 5 1 5 1 5 1 5TM 7 1 4 1 3 1 3
2000Observed Predicted Observed Predicted
ETM +1 36 36 37 37ETM +2 22 22 23 23ETM +3 14 18 15 19ETM +4 9 12 10 12ETM +5 7 9 7 9ETM +7 6 9 6 9
1987 1989 1997
2001
82
techniques help correct for calibration and atmospheric effects, but a further step must be
taken to standardize information before change detection analysis can take place. Huang
et al. (2003) showed that converting digital numbers (DN) to at-satellite reflectance not
only provides a physical parameter for describing land surfaces, but also provides more
relevant temporal information about the target (such as illumination). At-satellite
reflectance transformation does not introduce error and its use makes it possible to
standardize images that are separated spatially (mosaic) or temporally (change detection).
Markham and Barker (1986) described the methodology for obtaining at-sensor
reflectance from DN. This methodology was successfully applied to the five Landsat
scenes used in this thesis research. The first step (equation 7.12) consists of converting
DN to at-sensor radiance, a physical measure of the energy registered by the satellite
sensors:
Lλ = Gainλ * DN + Biasλ (equation 7.12)
where λ represents the spectral band and Gain and Bias are post-launch
calibration parameters provided in the Landsat scene header files. Gain and Bias can be
calculated from post-calibration coefficients as shown in equation 7.13 and 7.14.
CALMAX
MINMAXA Q
LLGain )( −= (equation 7.13)
MINA LBias = (equation 7.14)
where QCALMAX is the range of rescaled values (usually 255), LMIN is the spectral
radiance for DN = 0, and LMAX is the spectral radiance for QCALMAX. Table 7.9 summarizes
Gain and Bias pairs, and post-calibration LMIN and LMAX for each TM and ETM+ band.
83
The second step is to transform at-sensor radiance to at-sensor reflectance.
Reflectance is a measurement of the relationship between incoming solar irradiance and
exiting radiance from a surface (indicated by equation 6.1 and equation 7.3). The
bidirectional reflectance factor (BDRF; ρBDRF) simplifies the bidirectional reflectance
distribution function (BRDF; ρBRDF) to a unitless measure of diffuse radiance (Schott
1997) and corresponds to at-sensor reflectance (ρS) measurements:
ρBDRF = π ρBRDF = ρS (equation 7.15)
The ρBRDF term corresponds to equation 6.1 and can be substituted in equation
7.15 to obtain a general representation of unitless at-sensor reflectance:
),(),( 00
iiS E
Lφθφθπρ = (equation 7.16)
From the concept expressed by equation 7.16, Markham and Barker (1986) developed
equation 7.17 to calculate the resulting at-sensor reflectance for Landsat imagery:
θπρ
cos
2
SUN
SS E
dL= ; (equation 7.17)
Table 7.9. Post-calibration parameters (after Irish 1998).
Gain Bias Lmin LmaxTM1 0.6024314 -1.5200000 -1.5200000 152.1000070TM2 1.1750981 -2.8399999 -2.8399999 296.8100156TM3 0.8057647 -1.1700000 -1.1700000 204.2999985TM4 0.8145490 -1.5100000 -1.5100000 206.1999950TM5 0.1080784 -0.3700000 -0.3700000 27.1899920TM6 0.0569804 -0.1500000 -0.1500000 14.3800020
ETM+1 0.7756863 -6.1999969 -6.1999969 191.6000096ETM+2 0.7956862 -6.3999939 -6.3999939 196.4999871ETM+3 0.6192157 -5.0000000 -5.0000000 152.9000035ETM+4 0.6372549 -5.1000061 -5.1000061 157.3999934ETM+5 0.1257255 -0.9999981 -0.9999981 31.0600044ETM+7 0.0437255 -0.3500004 -0.3500004 10.8000021
84
where ESUN represents the incoming solar irradiance for a specific TM and ETM+
band (table 7.10), and d is the date-related Earth-Sun distance. It is necessary to factor in
a correction of up to 3.5% for changes in solar irradiance due to the eccentric orbit of the
Earth (table 7.11). The cosine of the solar zenith (cosθ) is introduced to remove the angle
of illumination effect.
It is important to emphasize that equation 7.17 assumes that the reflector is a flat
Lambertian surface (Markham and Barker 1986). However, only fresh snow cover
presents Lambertian reflectance characteristics. Topographic normalization and
anisotropic reflectance corrections may be applied to remove relief and reflective errors
(Colby 1991) using DEMs and Minnaert coefficients to normalize the measured at-sensor
radiance. Due to the lack of accurate DEMs for the study area, these corrections could not
be performed and the inherent errors must be taken into account during the analysis of the
classification results. (Appendix A provides a case study that discusses the application of
topographic normalization and anisotropic reflectance to a small subset of the study area).
Table 7.10. Mean Solar Exo-atmospheric irradiances (Wcm-2µm-1) (values after Irish 1998).
Landsat 4 Landsat 5 Landsat 7TM1 1957.00 1957.00 ETM+1 1969.00TM2 1825.00 1826.00 ETM+2 1840.00TM3 1557.00 1554.00 ETM+3 1551.00TM4 1033.00 1036.00 ETM+4 1044.00TM5 214.90 215.00 ETM+5 225.70TM6 80.72 80.67 ETM+7 82.07
Mean Solar Exoatmospheric irradiances
Date Julian Day Distance d Sun ElevationTM 1987 08/02/87 214 1.014627391 39.23TM 1989 07/30/89 211 1.015023836 40.00TM 1997 06/10/97 161 1.015323647 37.79ETM+ 2000 06/26/00 178 1.016587389 39.85ETM+ 2001 07/31/01 212 1.014896088 42.59
Table 7.11. Earth-Sun distance (astronomic units AU) and Sun elevation (degrees) at the time of acquisition (distance d calculated after Irish 1998).
85
Chapter 8
Spectral Analysis and Classification
Spectral Characteristics of the Subscenes
After preprocessing, descriptive statistics were calculated for the resulting five
Landsat scene subsets to obtain information on the spectral characteristics of each image.
In particular, the statistics concerning band saturation provided useful information
regarding the separability of snow facies in the accumulation zone (as discussed in
chapter 6). Table 8.1 and figure 8.1 compare the maximum reflectance measured by the
TM and ETM+ sensors for the five subscenes. It can be observed that in the ETM+
scenes (2000 and 2001) the maximum reflectance measured is significantly lower that
that measured by TM spectrometers. This is due to the high-gain mode setting employed
on the ETM+ spectrometer and it confirms the theoretical spectral response of ETM+
described in chapter 6. This observation also raises a question related to the amount of
information that has actually been compressed and obliterated by the band saturation
effect.
Table 8.2 and figure 8.2 show the statistics describing the band saturation effect
for the subscenes. ETM+ visible (VIS) and near-infrared (NIR) domains are significantly
saturated and will drastically affect the possibility to identify different snow facies in the
accumulation zone. The spectral signatures necessary to perform a supervised
86
classification were thus selected from the TM images, which provided a wider reflectance
range, and less band saturation, and thus preserved more spectral information to identify
snow and ice facies.
Figure 8.1. Maximum in-scene reflectance registered by TM and ETM+ sensors in the five subscenes.
0%
20%
40%
60%
80%
100%
120%
TM1 TM2 TM3 TM4 TM5 TM7
Band
Ref
lect
ace
19871989199720002001
Band TM 1987 TM 1989 TM 1997 ETM+ 2000 ETM+ 20011 34.87% 35.11% 36.68% 44.51% 41.77%2 80.49% 79.95% 83.53% 51.77% 48.85%3 66.17% 65.29% 68.39% 48.67% 45.93%4 101.38% 100.51% 105.19% 75.70% 71.91%5 64.67% 63.71% 66.84% 67.34% 65.67%7 70.57% 72.29% 70.69% 61.73% 58.26%
Table 8.1. Maximum in-scene reflectance registered by the TM and ETM+ sensors for the five subscenes. Note: Reflectance values greater than 100% are possible due to path radiance and background effect illumination
87
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
TM1 TM2 TM3 TM4 TM5 TM7Band
Perc
ent o
f sat
urat
ed p
ixel
s
TM 1987TM 1989TM 1997ETM+ 2000ETM+ 2001
Figure 8.2. Percent saturated pixels for each band of the subscenes.
Band TM 1987 TM 1989 TM 1997 ETM+ 2000 ETM+ 20011 34.21% 43.59% 38.07% 29.47% 38.20%2 5.36% 6.98% 4.16% 25.35% 33.25%3 14.59% 17.70% 15.58% 28.53% 36.96%4 0.73% 0.70% 0.33% 8.44% 11.06%5 0.02% 0.02% 0.01% 0.01% 0.01%7 0.00% 0.00% 0.00% 0.00% 0.00%
Table 8.2. Percent saturated pixels for each band of the subscenes.
88
Spectral Endmembers Definition
Classification of satellite imagery is the process of remotely identifying target
types of ground covers; the information preserved in the pixels of multispectral imagery
(spectral signatures) is used to match spectral classes (groups of pixels with similar
spectral signature) to information classes (categories of land cover under investigations)
(Mather 1999). The purpose of this research was to identify and classify a set of
landcovers that may be used to describe the dynamics of glaciers.
Hall et al. (1987) and Hall et al. (1988) performed classifications of Landsat TM
scenes of the Grossglockner Glacier (Austria) using three information classes (Zone I, II,
III) to study environmental changes in alpine glaciers. For the accumulation zone Hall et
al. (1987) recognized a fresh snow zone (Zone III) and a wet snow zone (Zone II); the
ablation zone was classified as glacier ice (Zone I). Similarly, Rosenthal and Dozier
(1996) identified two major accumulation facies for the Sierra Nevada (USA) and
described them as winter snow (corresponding to fresh snow) and spring snow (a wet
aged snow facies). Klein and Isacks (1999) adopted a similar classification scheme to
analyze changes of the Zongo glacier (Cordillera Real, Bolivia), and argued that the
spectral classes associated with the glacier facies of Zongo Glacier may be exported to
perform a regional scale classification. Klein and Isacks (1999) suggested that it was
possible to apply locally developed spectral signatures at a regional scale. Extrapolation
of such spectral signatures is the basis of the study of this research.
Five information classes, based on a comparative reconstruction of the Klein and
Isacks (1999) study of Zongo Glacier were developed to classify the snow and ice cover
of the Cordillera Real. The spectral classes were based on a set of in-scene training sites
89
matching the glacier facies identified by Klein and Isacks (1999). Figure 8.3 shows the
three glacier facies Hall et al. (1987) identified on Zongo Glacier after Klein and Isacks
(1999). Figure 8.4 is a sample of some of the training sites over Zongo Glacier from the
1987 Landsat TM Image.
To increase the classifier algorithms’ performance in discriminating the glacier
facies, three more information classes were added (Klein and Isacks 1999): vegetation,
rock-soil, and shadows. The shadow class is necessary to obtain correct spectral un-
mixing in shaded areas, while vegetation and rock-soil are necessary to avoid
misclassification of these facies when present nearby or within glaciated covers (Klein
and Isacks 1999). Figure 8.5 shows the final spectral signatures applied to perform snow
Zone IZone II
Zone III
Figure 8.3. Accumulation (Zone III and Zone II) and ablation (Zone I) zones of Zongo Glacier as identified by Klein and Isacks (1999).
Figure 8.4. Sample of training sites used to obtain the spectral signatures for the information classes.
90
and ice cover classification on a regional (Cordillera Real) and temporal (multiple scene)
scale.
Binary classification
The simplest approach to identify snow and ice cover is based on binary
classifiers. The two resulting information classes isolate snow and ice covers from all
other materials. This methodology is useful as a first order classification to produce
snow-ice masks. Dozier (1989) proposed a technique based on the spectral characteristics
of snow and ice to develop a snow/ice mask from Landsat TM imagery. Klein and Isacks
Figure 8.5. Spectral signatures obtained from the training sites on Zongo Glacier.
Accumulation - Zone III
Accumulation - Zone II
Ablation - Zone I
Rock/Soil
Vegetation
Shadows
91
(1999) applied this technique to remove unwanted landcovers from Zongo Glacier. The
procedure is also particularly effective in identifying and removing clouds from glaciated
regions.
Dozier (1989) observed that snow and ice can be isolated from other land covers
on the base of the three following conditions:
Reflectance of TM band 1 greater than 15% (or a selected threshold between 15%
and 20%); this threshold isolates surfaces with high brightness, including non-snow
surfaces that are removed by the following two rules;
Reflectance of TM band 5 lower than 25% (or a selected threshold between 20%
and 25%); this rule distinguishes between clouds and snow, being that clouds are highly
reflective in the MIR domain, and snow is highly absorbing.
Normalized Difference Snow Index (NDSI) calculated from reflectance values
must be greater than 0.4; this procedure filters out bright soils, rocks and clouds.
The selection of the thresholds can be obtained by analysis of the atmospheric
conditions and spectral response of the sensors at the acquisition time, or by in-scene
observations of reflectance values (Dozier 1999). This binary classification was
effectively applied to the five Landsat scenes to estimate regional variations in snow and
ice landcover over time. Figure 8.6 shows a sample of the resulting binary classification
for Zongo Glacier. The binary classification was applied to the five project images but
the 2001 Landsat ETM+ was excluded from the figure for reasons explained in the next
chapter. The results are also discussed in the next chapter, but it is already evident from
figure 8.6 that from 1987 to 1997 there is a visible retreat of the glacier terminus, and a
loss of snow cover along the edges of the glaciated areas.
92
Supervised Classification: Minimum Distance
The minimum distance classifier was selected from among the possible
supervised classification techniques. One of the driving reasons for the selection of this
method was its computational power, being the fastest decision rule to calculate (ERDAS
1997). Another reason for selecting this method was the simple implementation of the
algorithm that does not require detailed statistical information about the endmembers; the
minimum distance classifier calculates the Euclidean distance of an unclassified pixel
from the spectral vectors of endmembers; endmembers can be obtained from spectral
libraries as predefined spectral signatures, and fed to the minimum distance algorithm
Figure 8.6. Binary classification of Zongo Glacier using the snow/ice mask after Dozier (1989) applied to four Landsat scenes.
Zongo Glacier terminus
93
without involving complex in-scene statistics (such is the case with the maximum
likelihood classifier) (RSI 2001). Since the endmembers from this project had to be
exported from one scene to the others, the minimum distance classifiers offered the best
and simplest classification solution. Figure 8.7 and figure 8.8 show samples of the
resulting minimum distance classification of Zongo Glacier.
The ETM+ band saturation is clearly visible in figure 8.8. The saturation
obliterates information for bright reflectors such as fresh snow; as a result, the classifiers
cannot detect the difference between Zone III and Zone II and returns Zone III as a global
class for the accumulation endmembers.
Ablation – Zone I
Accumulation – Zone III
Vegetation
Accumulation – Zone II Rock and Soil
Ablation – Zone IAblation – Zone I
Accumulation – Zone IIIAccumulation – Zone III
VegetationVegetation
Accumulation – Zone IIAccumulation – Zone II Rock and SoilRock and Soil
Figure 8.7. Minimum distance supervised classification of Zongo Glacier from Landsat TM 1987.
Ablation – Zone I
Accumulation – Zone III
Vegetation
Accumulation – Zone II Rock and Soil
Figure 8.8. Minimum distance supervised classification of Zongo Glacier from Landsat ETM+ 2000.
94
Spectral unmixing
Klein and Isacks (1999) found that the spectral un-mixing technique performed
well in defining the position of the transient snowline (or equilibrium line). Interpretation
of the resulting classification is more complex than that of supervised classifiers due to
the fact that the resulting information layers represent transitional fractions of
endmembers rather than absolute hard class membership. Each classified pixel, thus,
needs to be interpreted as a function of its endmember fractions. It is important to observe
that spectral unmixing is highly effective in obtaining information in shaded area and
transitional zones where hard classifiers, such as minimum distance, have to output an
absolute decision about pixel membership. Through spectral un-mixing, for example, it is
possible to obtain the fractional amount of snow and ice present in proximity to (or
within) vegetation, rock, and soil landcovers, or in shaded areas. Rosenthal and Dozier
(1996) successfully applied the spectral un-mixing technique to obtain detailed
information relative to snow fraction in the Sierra Nevada (USA).
Figure 8.9 and figure 8.10 represent the results of a spectral un-mixing
classification performed on the 1987 Landsat TM. The results of the classification match
the results obtained by Klein and Isacks (1999) and confirm that the spectral signatures of
the endmembers are correctly calibrated to describe the information classes for the
accumulation and ablation zones. From the spatial transect (figure 8.10) it can be
observed that the transient snowline is estimated at the intersection of the accumulation
and ablation endmember fraction curves. In spatial terms, this spectral intersection
represents the location at which there is a transition from snow to ice cover, and thus
from accumulation to ablation.
95
Spectral un-mixing failed to give meaningful results when the endmembers
obtained from the Landsat TM scenes were exported to the Landsat ETM+ scenes due to
the distortion induced by the band saturation on the spectral signatures. This result
enforced the idea that ETM+ spectrometers in high-gain mode are not fit for snow and ice
facies classification based on spectral information.
A
B
A
B
Ablation endmember Accumulation endmember
Figure 8.9. Accumulation and ablation endmembers of Zongo Glacier, and AB transect (figure 8.10). The fraction is represented in grey scale with white as the highest and black as lowest .
Transient snowline
96
Figure 8.10. Accumulation and ablation endmember fractions along the AB transect (figure 8.9).
Accumulation Ablation
Transient snowline
A BSpatial transect
97
Chapter 9
Results
Zonal statistics provide a meaningful measure of the magnitude and spatial
pattern of the environmental changes of the glaciers of the Cordillera Real. Zonal
functions are used to calculate raster datasets where each pixel value depends on the
value of the source cell (information layer) and the association that the source cell has
within a cartographic zone (zone layer). In particular, zonal functions can be used to
calculate descriptive statistics of classified satellite imagery based on predefined
cartographic areas (ESRI 2003). For this project, the binary and the supervised minimum
distance classification outputs provided the information layers, and Jordan’s 1984 glacier
inventory (Jordan 1991) the zone layer. Each zone was defined by the glacier ID found in
Jordan’s 1984 inventory. ESRI ArcGIS 8.3 Spatial Analyst offered a powerful tool to
obtain zonal statistics of the classification results.
Binary Classification: Zonal Statistics
Zonal statistics applied to the snow and ice binary classification provided the
number of snow and ice covered pixels within each glacier zone. From the summary of
the statistics (table 9.1 and figure 9.1) it is clear that the amount of snow and ice cover in
98
the Cordillera Real was at minimal levels in 2000, with greater amounts of snow and ice
in 1987, 1997, 1989, 2001 (figure 9.1).
Relative snow and ice cover differences were calculated for the year pairs 1987-
1989, 1989-1997, 1997-2000, and 2000-2001 in order to quantify total and zonal surface
gains and losses. Table 9.2 illustrates the changes in glacial ice and snow cover for the
respective time periods by number and percent. It can be observed that between 1989 and
Figure 9.1. Total snow/ice cover of the Cordillera Real measured by binary classification.
1987
1989
1997
2000
2001
210
220
230
240
250
260
270
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
Table 9.1. Snow/ice area measured from the zonal statistics of the binary classification.
257,839,828220,392,666 236,320,889250,747,261233,632,341Total Snow/ice cover (m2)
317,439271,336 290,946308,707287,636Total Snow/ice cover (pixels) 20012000 199719891987Reference year
99
1997, and between 1997 and 2000, 55% and 63% of the glaciers of the Cordillera Real
lost snow and ice surface respectively. From 1987 to 1989, and from 2000 to 2001, the
number of glaciers that gained snow and ice surface was estimated to be 72% and 84%
respectively. The maximum surface loss was registered from 1997 to 2000 (15,928,223
m2), followed by the 1989-1997 negative balance (14,426,372 m2) for a total surface
reduction of 30,354,595 m2 from 1989 to 2000. Between 1987 and 1989, and between
2000 and 2001, the glaciers registered a net gain in area. The inter-scene percent of snow
and ice cover lost or gained is shown in table 9.3.
Reference year 1987 1989 1997 2000 2001Total Snow/ice cover (pixels) 287,636 308,707 290,946 271,336 317,439 Total Snow/ice cover (sqm) 233,632,341 250,747,261 236,320,889 220,392,666 257,839,828
1987-1989 1989-1997 1997-2000 2000-20001
Net areal balance (pixels) 21071 -17761 -19610 46103Net areal balance (sqm) 17,114,920 -14,426,372 -15,928,223 37,447,162
Percent change of snow/icecover from reference year 7.3% -5.8% -6.7% 17.0%
Table 9.3. Inter-scene snow/ice cover percent changes.
1987-1989 1989-1997 1997-2000 2000-20001Number of Glaciers w ith Snow/ice loss 135 495 563 54Percent of Glaciers w ith Snow/ice loss 15.1% 55.4% 63.0% 6.0%
Number of Glaciers w ith Snow/ice gain 640 300 215 748Percent of Glaciers w ith Snow/ice gain 71.7% 33.6% 24.1% 83.8%
Number of Glaciers w ith Snow/ice unchanged 118 98 115 91Percent of Glaciers w ith Snow/ice unchanged 13.2% 11.0% 12.9% 10.2%
Total area losses (pixels) -1653 -24208 -25987 -787Total area losses (sqm) -1,342,649 -19,662,948 -21,107,941 -639,241
Total area gains (pixels) 22724 6447 6377 46890Total area gains (sqm) 18,457,569 5,236,576 5,179,718 38,086,403
Net areal balance (pixels) 21071 -17761 -19610 46103Net areal balance (sqm) 17,114,920 -14,426,372 -15,928,223 37,447,162
Table 9.2. Inter-scene zonal statistics illustrating gain, losses and total net balance.
100
The losses between 1997 and 2000 corresponded to 6.7% of the cover registered
in 1997. Similarly, the losses between 1989 and 1997 were 5.8% of the 1989 glaciated
surface.
Minimum Distance: Zonal Statistics
A second analysis applied zonal statistics to the results of the supervised
minimum distance classification. The global trend of changes highlighted by the zonal
statistics applied to the supervised minimum distance classification was similar to the
binary classification results (table 9.4 and figure 9.2). Year 2000 presented the lowest
snow and ice cover extent followed by 1987, 1997, 1989 and 2001.From an inter-scene
point of view (table 9.5 and table 9.6), zonal statistics showed that the largest amount of
snow and ice cover was lost between 1989 and 1997, when 56% of the glaciers lost a
total of 14,471,046 m2 of snow and ice surface, corresponding to 5.6% of the 1989
surface extent; between 1997 and 2000, 57% of the glaciers showed a reduction of the
snow and ice surface for a total of 5,263,380 m2, and corresponding to 2.2% of year
1997. The periods from 1987 to 1989 and from 2000 to 2001 had snow and ice surface
gain 6.6% of 1987 and 13.4% of 2000 respectively.
Reference year 1987 1989 1997 2000 2001Total Snow/ice cover (pixels) 296,836 316,353 298,537 282,174 321,701Total Snow/ice cover (sqm) 241,105,041 256,957,724 242,486,678 229,195,832 261,301,637
Table 9.4. Snow/ice area measured from the zonal statistics of the minimum distance classification
101
Figure 9.2. Total snow/ice cover of the Cordillera Real measured by minimum distance classifications.
2000
2001
1997
1989
1987
210
220
230
240
250
260
270
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
102
Detailed change detection analysis (table 9.7a through 9.7d) provided information
about the interclass variations; the columns present the classes in the initial state, while
the rows represent the class distribution at final state. Class changes reflect environmental
dynamics and help to identify change anomalies and misclassifications. For example, a
summary of the net class differences between the scenes suggested that the surface gains
registered between 1987-1989 and 2000-2001, counterbalanced by a loss of rock-soil and
1987-1989 1989-1997 1997-2000 2000-20001Number of Glaciers w ith Snow/ice loss 129 500 577 60Percent of Glaciers w ith Snow/ice loss 14.4% 56.0% 64.6% 6.7%
Number of Glaciers w ith Snow/ice gain 646 282 190 735Percent of Glaciers w ith Snow/ice gain 72.3% 31.6% 21.3% 82.3%
Number of Glaciers w ith Snow/ice unchanged 118 111 123 96Percent of Glaciers w ith Snow/ice unchanged 13.2% 12.4% 13.8% 10.8%
Total area losses (pixels) -1304 -23895 -21887 -784Total area losses (sqm) -1,059,174 -19,408,714 -17,777,716 -636,804
Total area gains (pixels) 20821 6079 5524 40311Total area gains (sqm) 16,911,857 4,937,668 4,486,869 32,742,610
Net areal balance (pixels) 19517 -17816 -16363 39527Net areal balance (sqm) 15,852,683 -14,471,046 -13,290,847 32,105,806
Table 9.5. Inter-scene zonal statistics from minimum distance classification illustrating gain, losses and total net balance.
Reference year 1987 1989 1997 2000 2001Total Snow/ice cover (pixels) 296,836 316,353 298,537 282,174 321,701Total Snow/ice cover (sqm) 241,105,041 256,957,724 242,486,678 229,195,832 261,301,637
1987-1989 1989-1997 1997-2000 2000-20001Net areal balance (pixels) 19,517 -17,816 -16,363 39,527Net areal balance (sqm) 15,852,683 -14,471,046 -13,290,847 32,105,806
Percent change of snow/icecover from reference year 6.6% -5.6% -5.5% 14.0%
Table 9.6. Inter-scene snow/ice cover percent changes determined from minimum distanceclassification.
103
shadows classes, might not represent glacial activity but extended snow falls events (table
9.8 and figure 9.3).
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 89,329 1,466 23,063 16,534 23,215 153,607Vegetation 796 151 243 83 1,088 2,361Rock-Soil 11,267 563 104,735 1,336 12,474 130,375Accumulation 39,005 168 7,578 152,840 1,828 201,419Shadow 10,244 1,153 8,927 672 129,738 150,734
Class Total 150,641 3,501 144,546 171,465 168,343Class Changes 61,312 3,350 39,811 18,625 38,605Image Difference 2,966 -1,140 -14,171 29,954 -17,609
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 59.3% 41.9% 16.0% 9.6% 13.8%Vegetation 0.5% 4.3% 0.2% 0.0% 0.6%Rock-Soil 7.5% 16.1% 72.5% 0.8% 7.4%Accumulation 25.9% 4.8% 5.2% 89.1% 1.1%Shadow 6.8% 32.9% 6.2% 0.4% 77.1%
Class Total 100.0% 100.0% 100.0% 100.0% 100.0%Class Changes 40.7% 95.7% 27.5% 10.9% 22.9%Image Difference 2.0% -32.6% -9.8% 17.5% -10.5%
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 72,557,480 1,190,759 18,732,922 13,429,742 18,856,384 124,767,286Vegetation 646,551 122,650 197,377 67,417 883,728 1,917,722Rock-Soil 9,151,621 457,297 85,071,004 1,085,166 10,132,007 105,897,094Accumulation 31,681,811 136,458 6,155,231 124,144,290 1,484,793 163,602,583Shadow 8,320,689 936,524 7,250,956 545,832 105,379,691 122,433,692
Class Total 122,358,152 2,843,687 117,407,489 139,272,446 136,736,602Class Changes 49,800,672 2,721,038 32,336,485 15,128,156 31,356,911Image Difference 2,409,134 -925,965 -11,510,395 24,330,137 -14,302,910
Pixel Counts (1987-1989)
Percentages (1987-1989)
Area (Square Meters) (1987-1989)
Table 9.7a. Change detection statistics from 1987-1989
104
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 89,207 966 16,754 29,656 13,230 149,813Vegetation 1,934 106 1,216 114 1,864 5,234Rock-Soil 14,698 201 95,110 1,157 5,119 116,285Accumulation 21,083 135 3,449 168,192 1,656 194,515Shadow 26,685 953 13,846 2,300 128,865 172,649
Class Total 153,607 2,361 130,375 201,419 150,734Class Changes 64,400 2,255 35,265 33,227 21,869Image Difference -3,794 2,873 -14,090 -6,904 21,915
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 58.1% 40.9% 12.9% 14.7% 8.8%Vegetation 1.3% 4.5% 0.9% 0.1% 1.2%Rock-Soil 9.6% 8.5% 73.0% 0.6% 3.4%Accumulation 13.7% 5.7% 2.6% 83.5% 1.1%Shadow 17.4% 40.4% 10.6% 1.1% 85.5%
Class Total 100.0% 100.0% 100.0% 100.0% 100.0%Class Changes 41.9% 95.5% 27.0% 16.5% 14.5%Image Difference -2.5% 121.7% -10.8% -3.4% 14.5%
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 72,458,386 784,634 13,608,437 24,088,086 10,746,068 121,685,609Vegetation 1,570,892 86,099 987,696 92,597 1,514,034 4,251,317Rock-Soil 11,938,451 163,262 77,253,098 939,773 4,157,908 94,452,491Accumulation 17,124,667 109,654 2,801,450 136,613,952 1,345,086 157,994,809Shadow 21,674,891 774,074 11,246,414 1,868,175 104,670,596 140,234,150
Class Total 124,767,286 1,917,722 105,897,094 163,602,583 122,433,692Class Changes 52,308,900 1,831,624 28,643,996 26,988,631 17,763,095Image Difference -3,081,677 2,333,594 -11,444,603 -5,607,774 17,800,459
Pixel Counts (1989-1997)
Percentages (1989-1997)
Area (Square Meters) (1989-1997)
Table 9.7b. Change detection statistics from 1989-1997
105
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 75,309 1,308 11,914 28,875 19,005 136,411Vegetation 1,302 131 533 433 806 3,205Rock-Soil 23,027 1,008 89,608 8,574 13,123 135,340Accumulation 27,022 148 1,029 152,669 1,816 182,684Shadow 23,153 2,639 13,201 3,964 137,899 180,856
Class Total 149,813 5,234 116,285 194,515 172,649Class Changes 74,504 5,103 26,677 41,846 34,750Image Difference -13,402 -2,029 19,055 -11,831 8,207
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 50.3% 25.0% 10.2% 14.8% 11.0%Vegetation 0.9% 2.5% 0.5% 0.2% 0.5%Rock-Soil 15.4% 19.3% 77.1% 4.4% 7.6%Accumulation 18.0% 2.8% 0.9% 78.5% 1.1%Shadow 15.5% 50.4% 11.4% 2.0% 79.9%
Class Total 100.0% 100.0% 100.0% 100.0% 100.0%Class Changes 49.7% 97.5% 22.9% 21.5% 20.1%Image Difference -8.9% -38.8% 16.4% -6.1% 4.8%
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 61,169,735 1,062,423 9,677,147 23,453,719 15,436,811 110,799,835Vegetation 1,057,550 106,405 432,929 351,704 654,674 2,603,261Rock-Soil 18,703,681 818,748 72,784,098 6,964,232 10,659,157 109,929,915Accumulation 21,948,620 120,213 835,805 124,005,395 1,475,046 148,385,079Shadow 18,806,024 2,143,528 10,722,512 3,219,759 112,008,463 146,900,286
Class Total 121,685,609 4,251,317 94,452,491 157,994,809 140,234,150Class Changes 60,515,874 4,144,912 21,668,393 33,989,414 28,225,688Image Difference -10,885,775 -1,648,055 15,477,424 -9,609,730 6,666,136
Pixel Counts
Percentages
Area (Square Meters)
Table 9.7c. Change detection statistics from 1997-2000
106
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 106,199 1,506 30,755 18,830 41,507 198,797Vegetation 584 81 739 54 1,600 3,058Rock-Soil 5,341 428 83,925 231 13,076 103,001Accumulation 26,444 176 5,258 140,409 2,060 174,347Shadow 7,230 487 8,422 292 142,862 159,293
Class Total 145,798 2,678 129,099 159,816 201,105Class Changes 39,599 2,597 45,174 19,407 58,243Image Difference 52,999 380 -26,098 14,531 -41,812
Ablation Vegetation Rock-Soil Accumulation ShadowAblation 72.8% 56.2% 23.8% 11.8% 20.6%Vegetation 0.4% 3.0% 0.6% 0.0% 0.8%Rock-Soil 3.7% 16.0% 65.0% 0.1% 6.5%Accumulation 18.1% 6.6% 4.1% 87.9% 1.0%Shadow 5.0% 18.2% 6.5% 0.2% 71.0%
Class Total 100.0% 100.0% 100.0% 100.0% 100.0%Class Changes 27.2% 97.0% 35.0% 12.1% 29.0%Image Difference 36.4% 14.2% -20.2% 9.1% -20.8%
Ablation Vegetation Rock-Soil Accumulation Shadow Class TotalAblation 86,260,138 1,223,249 24,980,749 15,294,668 33,714,061 161,472,863Vegetation 474,354 65,792 600,253 43,862 1,299,600 2,483,861Rock-Soil 4,338,227 347,643 68,168,081 187,630 10,620,981 83,662,562Accumulation 21,479,139 142,956 4,270,811 114,047,210 1,673,235 141,613,351Shadow 5,872,568 395,566 6,840,770 237,177 116,039,660 129,385,739
Class Total 118,424,426 2,175,206 104,860,663 129,810,546 163,347,536Class Changes 32,164,288 2,109,413 36,692,582 15,763,336 47,307,877Image Difference 43,048,438 308,655 -21,198,101 11,802,805 -33,961,797
Pixel Counts (2000-2001)
Percentages (2000-2001)
Area (Square Meters) (2000-2001)
Table 9.7d. Change detection statistics from 2000-2001
107
Figure 9.3. Surface changes in km2 for each scenes pair obtained from the minimum distance classification.
-40
-30
-20
-10
0
10
20
30
40
50
1987-1989 1989-1997 1997-2000 2000-2001
Years
Surf
ace
chan
ges
(sq.
Km
)
AblationAccumulationShadowRock-SoilVegetation
Ablation Accumulation Shadow Rock-Soil Vegetation1987-1989 2,409,134 24,330,137 -14,302,910 -11,510,395 -925,9651989-1997 -3,081,677 -5,607,774 17,800,459 -11,444,603 2,333,5941997-2000 -3,261,184 -28,184,263 23,113,386 10,408,172 -2,076,1112000-2001 43,048,438 11,802,805 -33,961,797 -21,198,101 308,655
Table 9.8. Detail of class areal changes (sq. km)
108
The interclass changes of the ablation zones follow two main patterns (figure 9.4).
The first pattern is common for the pairs 1989-1997 and 1997-2000, where the general
loss of snow and ice cover correspond in the loss of ablation zone and a gain of rock-soil,
accumulation and shadows surface (which includes few misclassified water pixels); The
second pattern, common to the pairs 1987-1989 and 2000-2001, shows a general gain in
snow and ice surface, and is characterized by a net loss of ablation class in favor of
accumulation class.
Observing the variation of the rock-soil class (figure 9.3 and figure 9.4) it could
be argued that the snow and ice cover gains in the 1987-1989 and 2000-2001 pairs are not
the result of glacial processes, but the product of widespread episodes of late season snow
fall that simulate an increase of accumulation area and a reduction of coverage of all the
Figure 9.4. Percent change from ablation class to other classes.
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
Ablation Vegetation Rock-Soil Accumulation Shadow
Class
Perc
ent C
hang
e
Ablation 87-89
Ablation 89-97
Ablation 97-00
Ablation 00-01
109
other classes. This phenomenon is supported by field and remote sensed observation that
will be discussed in chapter 10.
Binary vs. Minimum Distance Classifier
The trend vectors of the two classifications methods are parallel in direction and
different in magnitude (figure 9.5); this result is significant because it is the outcome of
classifications based on two very different methodologies: the binary classification results
from constraint imposed on spectral reflectance, while the minimum distance classifier
operates on pattern recognition of spectral signatures.
The trend directions are parallel, so it can be argued that the surface changes
measured by the two classifications are, in fact, due to physical changes detected from the
Figure 9.5. Comparison of the total snow/ice cover of the Cordillera Real measured by binary and minimum distance classifications.
2001
2000
1997
1987
1989
210
220
230
240
250
260
270
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
Binary
Minimum Distance
110
five scenes. The difference in magnitude, in this case, is less relevant because the
direction of the change is the more significant variable for understanding environmental
dynamics. The difference in the magnitude can be attributed to the sensitivity of the
methodologies applied to detect the change. The binary classification was found to be
highly selective, because its rules operate as a bandpass filter with well defined cut-off
thresholds; this was particularly noticeable in shaded areas, where the binary classifier
underestimates snow and ice cover. On the other hand, the minimum distance classifier is
more sensitive and less selective, allowing the recognition of snow and ice cover in
shaded and mixed pixel areas, resulting in a larger number of pixels classified as snow
and ice cover. By varying the reflectance thresholds, the binary classification could be
tuned to match the minimum distance results.
Regional vs. Local Observations
Zongo Glacier (68.14°W, 16.27°S, approximately 3 km2), Chacaltaya Glacier
(68.12°W, 16.35°S, approximately 0.5 km2), and the unnamed glacier feeding Laguna
Glaciar (68.55°W, 15.83°S, approximately 4.3 km2) (from here referred as Laguna
Glaciar for simplicity) represent three retreating glaciers of the Cordillera Real that have
been observed under field monitoring since the 1980s. Mass balance records for Zongo
and Chacaltaya cover the period from 1991 to 2002 (Francou et al 2003). During the past
6 years, Laguna Glacial retreat has been observed in the field, but not measured, by the
Geography Department of Appalachian State University, NC.
Laguna Glaciar (figure 9.6) and Zongo Glacier (figure 9.7) present a surface
variation pattern that closely reflects the regional trend of the Cordillera Real. Chacaltaya
111
Glacier (figure 9.8), while showing a general retreat pattern, presents some anomalies.
The apparent surface increase of Chacaltaya Glacier registered from 1987 to 1997, and
the sudden drop in 2000 can be explained by the fact that the glacier surface is
considerably smaller than the other two glaciers, rendering the estimate of the real glacier
surface particularly sensitive to sudden environmental changes such as local snowfall
events. However, similar to the regional trend, the three glaciers presented a starting low
snow and ice cover surface in the year 1987, followed by an increase between 1987 and
1989, and a significant surface loss during the period 1997-2000. In the year 2001, a
significant increase of the snow and ice cover was recorded for the three glaciers
confirming the observations conducted at a regional scale.
3
3.05
3.1
3.15
3.2
3.25
3.3
3.35
3.4
3.45
3.5
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
Binary maskMinimum Distance
Figure 9.6. Laguna Glaciar: Comparison of the total snow/ice measured by binary and minimum distance classifications.
112
Figure 9.7. Zongo Glacier: Comparison of the total snow/ice measured by binary and minimum distance classifications.
2.3
2.35
2.4
2.45
2.5
2.55
2.6
2.65
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
Binary maskMinimum Distance
113
0.2
0.22
0.24
0.26
0.28
0.3
0.32
0.34
1986 1988 1990 1992 1994 1996 1998 2000 2002
Year
Tota
l sno
w/ic
e co
ver (
sq. k
m)
Binary maskMinimum Distance
Figure 9.8. Chacaltaya Glacier: Comparison of the total snow/ice measured by binary and minimum distance classifications.
114
Chapter 10
Discussion and Conclusion
Discussion
The results of the snow and ice classification presented in chapter 9 showed the
presence of two predominant trends in the dynamic evolution of the glaciers of the
Cordillera Real recorded by the five Landsat scenes. From 1987 to 1989 and from 2000
to 2001 the snow and ice cover seems to increase, while during the period 1989 to 1997
and from 1997 to 2000, a progressive reduction of the snow and ice cover is registered.
From the analysis of change detection statistics, it can be argued that the snow and ice
cover variations registered for the pairs 1987-1989 and 2000-2001 are, in fact, not the
result of glacial processes but extended short-term snow events. Fresh snow presents
similar spectral characteristics to the accumulation classes describes as Zone II and Zone
III. As a result, snow falls are classified as accumulation zone and the global snow and
ice cover for that scene is larger. Also, fresh snow that reaches elevations below the
glacier terminus obliterates the spectral signatures of other classes by covering ice, rock-
and soil and vegetation. These effects are visible in the global change detection statistics
for the pair 1987-1989 and 2000-2001, where the general increase of snow and ice cover
is counterbalanced by a reduction of the ablation and rock-soil classes. On the other hand,
retreat of glacier surfaces should be reflected as a general reduction of the snow and ice
115
cover balanced by a relative increase of non-glacier classes such as rock-soil, as visible
for the pairs 1989-1997 and 1997-2000.
These important observations obtained from in-scene analysis are confirmed by
independent satellite data and field measurements. An ASTER satellite scene acquired
June 29th, 2001 confirmed that a conspicuous snowfall event took place before Landsat
ETM+ acquired the July 31, 2001 scene of the Cordillera Real (figure 10.1). The
extended snowfall maintained a substantial cover at elevations lower than the glacier
termini, resulting in an apparent increase of the glacial surface. The snow event is also
reported in field measurements by Francou et al. (2003) who registered a positive mass
balance for Chacaltaya glacier in the year 2000. A similar snowfall event has been
assumed to describe the snow and ice cover increase for the pair 1987-1989. It was
observed that the pattern of changes for the ablation class detected for the pair 1987-1989
is consistently similar to that the pair 2000-2001 (chapter 9).
Figure 10.1. ASTER scene (left) acquired on 06/29/2001 and Landsat ETM+ scene (right) acquired on 07/31/2001 representing the central portion of the Cordillera Real after an extended snow fall event (USGS Global Visualization Viewer, http://glovis.usgs.gov).
116
It may thus be argued that the 1989 Landsat TM and 2001 Landsat EMT+ scenes
are not suitable to identify glacial retreat processes for the study area. The 1989 snowfall
recorded by the Landsat TM image appears to be more localized and the image could be
used for local studies of unaffected areas, while the 2001 event in the Landsat ETM+
presented a widespread regional disturbance. Consequently, within the five images
dataset, the 1987 and the 2000 scenes represent the closest portrait of undisturbed snow
and ice cover extend for the Cordillera Real, and will represent the reference to estimate
the total snow and ice cover change during the period of time.
Table 10.1 illustrates the snow and ice areal losses within the Cordillera Real
considering 1987 as the initial state and the year 2000 as the final; it was possible to
estimate a total snow and ice cover loss of 5.7% (binary classification) and 4.9%
(minimum distance) during this 13 year interval. The general trend is in agreement with
observations reported for example, by Thompson and Davis (1998), Kaser (1999),
Haeberli (2003), and Francou et al (2003). The observations conducted at local scale on
Zongo, Chacaltaya and Laguna Glaciar reflect the general regional trend, but also
emphasize the net dynamic contrast between larger glaciers (surface greater than 0.5 km2)
and smaller glaciers (less than 0.5 km2).
Classification TypeReference year 1987 2000 1987 2000Total Snow/ice cover (pixels) 287,636 271,336 296,836 282,174Total Snow/ice cover (sqm) 233,632,341 220,392,666 241,105,041 229,195,832
1987-2000 1987-2000Net areal balance (pixels) -16,300 -14,662Net areal balance (sqm) -13,239,675 -11,909,210
Percent change of snow/ice cover -5.7% -4.9%
Table 10.1. Inter-scene snow/ice cover percent losses.
Binary Minimum Distance
117
Francou et al. (2003) report that 80% percent of the glaciers of the Cordillera Real
have a surface area less than 0.5 km2 (79% in the subset of the Cordillera Real used for
this research thesis). Due to the ongoing climatic and environmental changes, these
glaciers are close to their critical mass, beyond which, the retreating process becomes
irreversible and the solid precipitation is not able to support the generation of new ice
mass (Kaser 1999, Haeberli et al. 1999, Thompson 2000, Francou et al. 2003).
Chacaltaya Glacier is a representative example of a glacier with surface area near 0.5
km2. It can be observed that the surface loss for Chacaltaya Glacier was estimated
between 11.9% and 18.6% (table 10.2), values significantly higher than the regional
estimates. The snow and ice cover loss measurements performed by Landsat imagery for
Chacaltaya is in agreement with field observations that predict the extinction of the
glacier within the next 30 years (Thompson and Davis 1998, Thompson 2000, Francou et
al. 2003).
The higher sensitivity of small glaciers such as Chacaltaya is also supported by
statistical observations performed on a subset of glaciers of the Cordillera Real
characterized by surface areas between 0.25 km2 and 0.75 km2. Depending on the
classification methodology (binary or minimum distance), the zonal statistics applied to
the 122 glaciers subset (representing 14% of the total number) showed a total loss of
Binary Mask Minimum DistanceNet areal balance (pixels) -39 -63Net areal balance (sqm) -31,678 -51,172
Percent change of snow/icecover from reference year -11.9% -18.6%
Chacaltaya1987-2000
Table 10.2. Inter-scene snow and ice cover change estimates for Chacaltaya Glacier.
118
snow and ice cover of 9% to 10%. Also, of the 122 glaciers, 75% to 79% presented snow
and ice cover losses, with an average surface loss of 16%.
On the other hand, Zongo Glacier (table 10.3), Laguna Glaciar (table 10.4), and a
subset of 20 glaciers with area between 2 km2 and 5 km2 presented a surface loss close to
the regional estimate. The total surface loss was estimated between 2% to 3%, but 76% to
80% of the 20 glaciers showed an average surface loss of 4%. These results are in
agreement with field measurements performed by Francou et al. (2003) on Zongo
Glacier, and confirm that these glaciers, while retreating, could still recover ice mass in
favorable environmental and climatic conditions.
Technical, and Physical Limitations of the Results
Satellite imagery from the Landsat platforms TM and ETM+ have been used in
many projects to study single glaciers and measure snow and ice cover extent in alpine
environments. The relationship between snow and ice cover changes detected by satellite
Binary Mask Minimum DistanceNet areal balance (pixels) -215 -140Net areal balance (sqm) -174,634 -113,715
Percent change of snow/icecover from reference year -6.9% -4.4%
Zongo Glacier1987-2000
Table 10.3. Inter-scene snow and ice cover change estimates for Zongo Glacier.
Table 10.4. Inter-scene snow and ice cover change estimates for Laguna Glaciar.
Binary Mask Minimum DistanceNet areal balance (pixels) -79 -219Net areal balance (sqm) -64,168 -177,883
Percent change of snow/icecover from reference year -2.0% -5.5%
Laguna Glaciares1987-2000
Glacial
Binary Mask Minimum DistanceNet areal balance (pixels) -79 -219Net areal balance (sqm) -64,168 -177,883
Percent change of snow/icecover from reference year -2.0% -5.5%
Laguna Glaciares1987-2000
Glacial
119
platforms at a regional scale and the estimate of glacier mass balance is, however, subject
to complex interactions between numerous environmental variables (Haeberli 2003).
Remote sensing based estimates of mass balance changes provide the advantage of being
faster than regional scale field studies, a major advantage when considering accelerating
global warming processes that are shortening the life of a number of glaciated regions of
the world. In addition, many alpine glaciers are remote and difficult to access on foot.
Also, in Bolivia, difficult political events during the latter half of the 2003 have further
reduced access for field studies in the Cordillera Real. This research study has found a list
of issues that affect the interpretation of quantitative estimates of glacier mass balance at
the regional scale from snow and ice cover change detection. The issues identified were
of both a technical and a physical nature.
The first problem encountered in this research study was related to the technical
limitations of the Landsat TM and ETM+ spectrometers. Landsat platforms were
developed to study vegetation; thus, any attempt to implement Landsat imagery for other
purposes may go beyond the capability of the sensors. This research emphasized that
Landsat platforms present important spectral limitations that affect the accuracy of the
detection of snow and ice facies within the tropics. Landsat TM sensors saturate in the
visible and NIR domain over snow cover, but maintain sufficient spectral response to
allow the identification of major glacial facies such as fresh snow, firn, and glacial ice.
Landsat ETM+ performance is dramatically affected by LTAP defined gain rules; this
study showed that ETM+ spectrometers in high-gain mode failed to discriminate the
accumulation facies (fresh snow to firn) due to the high saturation in the visible and NIR
bands. As a result, the identification of snow and ice cover is inherently limited by
120
classification errors. Also, it was observed that snow and ice spectral libraries developed
from TM imagery perform poorly when applied to ETM+ scenes, rendering impossible
the development of glacial models based on physical measurements such as reflectance
beyond the available Landsat TM dataset.
Another limitation in the estimation of snow and ice cover to study glacier
dynamic is of a physical nature. Satellite scenes are single shot images that represent the
Earth within well defined time and space coordinates, and do not contain information to
describe dynamic processes. In the specific case of glaciers, it may be correctly argued
that the temporal resolution of Landsat satellites is higher than the rate of change of the
glaciers. Unfortunately, physical atmospheric processes, such as cloud cover, greatly
affect the availability of clear images, especially in alpine environments. Also, if seasonal
constraints are added to the selection of suitable imagery for glaciological application, the
number of possible images available in the Landsat dataset may be reduced from a few to
none in any given year. In the case of Landsat ETM+, acquisition plans such as LTAP
may further reduce availability of imagery for remote areas.
Another physical limitation is related to atmospheric events such as extended
snowfall events that hide the glaciated area with a uniform cover. In extreme cases, such
as for the 2001 Landsat image in this study, early winter snowfall obliterated the real
extent of accumulation and ablation zones and provided false results about the effective
snow and ice cover extent of the glaciers of the Cordillera Real. In less extreme cases,
such as for the 1989 scene (and partially for the 1997 scene), snowfall may affect only a
portion of the study area, or provide misleading detection of the equilibrium line of the
glaciers. Again, the sum of the physical limitations may drastically reduce the availability
121
of suitable imagery. As a consequence, a large amount of ancillary data (precipitation and
snow falls, cloud cover, etc.) is required to determine the most appropriate year, season
and day to acquire a suitable satellite image, which may not correspond to the acquisition
plans, temporal resolution, and availability of a specific satellite platform dataset.
Ancillary data may not be available for remote areas where remote sensing studies are
best applied, leading to a try-fail process for the selection of images from a dataset. All
the physical limitations mentioned were encountered during the acquisition and the
analysis of the data for this research study.
While methodological limitations concerning classification techniques can be
improved from ground control samples, and glacier dynamic estimation models can be
optimized with further understanding of the processes involved in mass balance
variations, the lack of suitable data and spectral capability cannot be readily fixed,
especially when the research needs to address past events.
Conclusions
The research conducted by me and reported herein produced an estimate of the
snow and ice cover changes affecting the glaciers of the Cordillera Real. The results are
in line with field observations of a gradual but quantifiable retreat from 1987. The snow
and ice cover of the Cordillera Real is highly sensitive to changes in regional weather
patterns such as those induced by El Niño events, but the progressive retreat of the
glaciers may also be related to long-term global change processes in the troposphere such
as the increasing planetary temperature.
122
The glaciers of the Cordillera Real developed in a tropical climate which is
characterized by a homogenous atmosphere and well defined temperature ranges.
Consequently, it could be argued that the measured loss of snow and ice surface, the
reduction of glacial mass, and the migration of the equilibrium line at higher elevation,
reflect re-adjustments of tropical environmental balances at regional scale. While the
cause of these environmental changes is currently under debate, it is important to observe
that the changes are significantly visible and have been monitored and measured.
Remote sensing based technologies may be implemented to identify and classify
glacier facies of remote areas, and to estimate environmental changes from the
measurement of the variation of physical properties of the glacier surface. A number of
limitations, though, affect the quality of the estimates, and must be taken into account
when discussing the resulting environmental change estimates. The error of estimate is
characteristic of any remotely sensed measurement due to the nature of the measurement
itself, which is not directly taken from on the object, but inferred from afar.
The accuracy of remotely sensed snow and ice cover is difficult to assess due to
the transitional nature of snow, and the continued overlapping variation of its surface
extent which requires continuous in-field monitoring. This research focused on the
necessity of developing a methodology to readily obtain measurements of snow and ice
changes to describe glacier retreat in remote areas, accepting the inherent error of
estimate resulting from the study of unexplored and extended areas.
This study confirmed that, while not developed for glaciological applications,
Landsat platforms are a valuable resource to study snow and ice cover, but present
serious limitations in snow covered tropical alpine environments. Mass balance variations
123
of the glaciers have been recognized as an optimum indicator of global environmental
changes; glaciers are also an important source for water for hydroelectric power and
water supply for a significant amount of population in alpine areas worldwide. It is thus
necessary to plan and develop satellite platforms dedicated to measure spatial and
temporal variations of the snow and ice cover of the glacier; dedicated satellite systems
would make it possible to estimate quickly and more accurately the environmental
variations driven by global changes processes, and would provide higher quality data to
ensure the development of prompt response plans to mitigate those changes that could
have negative impacts on human and natural environments.
124
Bibliography
References Cited
Bamber, J. 2003. A review of current and future remote sensing techniques for mass balance determination. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Bennet, M.R., N.F. Glasser. 1996. Glacial Geology: Ice sheets and Landforms. John Wiley & Sons, New York.
Bindschadler, R., J. Dowdeswell, D. Hall and J. Winther. 2001. Glaciological applications with Landsat-7 imagery: Early assessment. Remote Sensing and Environment 78: 163-179.
Bishop, M.P., and J.D. Colby. 2002. Anisotropic reflectance correction of SPOT-3 HRV imagery. International Journal of Remote Sensing 23(10):2125-2131
Braithwaite, R.J. 2002. Glacier mass balance: the first 50 years of international monitoring. Progress in Physical Geography 26 (1): 76-95.
Braithwaite, R.J. and Y. Zhang. 1999. Modeling changes in glacier mass balance that may occur as a result of climate changes. Geografiska Annaler 81 A: 489-496.
Chander, G., B. Markham. 2003. Revised Landsat 5 TM radiometric calibration procedures and post-calibration dynamic ranges. White Paper, USGS Landsat Project Website. Online at http://landsat7.usgs.gov/documents/L5TMCal2003.pdf.
Chavez, P.S. 1988. An improved dark-object subtraction technique for atmospheric scattering correction of multispectral data. Remote Sensing of Environment 24: 459-479.
Chavez, P.S. 1989. Radiometric calibration of Landsat Thematic Mapper multispectral images. Photogrammetric Engineering and Remote Sensing 55(9): 1285-1294.
Chavez, P.S. 1996. Image-based atmospheric corrections – Revisited and improved. Photogrammetric Engineering and Remote Sensing 62(9): 1025-1036.
125
Colby, J.D. 1991. Topographic normalization in rugged terrain. Photogrammetric Engineering and Remote Sensing 57(5): 531-537.
Colby, J.D. and P.L Keating. 1998. Land cover classification using Landsat TM imagery in the tropical highlands: the influence of anisotropic reflectance. International Journal of Remote Sensing 19(8): 1479-1500.
Davis, R.E., A.W. Nolin, R. Jordan, and J. Dozier. 1993. Toward predicting temporal changes of the spectral signature of snow in visible and near-infrared wavelengths. Annals of Glaciology 17: 143-148.
d'Orbigny, A.D. 1835-1847. Voyage dans l'Amérique Méridionale (le Brésil, la République orientale de l'Uruguay, la République Argentine, la Patagonie, la République du Chili, la République de Bolivia, la République du Pérou), executé pendant les années 1826, 1827, 1828, 1829, 1830, 1831, 1832 et 1833 [Voyage to South America (Brazil, Uruguay, Argentina, Patagonia, Chile, Bolivia, Perú), carried out during the years 1826-1833]. Paris, Pitois-Levrault and Company 9v.
Dozier, J. 1989. Spectral signature of alpine snow cover from Landsat Thematic Mapper. Remote Sensing of Environment 28: 9-22.
Dozier, J., and D. Marks. 1987. Snow Mapping and Classification from Landsat Thematic Mapper data. Annals of Glaciology 9: 97-103.
EDC. 2003. Landsat Pathfinder Program. Land Processes Distributed Active Archive Center (LP DAAC). Earth Observatory System Data Center (EDC). Online documents version at http://edcdaac.usgs.gov/pathfinder/pathpage.html
ERDAS. 1997. ERDAS Imagine, ERDAS Field Guide. Leica Geosystems GIS & Mapping, Atlanta, GA.
ESRI. 2003. ArcGIS 8.3 Online Help Documentation. ESRI Redlands, CA.
Fily, M., B. Bourdelles, J.P. Dedieu, and C. Sergent. 1997. Comaprison of In Situ and Landsat Thematic Mapper derived snow grain characteristics in the Alps. Remote Sensing of Environment 59: 452-460:
Finsterwalder, R. 1987. Map of the Cordillera Real Nord (Illampu). München, Deutsche Alpenverein, scale 1:50,000.
Finsterwalder, R. 1990. Map of the Cordillera Real Süd (Illimani). München, Deutsche Alpenverein, scale 1:50,000.
Francou, B. 1993. Une représentation factorielle des cryosphères d'altitude dans le monde in Géomorphologie et aménagement de la montagne--melanages in hommage à P. Gabert [A quantitative description of the high-altitude cold regions in the world in Geomorphology and development of the mountain--a series of
126
papers dedicated to Pierre Gabert]: Centre National de la Recherche Scientifique, Centre de Geomorphologie de Caen, Bulletin 42: 75-86.
Francou, B., M. Vuille, P. Wagnon, J. Mendoza and J. Sicart. 2003. Tropical climate change recorded by a glacier in the central Andes during the last decade of the 20th century: Chacaltaya, Bolivia, 16ºS. Journal of Geophysical Research, 108(0).
Gao, J. and Y. Liu. 2001. Application of remote sensing, GIS and GPS in glaciology: A review. Progress in Physical Geography 25 (4): 520-540.
Haeberli, W. 2003. Systematic long-term mass-balance observations in the Andes as a high-priority need in worldwide climate-related glacier observation programs. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Haeberli, W., R. Frauenfelder, M. Hoelzle, and M. Maisch. (1999). On rates and acceleration trends of global glacier mass changes. Geografiska Annaler 81 A:585-591.
Haeberli, W. and M. Hoelzle. 1994. Application of inventory data for estimating characteristics of and regional climate change effects on mountain glaciers – A pilot study with the European Alps. IGS symposium on the role of the Cryosphere on global change, Columbus Ohio.
Hall, D.K., A.T.C. Chang, J.L. Foster, C.S. Benson, and W.M. Kovalick. 1989. Comparin of In Situ and Landsat Derived Reflectance of Alaskan Glaciers. Remote Sensing of Environemnt 28: 23-31.
Hall, D.K., A.T.C. Chang, and H. Siddalingaiah. 1988. Reflectance of glaciers as calculated using Landsat-5 Thematic Mapper data. Remote Sensing of Environment 25: 311-321.
Hall, D.K., J.L. Foster, D.L. Verbyla, and A.G. Klein. 1998. Assessment of snow-cover mapping accuracy in a variety of vegetation-cover densities in central Alaska. Remote Sensing of Environment 66: 129-137.
Hall, D.K., J.P. Ormsby, R.A. Bindschadler, and H. Siddalingaiah. 1987. Characterization of snow and ice reflectance zones on glaciers using Landsat Thematic Mapper data. Annals of Glaciology, 9: 104-108.
Hall, D.K., G.A. Riggs, and V.V. Salomonson. 1995. Development of methods for mapping global snow cover using moderate resolution imaging spectroradiometer data. Remote Sensing of Environment 54: 127-140.
Herrmann, A. 1993. Review of Jordan, Ekkehard, 1991. The glaciers of the Bolivian Andes, a photogrammetric-cartographical inventory of the Bolivian glaciers as a
127
basis for climatic interpretation and potential economic use. Catena 20(3): 351-353.
Hooker, B.L. and B.B. Fitzharris. 1999. The correlation between climatic parameters and the retreat and advance of Franz Josef Glacier, New Zealand. Global and Planetary Change 22: 39-48.
Huang, C., L. Yang, C. Homer, B. Wylie, J. Vogelmann and T. DeFelice. 2003. At-satellite reflectance: a first order normalization of Landsat 7 ETM+ images. Raytheon ITSS EROS Data Center, Sioux Falls, SD. U.S. Geological Survey online document at http://landcover.usgs.gov/pdf/huang2.pdf .
IPCC, 2001. Climate Change 2001: The Scientific Basis. Published for the Intergovernmental Panel on Climate Change. Cambridge University Press, NY.
Irish, R. 1998. Landsat 7: Science Data Users Handbook. Goddard Space Center. Greenbelt, MD. Online at http://ltpwww.gsfc.nasa.gov/IAS/handbook/handbook_toc.html.
Jordan, E. 1991. Die Gletscher der bolivianischen Anden, eine photogrammetrisch-kartographische Bestandsaufnahme der Gletscher Boliviens als Grundlage für klimatise Deutungen und Potential für die wirtschaftliche Nutzung [The glaciers of the Bolivian Andes, a photogrammetric-cartographical inventory of the Bolivian glaciers as a basis for climatic interpretation and potential for economic use]. Stuttgart, Franz Steiner Verlag, 401p. maps.
Jordan, E. 1999. Glaciers of Bolivia. Satellite Image Atlas of the World. USGS professional paper 1386-I-5. http://pubs.usgs.gov/prof/p1386i/index.html. Accessed on 12/12/03.
Jordan, E., C. Brockman, A. Fernandez, R. Alvarez and K. Jacobsen. 1980, The glacier inventory of Bolivia in world glacier inventory, proceedings of the Riederalp (Switzerland) workshop. International Association of Hydrological Sciences, IAHS-AISH 126: 25-32.
Jordan, E. and R. Finsterwalder. 1992. Observaciones respecto al mapa Cordillera Real Norte (Illampú) 1:50,000--una contribución a la representación cartográfica y a la glaciología e historia de los glaciares de los Andes bolivianos [Explanatory notes to the map Cordillera Real North (Illampu) 1:50,000--A contribution to the cartographic representation, glaciology, and glacial history of the Bolivian Andes]. La Paz, Instituto Geográfico Militar: 178p. [In Spanish, German, English, and French.]
JPL, 2003. SAR Interferometry. Document, Courtesy of the Jet Propulsion Laboratory. http://www-radar.jpl.nasa.gov/sect323/InSar4crust/SarInterferometry.html Accessed on 06/2003.
128
Kaser, G. 1999. A review of modern fluctuation of tropical glaciers. Global and Planetary Change 22: 93-103.
Keller, K., G. Casassa, A. Rivera, R. Forsberg and N. Gundestrup. 2003. Mass balance of glaciers in Patagonia by means of airborne laser altimetry. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Klein, A.G. and B.L. Isacks. 1998. Alpine glacial geomorphological studies in the central Andes using Landsat thematic mapper images. Glacial Geology and Geomorphology, rp01/1998. http://ggg.qub.ac.uk/ggg/papers/full/1998/rp011998/rp01.htm
Klein, A.G. and B.L. Isacks. 1999. Spectral mixture analysis of Landsat thematic mapper images applied to the detection of transient snowline on tropical Andean glaciers. Global and Planetary Change 22: 139-154.
Knap, W.H., C.H. Reijmer. 1998. Anisotropy of the reflected radiation field over melting glacier ice: measurements in Landsat TM bands 2 and 4. Remote Sensing of Environment 65: 93-104.
Lange, H., J. Araos and A. Rivera. 2003. Ice elevation changes and ice velocities of glaciar Juncal Norte, Chile. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Markham, B.L., G. Chander. 2003. Revised Landsat 5 TM Radiometric Calibration Procedures and Post-Calibration Dynamic Ranges. http://landsat7.usgs.gov/documents/L5TMCal2003.pdf, White Paper, USGS Landsat Project Website.
Markham, B.L., J.L Barker. 1986. Landsat MSS and TM post-calibration dynamic ranges, exoatmospheric reflectances and at-satellite temperatures. EOSAT Landsat Data User Notes. Lanham, MD.
Martin, C. 2004. The phase diagram of water. London South-Bank University. Online document at http://www.lsbu.ac.uk/water/phase.html.
Mather, P.M. 1999. Computer processing of remotely-sensed images: An introduction. John Wiley & Sons Inc. New York, NY.
McKnight, T.L. 2000. Physical Geography: a landscape appreciation. Prentice-Hall Inc. Upper Saddle River, NJ.
Mercer, J.H. 1967. Glaciers of Bolivia and of Chile north of latitude 23ºS; in Southern Hemisphere glacier atlas. U.S. Army Natick Laboratories, Earth Sciences Laboratory, Series ES-33, Technical Report 67-76-ES: 65-81.
129
Minnaert, J.L. 1941. The reciprocity principle in lunar photometry. Astrophysics Journal 93:403-410.
Moran, M.S., R.D. Jackson, P.N. Slater and P.M. Teillet. 1992. Evaluation of simplified procedures for retrieval of land surface reflectance factors from satellite sensor output. Remote Sensing of Environment 41:169-184:
Nave, C.R. 2000. Mie Scattering. Georgia State University. Online document at http://hyperphysics.phy-astr.gsu.edu/hbase/atmos/blusky.html.
NASA. 2003. Landsat 7: Documentation. National Aeronautics and Space Administration. Online version at http://landsat.gsfc.nasa.gov/main/documentation.html.
Nolin, A.W., J. Dozier. 2000. A hyperspectral method for remotely sensing the grain size of snow. Remote Sensing of Environment 74: 207-216.
Nolin, A.W., J. Dozier, and L.A.K Mertes. 1993. Mapping alpine snow using a spectral mixture modeling technique. Annals of Glaciology 17: 121-124.
Painter, T.H., J. Dozier, D.A. Roberts, E.D. Robert, and R.O. Green. 2003. Retrieval of subpixel snow-covered area and grain size from imaging spectrometer data. Remote Sensing of Environment 85: 64-77.
Paul, F., A. Kaab, M. Maisch, T. Kellenberger and W. Haeberli. 2002. The new remote-sensing-derived Swiss glacier inventory: I. Methods. Annals of Glaciology 34: 355-361.
Riaño, D., E. Chuvieco, J. Salas, and I. Aguado. 2003. Assessment of different topographic corrections in Landsat TM data for mapping vegetation types. IEEE Transactions on Geoscience and Remote Sensing 41(5):1056-1061
Rignot, E., A. Rivera and G. Casassa. 2003. Volume changes and ice dynamics of the Patagonian icefields from SAR interferometry. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Rivera, A. and G. Casassa. 1999. Volume changes on Pio XI glacier, Patagonia: 1975-1995. Global and Planetary Change 22: 233-244.
Rivera, A., T. Benham, G. Casassa, J. Bamber and J. Dowdeswell. 2003. Glaciological application of ASTER DEM from Campo de Hielo Norte, Patagonia, Chile. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Rosenthal, W, J. Dozier. 1996. Automated mapping of montane snow cover at subpixel resolution from Landsat Thematic Mapper. Water Resources Research, 32(1): 115-130.
130
RSI. 2001. ENVI User’s Guide. Research System Inc. Boulder, CO.
Schnirch, M., C. Schneider, G. Casassa and R. Kilian. 2003. Glacier inventory of Peninsula Munoz Gamero, Patagonia, and glacier change during recent decades. Symposium on mass balance of Andean glaciers and 1st mass balance workshop on Andean glaciers. Centro de Estudios Cientificos, Valdivia (Chile).
Schott, J. 1997. Remote Sensing: The image chain approach. Oxford University Press. New York.
Skvarca, P., M. Stuefer and H. Rott. 1999. Temporal changes of Glacier Mayo and Laguna Escondida, southern Patagonia, detected by remote sensing data. Global and Planetary Change 22: 245-253.
Smith, J., T. Lin, and K. Ranson. 1980. The Lambertian assumption and Landsat data. Photogrammetric Engineering and Remote Sensing 46:1183-1189.
Teillet, P.M., B. Guidon, and D.G. Goodenough. 1982. On the slope-aspect correction of multispectral data. Canadian Journal of Remote Sensing 51:229-235.
Teillet, P.M., J.L. Barker, B.L. Markham, R.R. Irish, G. Fedosejevs, and J.C. Storey. 2001. Radiometric cross-calibration of Landsat-7 ETM+ and Landsat-5 TM sensors based on tandem data sets. Remote Sensing of Environment 78: 39-54.
Thompson, L.G. 2000. Ice core evidence for climate change in the Tropics: Implications for our future. Quaternary Science Review 19: 19-35.
Thompson, L.G. 2003. Personal interview at Appalachian State University, NC.
Thompson, L.G. and M.E. Davis. 1998. A 25,000 year tropical climate history from Bolivian ice cores. Science 282 (5395): 1858-1865.
Troll, C. and R. Finsterwalder. 1935. Die Karten der Cordillera Real und des Talkessels von La Paz and die Diluvialgeschichte der zentralen Anden [The maps of the Cordillera Real and the valley of La Paz and the Pleistocene history of the Central Andes]. Petermanns Geographische Mitteilungen 81(11): 393-399; 81(12): 445-455.
USGS. 1999. Satellite Image Atlas of Glaciers of the World. U.S. Geological Survey Professional Paper 1386-I. Online at http://pubs.usgs.gov/prof/p1386i/
USGS. 2003. Landsat-7 Level-0 and Level-1 Data Sets Document. U.S. Geological Survey. Online at http://eosims.cr.usgs.gov:5725/DATASET_DOCS/landsat7_dataset.html.
Warren, S. G. 1982. Optical properties of snow. Review of Geophysics and Space Physics, 20: 67-89.
131
Xiao, X., Z. Shen and Z. Qin. 2001. Assessing the potential of Vegetation sensor data for mapping snow and ice cover: a Normalized Difference Snow and Ice Index. International Journal of Remote Sensing 22(13):2479-2487.
Zheng, Q., C.M. Cao, X. Feng, F. Liang, X. Chen, W. Sheng. 1984. Study on spectral reflection characteristics of snow, ice and water of northwest China. Scienctia Sinica 27:647-656.
References Consulted
Bishop, M.P., J.F. Shroder, and J.D. Colby. 2003. Remote sensing and geomorphometry for studying relief production in high mountains. Geomorphology 55: 345-361.
Demuth, M. and A. Pietroniro. 1999. Inferring glacier mass balance using Radarsat: Results from Peyto Glacier, Canada. Geografiska Annaler 81 A: 521-540.
Hall, D.K., K.J. Bayr, W. Schoner, R.A. Bindschadler, and J.Y.L. Chien. 2003. Consideration of the error inherent in mapping historical glacier positions in Austria from the ground and space (1893-2001). Remote Sensing of Environment 86: 566-577.
Kaab, A., F. Paul, M. Maisch, T. Kellenberger and W. Haeberli. 2002. The new remote-sensing-derived Swiss glacier inventory: II. First Results. Annals of Glaciology 34: 362-366.
Li, Z, W. Sun and Q. Zeng. 1998. Measurements of glacier variation in Tibetan Plateau using Landsat data. Remote Sensing of Environment 63:258-264.
McGuffie, K. and A. Henderson-Sellers. 1997. A climate modeling primer. John Wiley & Sons, New York.
Meier, M.F. 1973. Evaluation of ERTS imagery for mapping and detection of changes in snow cover on land and on glaciers. Symposium on significant results obtained from earth resources technology satellite-1, NASA SP-327: 863-875.
Ribstein, P., E. Tirau, B. Francou and R. Saravia. 1993. Tropical climate and glacier hydrology: A case study in Bolivia. Journal of Hydrology 165: 221-234.
Thompson, L.G. 1997. Tropics on Ice. Earth 6(5):38-46.
Winther, J.G., D.K. Hall. 1999. Satellite-derived snow coverage related to hydropower production in Norway: present to future.
132
Appendix A.
Anisotropic Reflectance Correction and Topographic Normalization of
rugged terrains and snow covered surfaces.
Concepts of Anisotropic Reflectance and Topographic normalization
One of the most complex questions remote sensing research needs to address is
the question of how the energy leaving a surface is angularly distributed along the
hemisphere centered on the point of reflection (Schott 1997). Depending on the physical
properties of the illuminated surface, the incoming energy could be perfectly reflected in
one direction (specularly, such as in a mirror), perfectly diffused in all directions, or as in
nature, vary between nearly diffuse to nearly specular.
An ideal flat surface that diffuses electromagnetic radiation perfectly in all
directions is defined as Lambertian. As a result, the radiation measured from any
observation angle is the same, and the surface is said to present isotropic reflectance. In
this situation, the amount of the reflected energy is only a function of the angle of
incidence and of the magnitude of the incoming radiation. In remote sensing research,
surfaces are often assumed to be Lambertian in order to simplify models and reduce
computation time. A flat surface that presents nearly diffused to nearly specular reflection
will reflect energy unevenly. Measurements of reflected radiation will vary with the
observation angle by a complex function. In this case, the surface is defined non-
Lambertian and it presents anisotropic reflectance. The reflected energy measured by an
133
observer (such as a satellite sensor) will depend on the geometry of the source (incidence
angle, azimuth and magnitude), and by the geometry of the observer (exitance angle and
azimuth) (Colby and Keating 1998; Riaño et al. 2003). The bidirectional reflectance
distribution function (BRDF), in conjunction with anisotropic reflection models, is used
to describe bidirectional reflectance values for all combinations of input-output angles
(Schott 1997).
Most of the targets and landcovers on the surface of the Earth present a certain
degree of anisotropic reflectance, which could more easily be corrected if all the surfaces
were ideally flat, however, the irregular topography of the Earth’s surface introduces a
further complication. In fact, even if source and observer maintain a constant geometry,
the orientation of the targets may vary depending on local slope and aspect. As a result,
the amount of solar energy reflected toward the sensor will vary as a function of (1)
source and observer geometry, and (2) slope and aspect of the local terrain (Colby 1991;
Colby and Keating 1998). This phenomenon, known as topographic effect, make surfaces
appear brighter when exposed to direct illumination and darker otherwise. This
differential illumination is perceived visually as relief in photographic pictures and
optical satellite images (Colby and Keating 1998). Topographic normalization techniques
aim to eliminate, or reduce, the topographic effect to obtain surface brightness values as
if the surface was flat.
Topographic Normalization: Methods and Techniques
Three techniques available to perform topographic normalization are band
ratioing, Lambertian models, and non-Lambertian models. While band ratioing relies on
134
in-scene information, Lambertian and non-Lambertian methods require additional data in
the form of a digital elevation model (DEM) to extract slope and aspect of the local relief.
Also, local knowledge about the landcover may improve the efficiency and accuracy of
non-Lambertian models.
Topographic normalization techniques based on band ratioing assume that the
relative reflectance of spectral bands for a specific landcover is independent of
illumination, thus the ratio of two spectral bands will be a constant not affected by the
topographic effect (Schott 1997; Mather 1999). Band ratioing has been considered to be a
suitable methodology to normalize the terrain effect because it presents low
computational cost and does not require additional data; nevertheless, researchers argued
that variations of the reflectance of a surface may be misinterpreted since different
surfaces may present similar reflective properties, thus leading to misclassification
(Colby 1991). Also, it is argued that the assumption of constant proportionality between
spectral bands is not supported by physical models. In fact, in case of diffuse radiance,
the pattern of the reflected energy depends on wavelength and incident angle (Riaño et al.
2003).
The Lambertian model implemented to perform topographic normalization
assumes that the surface investigated diffuse the reflected energy uniformly across an
ideal hemisphere centered on the reflection point. Also, Lambertian models state that the
intensity of the energy Ln measured from any observation angle is proportional to the
incoming incident radiation L, and inversely proportional to the cosine of the incident
angle (equation A.1) (Schott 1997; Colby and Keating 1998; Mather 1999; Riaño et al.
2003).
135
iLLn cos
= (equation A.1)
Three main problems arise when considering natural surfaces to be Lambertian.
First, it can be argued that while there are natural landcovers that present Lambertian
reflection, such as fresh snow (Knap and Reijmer 1998), most of the Earth’s landcovers
are characterized by anisotropic reflectance, such as vegetated areas. Second, Earth’s
surface can be characterized by complex topography, thus the incident angle cannot be
simply represented by the source geometry (sun elevation), but must be corrected for the
observer position considering slope and aspect at the incident point. The most accurate
Lambertian topographic normalization methodologies take into account both source and
observer geometries. Nevertheless, a third flaws affects the Lambertian models. Being
inversely proportional to the cosine of the incident angle, the Lambertian model over
estimates radiance calculated for incident angles proximate to vertical (reflection plane
close to perpendicularity). This phenomenon is due to the fact that the cosine approaches
zero when the incident angle is close to vertical, thus amplifying the estimated reflectance
radiance toward infinity. This effect is particularly visible for sun-facing slopes, or for
flat surfaces recorded by nadir sensors and high solar elevation (Schott 1997; Colby and
Keating 1998).
Non-Lambertian models are based on semi-empirical models, and include the
Minnaert correction (Minnaert 1941), the Teillet correction (Teillet et al. 1982), and the C
correction (Teillet et al. 1982). Riaño et al. (2003) evaluated the application and response
of different non-Lambertian methods. Colby (1991) assessed the efficiency of the
Minnaert correction for topographic normalization of rugged terrain. Later, Colby and
136
Keating (1998) and Bishop and Colby (2002) further discussed and successfully applied
the Minnaert correction to normalize topography in tropical montane regions, and in the
Himalayas. The Minnaert correction can be derived using the backward radiance
correction transformation (BRCT) (Smith et al. 1980), and is based on the estimation of
semi-empirical constants, called Minnaert coefficients, that are dependent on the spectral
bands (wavelength) and on the physical properties of the landcover (Colby 1991). For a
fast, less accurate result, the Minnaert coefficients necessary to normalize each spectral
band could be estimated on a scene-wide scale. A more accurate approach, which takes
into account local knowledge about the physical properties of the landcovers investigated,
was used to estimate Minnaert coefficients that correct spectral bands on a landcover
basis, thus increasing considerably the accuracy and performance of the methods.
Case Study: Image Processing
The Minnaert correction was found to be particularly
suitable to normalize Landsat imagery of rugged terrains in
tropical environments (Colby and Keating 1998). The Minnaert
correction was chosen to perform the topographic normalization
of two subset areas of a 1987 Landsat TM scene of the Cordillera
Real. The first subset (about 59 km2) represents a portion of the
Ancohuma-Illampu massif and contains mainly snow and ice
covers (figure A.1), while the second subset (about 43 km2)
Figure A.1. Snow and ice cover sample of approximately 59 km2.
Snow and Ice
137
represented a portion of rugged terrains
characterized by barren soils and low sparse
vegetation (figure A.2). A digital elevation model
(DEM) of the year 2003 was obtained from the
available free dataset of the satellite ASTER.
The Minnaert correction theory assumes
that the pattern of scattered radiation can be
described by the bidirectional reflection
distribution function (BRDF) which requires the calculation of incidence and exitance
angles. The Minnaert constants k can be used as an approximate representation of the
BRDF.
Equation A.2 represents a non-Lambertian reflection model based on Minnaert
correction; L is the incoming radiance, i is the angle of incidence, e is the angle of
exitance, and k is a Minnaert constant.
eieLL kkn coscos
cos= (equation A.2)
For each spectral band n, a Minnaert constant kn can be estimated by regression of pre-
processed layers obtained from the DEM and the Landsat scene (Colby 1991). To begin,
it is necessary to rewrite Equation A.2 as follow:
eiLeL kkn coscoscos = . (equation A.3)
Then a mathematical transformation enable k to be represented as a linear coefficient:
eikLeL n coscosloglogcoslog += . (equation A.4)
Figure A.2. Soils and vegetation sample of approximately 43 km2.
Soils and low vegetation
138
Equation A.4 is in the form of a linear equation y=kx+b where y=logLcose,
x=logcosicose, and b=logLn. The x term (independent variable) and y term (dependent
variable) can be used in a regression model to estimate the k constant for each spectral
band.
The available DEM was rectified, co-registered, and re-sampled to 28.5 m to
match the spatial characteristics of the Landsat scene. The DEM was then processed to
obtain slope and aspect layers. The calculation of slope and aspect represent a necessary
step to compute the correct angle of incidence i and angle of exitance e for each pixel of
the scene. In fact, the angles of incidence and exitance depend on the geometry of the
source (solar illumination), the position of the observer (nadir satellite sensor in this
case), and the local topography. The next step involved the calculation of the cosine of
the angle of incident (cosi) and of the angle of exitance (cose). The product of the cosines
(cosicose) and the logarithm of the product (logcosicose) were also calculated.
The Landsat scene had been previously pre-processed to obtain radiance layers L
and calculate the logarithm of the product between the cosine of the angle of exitance and
the radiance for each spectral band (logLcose).
Finally, the Minnaert coefficients kn (table A.1 and table A.2 ) were obtained by
regression using the member logcosicose as the independent variable (x term), and the
member logLcose as the dependent variable (y term). This procedure was performed on
both subsets, and for all the available spectral bands (visible, near infrared, and mid-
infrared).
139
k 1 0.670794658955001 k 1 0.725530599314385k 2 0.841102853169116 k 2 0.932826820287288k 3 0.800794183002669 k 3 0.897134115851479k 4 0.766390160049200 k 4 0.963814659244721k 5 0.811673282184080 k 5 0.704773225411237k 7 0.847638785533782 k 7 0.355635525943596
Table A.1. Rugged terrain subset Table A.2. Ancohuma-Illampu glaciers
Soils and low vegetationMinnaert coefficientsSnow and ice cover
Minnaert coeffcients
Case Study: Results and Discussion
Figures A.3 and A.4 present the subsets before and after topographic
normalization. A first order analysis of the results based on visual inspection reveals that
the soil and low vegetation normalized sample is characterized by a general optical
flattening of the topography. This is in agreement with the concept of topographic
normalization which aims to reduce topography to a flat horizontal surface, and
homogenize brightness variations induced by slope and aspect.
Figure A.3. Soils and vegetation sample before and after topographic normalization
Before After
140
The visual difference between corrected and uncorrected snow and ice cover
samples, instead, appear to be less pronounced. This is mostly attributed to the saturation
of the Landsat sensors in the visible and near-infrared bands which induces brightness
cut-off, and thus the optical flattening of the topography in the original sample. Also,
fresh and non-diagenized snow present a strong Lambertian reflection (Knap and Reijmer
1998), which partly masks the terrain effect in optical imagery.
To achieve a better understanding of the results, it is necessary to proceed with a
second order analysis to compare the data characteristics before and after the
normalization process. The hypothesis is that the original subset should present some
correlation between the brightness values and the incidence angle, in particular, the
reflectance registered by the sensor should be a function of the incidence angle. On the
other hand, in the normalized subsets, the brightness value should be largely uncorrelated
to the incident angle, since the normalization process should remove variation of
reflectance due to topography. To validate this test, it is possible to perform a regression
between the cosine of the incident angle (cosi) (independent variable) and the reflectance
Figure A.4. Snow and ice cover sample before and after topographic normalization.
Before After
141
(dependent variable) before and after normalization. The regression analysis was
achieved with IDIRISI Kilimanjaro (Clark Labs at Clark University). and the results for
the soil and low vegetation covers (band 3) are presented in figure A.5 and figure A.6.
It is possible to observe that, in fact, the coefficient of correlation between the
cosine of the incident angle (cosi) and the reflectance drops from 34.12% (figure A.5) to
6.19% after the topographic normalization (figure A.6). This change is significant and
indicates that the topographic normalization has reduced the geometric correlation
between incident angle and brightness values.
In the case of the snow and ice cover subset (figure A.7 and figure A.8 showing
band TM 3), the results needs to be interpreted in lieu of the characteristics of the
landcover. It can be observed that the normalized dataset presents a low degree of
correlation (4.33%), but it could be argued that the natural Lambertian properties of the
snow already reduces the correlation between incidence angle and reflectance (19.90% in
the source subset), and also, that due to the effect of sensor saturation over snow and ice
covers, the radiance values returned by the normalization process are based on cut-off
brightness values and thus, the normalized dataset do not represent the real physical
properties of snow and ice.
142
Figure A.5. Regression analysis for the soil and low vegetation subset (band TM 3) before normalization.
Figure A.6. Regression analysis for the soil and low vegetation subset (band TM 3) after normalization.
143
A further example (figure A.9) showing the regression of the cosine of the
incident angle and reflectance of TM band 5 seems to support the near Lambertian
reflection of snow covers. The TM Band 5 does not saturate over snow fields, thus the
reflectance values are in good agreement with the optical properties of snow. It can be
Figure A.8. Regression analysis for the snow and ice cover subset (band TM 3) after normalization.
Figure A.7. Regression analysis for the snow and ice cover subset (band TM 3) before normalization
144
observed that for TM band 5 the coefficient of correlation in the source subset is only
9.56%, showing a very low correlation between angle of incidence and reflectance.
After normalization, the coefficient of correlation drops to 4.11%, a value not
significantly different from the uncorrected dataset.
Limitations and Final Consideration
Topographic normalization and anisotropic reflectance correction techniques are
of utmost importance to study rugged terrain and alpine environments. The case study
highlighted that non-Lambertian normalization based on Minnaert coefficient is effective
in reducing the terrain effect over soils and vegetation in Landsat imagery, but did not
significantly improve the physical information over snow and ice covers. This was
mainly due to (1) the Lambertian properties of fresh snow which naturally reduce the
optical terrain effect, and (2) the loss of information in the visible and near infrared
Figure A.9. Regression analysis for the snow and ice cover subset (band TM 5) before normalization.
145
Landsat bands caused by sensor saturation. Also, studies of the optical properties of snow
and ice have shown that aging snow cover presents a strong forward scattering that can
be efficiently corrected with empirical expressions (or C corrections) (Knap and Reijmer
1998). Rosenthal and Dozier (1996) argued that the calculation of slope and aspect
amplifies the digital noise present in digital elevation models and introduces considerable
error in the classification of snow facies, instead they suggested the use of the
information obtained from elevation models to specify boundary conditions for radiative
transfer calculation by defining facies ranges from bi-spectral plots.
An important limitation of the non-Lambertian methodology resides in the use of
reliable and temporally consistent DEM datasets which are not always available for
remote areas such as in the case of the Cordillera Real. Future airborne and satellite
missions, and the release of classified information such as the Shuttle Radar Topography
Mission (SRTM) dataset, may improve the quality, availability, and worldwide coverage
of high resolution DEM to be implemented for topographic normalization.
146
Biographical Information
Christian Degrassi was born in Grado, Italy, on December 7th, 1969. He was
awarded the Bachelor of Science Degree in Geology from the University of Trieste, Italy
in 1994. In 1996 he was hired by Schlumberger Oilfield Services, Texas, where he had
the opportunity to travel and work worldwide in a high technology oriented environment.
In 2001 he began study toward a Master of Arts in Geography. The M.A. was awarded in
2004.