probabilistic large-area mapping of seagrass species distributions

23
Copyright # 2006 John Wiley & Sons, Ltd. Received 1 August 2005 Accepted 27 December 2005 AQUATIC CONSERVATION: MARINE AND FRESHWATER ECOSYSTEMS Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007) Published online 14 July 2006 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/aqc.772 Probabilistic large-area mapping of seagrass species distributions K.W. HOLMES 1,2,3, *, K.P. VAN NIEL 1,2 , G.A. KENDRICK 1,3 and B. RADFORD 2 1 Cooperative Research Centre for Coastal Zone, Estuary and Waterways Management, Indooroopilly, Queensland 4068, Australia 2 School of Earth and Geographical Sciences, The University of Western Australia, Crawley, WA 6009, Australia 3 School of Plant Biology, The University of Western Australia, Crawley, WA 6009, Australia ABSTRACT 1. Aerial photograph classification was used to map perennial thick canopy seagrass presence/ absence over a large area (85 km 2 ) off the coast of Western Australia. Within those areas mapped as seagrass, a geostatistical nonparametric interpolation method was applied to map the probability of seagrass species presence from underwater tow video. Multiple species mixtures were mapped at fixed probability thresholds of 0.95, 0.75, 0.50, and 0.25. Taxa included Amphibolis spp., Posidonia coriacea, P. sinuosa, P. australis and ephemeral species (Halophila and Zostera tasmanica (newly named as Heterozostera polychlamys)). 2. The most commonly occurring species were respectively Amphibolis spp., Posidonia coriacea, P. sinuosa, P. australis, and the ephemeral species. Amphibolis, P. coriacea, and the ephemeral species were mapped predominantly as mixed assemblages (71–89% mixed), whereas P. sinuosa and P. australis were typically mapped as single species. 3. Different species growth habits led to distinctive differences in large area distributions. All species were highly variable over short distances (5500 m), and spatial dependence persisted over more than 5 km. However, Posidonia sinuosa meadows were oriented with the longest axis running north–south, and a shorter axis running east–west perpendicular to the coastline (spatial dependence to 2.8 km and 0.8 km, respectively). The ephemeral species were less successfully mapped, largely owing to the potentially different growth patterns of the grouped species, and because their full extent could not be captured by the aerial photograph classification. 4. The individual biology of each species results in unique landscape features where Posidonia sinuosa forms larger continuous and predominantly monospecific meadows, whereas the more common Amphibolis and P. coriacea form multi-species patchy meadows. These mapped features suggest that the emergence of species patterns in seagrass landscapes is influenced by differences in clonal growth among seagrass species. 5. Probabilistic species mapping provided information unavailable from discretely classified maps, and facilitates targeted sampling for improving map accuracy, and for more realistically evaluating *Correspondence to: K.W. Holmes, School of Earth and Geographical Sciences, The University of Western Australia, 35 Stirling Highway, Crawley, WA 6009, Australia. E-mail: [email protected]

Upload: kw-holmes

Post on 06-Jun-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Probabilistic large-area mapping of seagrass species distributions

Copyright # 2006 John Wiley & Sons, Ltd. Received 1 August 2005Accepted 27 December 2005

AQUATIC CONSERVATION: MARINE AND FRESHWATER ECOSYSTEMS

Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Published online 14 July 2006 in Wiley InterScience(www.interscience.wiley.com). DOI: 10.1002/aqc.772

Probabilistic large-area mapping of seagrassspecies distributions

K.W. HOLMES1,2,3,*, K.P. VAN NIEL1,2, G.A. KENDRICK1,3 and B. RADFORD2

1Cooperative Research Centre for Coastal Zone, Estuary and Waterways Management, Indooroopilly,

Queensland 4068, Australia2School of Earth and Geographical Sciences, The University of Western Australia, Crawley, WA 6009, Australia

3School of Plant Biology, The University of Western Australia, Crawley, WA 6009, Australia

ABSTRACT

1. Aerial photograph classification was used to map perennial thick canopy seagrass presence/absence over a large area (85 km2) off the coast of Western Australia. Within those areas mapped asseagrass, a geostatistical nonparametric interpolation method was applied to map the probability ofseagrass species presence from underwater tow video. Multiple species mixtures were mapped atfixed probability thresholds of 0.95, 0.75, 0.50, and 0.25. Taxa included Amphibolis spp., Posidoniacoriacea, P. sinuosa, P. australis and ephemeral species (Halophila and Zostera tasmanica (newlynamed as Heterozostera polychlamys)).2. The most commonly occurring species were respectively Amphibolis spp., Posidonia coriacea,

P. sinuosa, P. australis, and the ephemeral species. Amphibolis, P. coriacea, and the ephemeral specieswere mapped predominantly as mixed assemblages (71–89% mixed), whereas P. sinuosa andP. australis were typically mapped as single species.3. Different species growth habits led to distinctive differences in large area distributions. All

species were highly variable over short distances (5500m), and spatial dependence persisted overmore than 5 km. However, Posidonia sinuosa meadows were oriented with the longest axis runningnorth–south, and a shorter axis running east–west perpendicular to the coastline (spatial dependenceto 2.8 km and 0.8 km, respectively). The ephemeral species were less successfully mapped, largelyowing to the potentially different growth patterns of the grouped species, and because their fullextent could not be captured by the aerial photograph classification.4. The individual biology of each species results in unique landscape features where Posidonia

sinuosa forms larger continuous and predominantly monospecific meadows, whereas the morecommon Amphibolis and P. coriacea form multi-species patchy meadows. These mapped featuressuggest that the emergence of species patterns in seagrass landscapes is influenced by differences inclonal growth among seagrass species.5. Probabilistic species mapping provided information unavailable from discretely classified maps,

and facilitates targeted sampling for improving map accuracy, and for more realistically evaluating

*Correspondence to: K.W. Holmes, School of Earth and Geographical Sciences, The University of Western Australia, 35 StirlingHighway, Crawley, WA 6009, Australia. E-mail: [email protected]

Page 2: Probabilistic large-area mapping of seagrass species distributions

species and mixed species distribution predictions. The kriging approach, although not well suitedfor all types of vegetation data, performed well for clonal seagrasses.Copyright # 2006 John Wiley & Sons, Ltd.

KEY WORDS: seagrass; probabilistic mapping; geostatistics; landscape scale; Western Australia; spatial pattern;

tow video; aerial photography

INTRODUCTION

Seagrass meadows play a critical role in maintaining marine biological productivity and biogeochemicalcycles (Hemminga and Duarte, 2000), are recognized for supporting high species diversity and commercialfisheries (Carruthers et al., 2002), and are important indicators of disturbance in coastal marine systems(Cambridge et al., 1986; Short and Wyllie-Echeverria, 1996; Kendrick et al., 1999, 2000, 2002; Duarte,2002; Fourqurean et al., 2003; Green and Short, 2003). The relatively few species of seagrasses in mostmarine ecosystems makes them prime candidates for testing ecological theories involving spatial andtemporal variability of plants and biogeochemical cycling, scaling issues, competition, environmentalcontrols on vegetation distribution, and organism–habitat interactions. Seagrass research has increaseddramatically over the last 20 years, but like many other life science disciplines it has historically focused onphysiological description and has only recently been evolving toward more quantitative, process-orientedresearch (Duarte, 1999). Large gaps remain in our knowledge of seagrass ecosystems which limit our abilityto predict future trends for sustainable management and conservation (Duarte, 1999). A better grasp onpatterns of regional (tens to hundreds of square kilometres) seagrass distribution is needed to both defineand address research and management questions concerning regional relationships among seagrasses,predictive relationships among seagrass species and environmental variables, interactions betweenseagrasses and seafloor habitats, and future scenarios of land-use change (Short and Neckles, 1999;Green and Short, 2003). Quantifying seagrass species distribution is a first step toward understanding theprocesses potentially driving these patterns, enabling prediction of vegetation changes and distributions inunder-sampled areas (Levin, 1992).

In areas with high water clarity, remote sensing approaches have commonly been used to map seagrassextent, including aerial photography (Pasqualini et al., 1998; Kendrick et al., 2000, 2002; Cuevas-Jimenezet al., 2002; Frederiksen et al., 2004), other airborne imagery such as CASI, MEIS-II and Ocean PHILLS(Mumby and Edwards, 2002; Dierssen et al., 2003; Vis et al., 2003) and satellite imagery such as LandsatTM, MSS, SPOT, IKONOS and others (Ferguson and Korfmacher, 1997; Alberotanza et al., 1999;Mumby and Edwards, 2002; Andrefouet et al., 2004). A variety of image classification techniques have beenapplied, ranging from manual digitization (e.g. Kendrick et al., 2000) through supervised (e.g. Pasqualiniet al., 1998) and unsupervised methods (e.g. Cuevas-Jimenez et al., 2002) to radiative transfer modelling(e.g. Dierssen et al., 2003).

These methods have proved highly effective for mapping dense canopy seagrass presence or absence inparticular cases, and have facilitated the delineation of seagrass distributions over hundreds of squarekilometres. However, they are not useful for distinguishing different species nor between seagrass andmacroalgae. For species mapping, a compromise must be reached between gathering sufficient detail in thefield to correctly identify species occurrence and georeference the location, and collecting an adequatenumber of well-distributed samples over the mapping area to permit the use of robust mapping techniques.As illustrated by Fonseca et al. (2002), a small number of detailed study sites taken over a large area, as iscommon in natural resource surveys, can result in highly misleading estimates of seagrass cover,particularly in heterogeneous landscapes. Seagrass species are identified by direct observation in shallowareas accessed by walking (e.g. Robbins and Bell, 2000) or using scuba (in situ identification or destructivesampling) (e.g. Pasqualini et al., 1998; Kirkman and Kirkman, 2000; Fourqurean et al., 2001), or

K.W. HOLMES ET AL.386

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 3: Probabilistic large-area mapping of seagrass species distributions

alternatively using non-destructive remote sensing technology such as underwater tow video (e.g. Robbinsand Bell, 2000; Brown et al., 2002; Kendrick et al., 2002), or remotely operated vessels (ROV) (e.g. McReaet al., 1999; Cochrane and Lafferty, 2002). Destructive sampling is the only way to get positiveidentification for all species, but video and ROV sampling have the advantage of providing contiguousobservations over larger areas. However, the unsystematic distribution of remote observations (densesample coverage along transect, no information between transects) can lead to large mapping errors.

As researchers pose more quantitative and process-oriented questions concerning seagrass ecology(Duarte, 1999), alternative mapping methods are needed to take advantage of the improving quality andquantity of submerged vegetation observations to more accurately depict reality, as well as to supply mapusers with estimates of mapping accuracy. Few options for quantitative mapping of sparse seagrass speciesobservations exist, particularly given the lack of well-understood environmental predictors for theoccurrence of seagrass species. In this study, we map the distribution of seagrass species over a largenearshore region (85 km2) in Western Australia where quantitative species maps were needed to assessimpacts on the ecosystem from commercial development. A combination of image-processing techniqueswere applied to map the extent of perennial seagrass, then, within vegetated areas, geostatisticalinterpolation methods were used to determine the presence of individual seagrass species as a probabilisticcontinuum.

BACKGROUND

Environmental setting

In the 1970s, diebacks of seagrass occurred in Cockburn Sound, just south of the present study in OwenAnchorage near Perth, Western Australia (Figure 1) (Kendrick et al., 2002). The cause was eventuallyattributed to industries along the shoreline, from which nutrient loading led to algal blooms that blockedlight from reaching the seafloor and thus affected underwater plant survival (Cambridge et al., 1986). Sincethat time, the extent of seagrass coverage and seagrass species composition have been monitored for earlywarning signs of the adverse effects of human activities in this economically important area. Shell-sanddredging is another local activity with a direct impact on seagrass distribution, physically removing someplants, but more importantly increasing water turbidity and sedimentation which may negatively impactlarger areas (Kendrick et al., 2000). Knowledge of spatial patterns and rates of change of seagrass loss andrevegetation is critical for environmental planning, recreational and commerce management.

Seagrasses occupy large areas of shallow (510m water depth), subtidal sand banks in Owen Anchorage.Success Bank to the north and Parmelia Bank to the south consist of unconsolidated carbonate sand, theupper surfaces of which range from 3 to 10m water depth. The Banks occupy approximately 10 000 ha, andare separated by a relatively deep lagoon (12–18m water depth). Common seagrass taxa include the generaPosidonia, Amphibolis, Zostera and Halophila. Amphibolis griffithii, A. antarctica, Posidonia australis,P. sinuosa and P. coriacea are perennial, canopy-forming species that dominate shallow areas. Posidoniasinuosa also occurs on the deeper flanks of Success and Parmelia Banks (8–15m depths) whereas Halophilaovalis and Zostera tasmanica (newly named asHeterozostera polychlamys in Kuo (2005)) are confined to theunderstorey, as well as occurring as ephemeral meadows in high-energy environments and deeper water(Kendrick et al., 2000). Aerial photography classification effectively separates the perennial, dense-canopyseagrass species (Amphibolis and Posidonia) from unvegetated sands (Kendrick et al., 2000, 2002). Theephemeral species (Halophila and Zostera) are difficult to delineate remotely. Identification to the specieslevel is only possible using in situ observation.

SEAGRASS DISTRIBUTION MAPPING 387

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 4: Probabilistic large-area mapping of seagrass species distributions

Application of geostatistical approaches for species mapping

Maps of vegetation typically depict non-overlapping polygons of vegetation classes (Goodchild, 1994), butthe presence of species or assemblages can also be viewed as continuous distributions or fields, mapped witha given probability of occurrence (Castrignano et al., 2000; Miller and Franklin, 2002). Assumptions aboutthe nature of the processes controlling a variable’s distribution are inherent in different mapping methods.It can be problematic to use some of the more common geostatistical methods (e.g. ordinary or simplekriging) which assume positive spatial autocorrelation between nearby observations to interpolatevegetation cover in unsampled areas because the ways in which plants propagate do not necessarily producepatterns easily represented by smooth surfaces. For example, in a tropical rainforest, seeds may be widelydistributed by birds and animals, leading to seemingly random species distributions over large areas whichare difficult to predict from the locations of the parent plants. Seagrasses, on the other hand, arepredominantly clonal, meaning the majority of their growth is from rhizome elongation (Kendrick et al.,2005; Sintes et al., 2005). They also release seeds that are widely dispersed by currents, but these tend to

Figure 1. Location map for the study area in Owen Anchorage, Western Australia.

K.W. HOLMES ET AL.388

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 5: Probabilistic large-area mapping of seagrass species distributions

take root in protected areas such as within established meadows, and have a low likelihood of survivalelsewhere (Hemminga and Duarte, 2000; but see Orth et al., in press).

Geostatistical methods have been applied to seagrass mapping, including analysis of spatial dependence(Fonseca, 1996), mapping seagrass density (Fourqurean et al., 2001; Durako et al., 2002), and mappingfertile seagrass shoot density (Campey et al., 2002). However, like all statistical methods, geostatistics arebuilt on a number of assumptions, including, for the most common interpolation techniques (ordinary andsimple kriging), a Gaussian data distribution. For non-Gaussian distributions (such as binominal ornominal data, as are often available for vegetation), indicator kriging was developed. Indicator krigingproduces spatially explicit estimates of the probability that a threshold is exceeded, similar to output fromlogistic regression (Goovaerts, 1997; Burrough and McDonnell, 1998). Logistic regression has been testedfor mapping the probability of seagrass presence (Kelly et al., 2001), but not individual species. Oneadvantage of all kriging algorithms is that estimates of modelling error are produced that provide a usefulspatially explicit map of the error due to sample distribution.

In this paper we map species distributions of seagrasses in Owen Anchorage, Western Australia, usingindicator block kriging. Perennial seagrass extent is first mapped from aerial photography using acombination of image classification techniques. Within those areas mapped as seagrass, probability maps ofthe occurrence of five taxa are constructed for a range of probability thresholds (25%, 50%, 75% and 95%)from seagrass observations obtained from sparse underwater video footage.

METHODS

Data collection

Aerial photographs of Owen Anchorage were captured on March 12, 2004, at a scale of 1:25 000. Anorthorectified mosaic was produced with spatial accuracy of approximately 0.4m horizontally, and 0.7mvertically, with 67% confidence. Original image pixel size was 0.25m, but was resampled to 2m to speedimage-processing.

Identification of seagrass species in situ was accomplished using video footage from an underwatercamera towed behind a boat travelling at 1.0–3.0 knots. Scuba diving was used at eight sites to collectspecimens and resolve ambiguities in the video interpretation. The video was viewed live on the boat, andwas recorded on a Sony miniDV recorder. The boat position, heading and speed were logged from adifferential GPS along with water depth from the vessel’s echo sounder. All of this information was linkedto the video frames at 5-second intervals along each transect. The camera position relative to the seabedwas controlled manually to maintain an optimal distance from the seabed for identifying species (51m).The location of the camera behind the boat’s GPS receiver was empirically modelled as a function of waterdepth and boat speed. The final model was simply ‘1.2�water depth’, and accounted for 86% of thevariance measured over an object with a known location.

The number of transects, transect orientation, and the video frames chosen for recording seagrass speciesinformation were determined by analysing seagrass video data collected in Owen Anchorage in 1999(Kendrick et al., 2002) and adjusted as new information was gathered during the 2004 field campaign.Areas known to have high variability in terms of species present or localized patterns of distribution weretargeted for higher density sampling. The video footage was collected along transects, which meant that alarge number of samples (interpreted video frames) were aligned in one direction (along transect), but therewere large gaps between transect lines. The orientation of the transects and spacing of interpreted videoframes was designed to: (1) reduce mapping error, (2) permit evaluation of differences in species distributionwith direction (anisotropy), (3) ensure sampling of less common species; and (4) vary the distance between

SEAGRASS DISTRIBUTION MAPPING 389

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 6: Probabilistic large-area mapping of seagrass species distributions

samples to provide information about species distribution at multiple spatial scales (Hewitt et al., 1998;Balestri et al., 2003).

Presence or absence of the following species of seagrass were recorded: Amphibolis antarctica, A. griffithii,Posidonia australis, P. coriacea, P. sinuosa, Halophila ovalis and Zostera tasmanica. A ‘sparse’ systematicdataset was constructed in which species were noted at every tenth GPS position (approximately every50m). These data were stratified by species, and 20 records from each species chosen as starting points foridentifying an additional nine georeferenced frames (creating a ‘clustered’ dataset) to provide informationabout species patch size and distribution at shorter distances.

The reliability of the dataset rests on the accuracy and repeatability of the video interpretations. Fromhigh-quality video footage (i.e. good lighting and camera angle, camera positioned 0.25m above canopy) allof the major species in this locale can be readily distinguished. However, owing to vessel movement,shadows, water column turbidity, and rapid changes in bathymetry, the video quality was not alwaysoptimal. The reproducibility of video interpretations was evaluated by a single interpreter viewingapproximately 60 typical video frames three times, over a period of several months, and by two additionalinterpreters independently. Species identification agreement was calculated as the average number ofinterpretations (or repetitions) in agreement for each frame in which at least one interpreter has identified aspecies. This is reported as a percentage by weighting the average by the number of frames considered. Theagreement between interpreters was calculated for seagrass species presence/absence, and for speciespresence only. Species presence/absence includes a large number of unvegetated frames on which allinterpreters agreed, inflating the reproducibility of observations. The small number of data replicates(three) means the average and standard deviations are not robust, but they give an indication of whichspecies were most readily recognized by multiple observers, and which were problematic.

Mapping seagrass extent

The extent of perennial (dense canopy) seagrass in Owen Anchorage was determined through classificationof the aerial photography using several image-processing techniques implemented on spatially nestedanalysis windows. Classification accuracy was improved in areas of low light, including deep water, bychanging the spatial extent of the area classified at one time (Woodcock and Hayward, 1992). OwenAnchorage was initially subdivided into nine sections with visually differing patterns of seagrassdistribution, and classified separately. Each area was then divided into 1-km2 tiles to improve the localclassification consistency. In areas where visual inspection revealed poor classification, the mapping regionswere subdivided into 200� 200-m tiles and re-classified. This provided a good compromise betweenclassification accuracy and the increased time necessary to complete classification using smaller tiles. Threeclassification methods were applied, including ISODATA (‘Iterative Self-Organizing Data AnalysisTechnique’) (Tou and Gonzalez, 1974) unsupervised classification, Spectral Angle Mapper (SAM)supervised classification (Kruse et al., 1993), and textural analysis (Schowengerdt, 1983; Jensen, 1986) withmanual correction from photographs taken at lower turbidity levels. ISODATA classification and texturalanalysis was implemented in ERDAS Imagine 8.5 (Erdas Incorporated, 2002), and Spectral Angle Mapperusing MultiSpec version 2.7 (Purdue Research Foundation, 2004).

Species mapping

Modelling spatial dependence

The seagrass species indicator data (species presence ¼ 1; species absence ¼ 0) were used to measure thelikelihood that any two observations separated by a given distance do not have the same species present,and assembled as indicator semivariograms (hereafter referred to as variograms). The larger the indicatorvariance (y-axis of the variogram), the more disconnected or discontinuous the species patterns. Differences

K.W. HOLMES ET AL.390

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 7: Probabilistic large-area mapping of seagrass species distributions

in the distribution of species with cardinal direction were tested, and used for mapping in the cases wherethe data showed strong directional trends.

The indicator variogram is calculated as the average squared semi-difference between every possible pairof observations (Goovaerts, 1997):

glðh; zkÞ ¼1

2NðhÞ

XNðhÞ

a¼1

½iðua; zkÞ � iðua þ h; zkÞ�2 ð1Þ

where gl is the measure of similarity, N(h) is the number of pairs of observations at distance h; z is theindicator threshold of species presence or absence with the coordinates ua. The threshold k was set to 0.5%to be sure that all species presence observations recorded from the video footage were included formapping.

The spatial dependence is modelled by fitting a line to the experimental variogram under the constraintsof positive semi-definiteness (Goovaerts, 1997). This model quantifies the distances over which informationfrom neighbouring samples can be used to improve estimates at unsampled map locations, and is input tothe interpolation routine. The public domain software package GSLIB (Deutsch and Journel, 1998) wasused for modelling and interpolation.

Single species maps

Maps of the probability of species presence were constructed using ordinary indicator block kriging.Geostatistical interpolation (kriging) optimizes estimation by separating spatial variability into three parts:(1) deterministic variation (a trend), (2) spatial autocorrelation unexplained by known physical processes,and (3) unexplained variability or noise (Burrough and McDonnell, 1998). The indicator kriging algorithmis a linear combination of the observations (species presence/absence) at all sample locations, and predictsthe likelihood of species presence from nearby samples in a local neighbourhood around the unsampledlocation. The predictions are weighted by correlative relationships quantified in the variogram model. Thegeneral equation for indicator kriging is (Goovaerts, 1997):

½Iðu; zkÞ�� ¼XnðuÞ

a¼1

laðu; zkÞ½Iðua; zkÞ� ð2Þ

where laðu; zkÞ is the weight from variogram modelling assigned to the indicator data Iðua; zkÞ:Kriged mapsshowing the probability of species presence were produced over the study area in 20� 20m pixels. Maps ofkriging variance were also produced which show the estimation error across the study region. The krigingvariance is a function of the spatial configuration of the samples and the semivariogram model, and givesan indication of the accuracy of mapped estimates.

Multiple species maps

The kriged maps of species probability were used to generate binary maps of species presence or absence atthe following probabilities: 0.25, 0.50, 0.75, and 0.95. Each map was assigned a unique index for speciespresence, and the maps for all species were combined in a geographic information system (GIS)environment to create a composite map of species for each probability threshold. The maps show thedistribution of all seagrass species with a given likelihood of species presence. The map of perennial seagrassdistribution classified from the aerial photography was used to constrain the species estimations to onlythose areas known to be vegetated.

SEAGRASS DISTRIBUTION MAPPING 391

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 8: Probabilistic large-area mapping of seagrass species distributions

RESULTS

Delineation of seagrass extent

Multiple methods were tested for image classification, and evaluated at test sites (deep water, shallow water,single species, multiple species, dredging affected) by assessing the correlation between image-processingclassification results and expertly corrected local classifications done using histogram segmentationtechniques. The majority of the shallow areas were mapped using unsupervised ISODATA classificationðr ¼ 0:94Þ: In areas of deeper water or areas with a strong light/shade gradient, ISODATA performancewas poor, but was improved by combining it with Spectral Angle Mapper (SAM) supervised classificationðr ¼ 0:82Þ: Where active sediment plumes from dredging activity obscured the seafloor, textural analysisðr ¼ 0:35Þ was used in combination with ISODATA and manual correction using aerial photographs takenduring other time periods when the water column was clear. Approximately 1140 ha (14%) of the studyarea were classified using ISODATA, 6990 ha (84%) with a combination of ISODATA and SAM, and153 ha (2%) with ISODATA, textural analysis, and manual correction.

Table 1. Summary of combinations of species present at the 95, 75, 50, and 25 % probability levels from the kriged maps. Unitscovering more than 1% of the study area are included. NA=Not applicable (not present in field area)

Seagrass species Area

ha % ha % ha % ha %

Total vegetated area 1983.48 24.3Total unvegetated area 6188.52 75.7

Within vegetated areas: 95%probability

75%probability

50%probability

25%probability

Unidentified seagrass 1636.6 82.5 966.84 48.7 487.76 24.6 129.04 6.5Amphibolis (A) 188.08 9.5 386.12 19.5 284.8 14.4 119.88 6.0Posidonia sinuosa (Ps) 70.92 3.6 250.96 12.7 367.96 18.6 404.16 20.4Posidonia coriacea (Pc) 66.6 3.4 194 9.8 210.44 10.6 161.08 8.1Pc, A 17.48 0.9 141.12 7.1 434.52 21.9 646.36 32.6Ps, A 2.24 0.1 23.04 1.2 111.88 5.6 173.08 8.7Posidonia australis (Pa) 1.56 0.1 20.96 1.1 64.08 3.2 78.64 4.0Ephemeral speciesa (E) NA NA 0.32 0.0 6.92 0.3 24.72 1.2Ps, Pc NA NA 0.12 0.0 1.84 0.1 16 0.8Pa, Ps NA NA NA NA 5.68 0.3 56.28 2.8Ps, Pc, A NA NA NA NA 4.24 0.2 103.72 5.2Pc, E NA NA NA NA 1.72 0.1 20.12 1.0Ps, E NA NA NA NA 1.28 0.1 9.24 0.5Pa, Pc, A NA NA NA NA 0.2 0.0 4.48 0.2Pa, A NA NA NA NA 0.16 0.0 1.84 0.1Pa, E NA NA NA NA NA NA 14.12 0.7A, Pc, E NA NA NA NA NA NA 7.08 0.4A, E NA NA NA NA NA NA 6.08 0.3Pa, Ps, A NA NA NA NA NA NA 2.08 0.1Pc, Ps, E NA NA NA NA NA NA 1.84 0.1Pa, Pc NA NA NA NA NA NA 1.08 0.1A, Ps, E NA NA NA NA NA NA 1.4 0.1Ps, Pc, A, E NA NA NA NA NA NA 0.64 0.0Pa, Ps, Pc NA NA NA NA NA NA 0.24 0.0Pa, Ps, Pc, A NA NA NA NA NA NA 0.2 0.0Pa, Ps, E NA NA NA NA NA NA 0.08 0.0

aEphemeral species ¼ Zostera tasmanica and Halophila ovalis.

K.W. HOLMES ET AL.392

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 9: Probabilistic large-area mapping of seagrass species distributions

Over the study area as a whole (8172 ha), 1983 ha (24%) were mapped as seagrass, and 6189 ha (76%) asunvegetated sediments or reef (Table 1; Figure 2). The majority of the seagrass mapped was located onSuccess and Parmelia Banks, where the water is shallower, therefore light is less limiting to perennialseagrass growth, and identification by aerial photography classification was relatively unambiguous.

Seagrass species observations

A total of 52 video transects covering 113 km were collected in Owen Anchorage. From the video footage,seagrass species presence/absence was recorded in 2160 frames at sparse intervals (approximately 1 every50m), plus 1247 clustered observations (approximately every 5m over 60-m intervals), resulting in a total of3407 samples (Table 2; see Figure 2 for sample locations). Posidonia coriacea was the most commonlyidentified species ðn ¼ 1216Þ; followed by Amphibolis griffithii ðn ¼ 1059Þ and P. sinuosa ðn ¼ 721Þ.

Figure 2. Locations of species presence/absence samples, interpreted from georeferenced underwater video footage.

SEAGRASS DISTRIBUTION MAPPING 393

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 10: Probabilistic large-area mapping of seagrass species distributions

Halophila ovalis and Zostera tasmanica, the ephemeral species, were less frequently observed ðn ¼ 319Þ:Overall, 91% of seagrass species identified from video were perennials (Posidonia and Amphibolis) and 9%were ephemeral seagrasses (Halophila and Zostera tasmanica).

The assessment of uncertainty from the video interpretation revealed fairly high consistency for oneinterpreter, and among interpreters for overall identification of vegetation presence and species(Table 3(a)). When frames in which no interpreter recorded a species were ignored (note change innumber of frames assessed), the lowest agreement among interpreters on species identity occurred forAmphibolis antarctica which was only identified by one of the three interpreters, and the highest forA. griffithii and P. coriacea, with 88% and 72% average agreement respectively (Table 3(b)). In some cases,there was higher agreement among the three independent interpreters than among the three repetitions by asingle interpreter. This emphasizes the difficulty of identifying plants to the species level from video, givenchanging conditions in light, camera angle, proximity to the plant, and water column turbidity.

The comparison among multiple interpretations by one person and several different people givesanecdotal evidence of which species were most difficult to identify. There were consistent problemsdistinguishing Amphibolis antarctica from A. griffithii from the video footage, so the two species werecombined, and modelled jointly as Amphibolis spp. in the final species maps. From past fieldwork, it is

Table 2. Counts (n) of seagrass species presence from underwater video footage

Video seagrass data n %

Total no. of frames 3407Regular distribution 2160Clustered samples 1247

Total sites with seagrassa 1461 67.6

Seagrass speciesAmphibolis antarctica 85 2.3Amphibolis griffithii 1059 29.2Posidonia australis 225 6.2Posidonia coriacea 1216 33.5Posidonia sinuosa 721 19.9Halophila ovalis 167 4.6Zostera tasmanica 152 4.2

Total seagrass id 3625 100.0

Perennial vs. ephemeral speciesPerennialPosidonia 2162 59.6Amphibolis 1144 31.6

EphemeralHalophila and Zostera 319 8.8

Other occurrencesReef 132Wrack 529Algae 295Epiphytes 112Patchy 210Bad video quality 311Uncertain ID 445

aCalculated from regularly distributed samples only.

K.W. HOLMES ET AL.394

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 11: Probabilistic large-area mapping of seagrass species distributions

known that A. griffithii is more common in this region. Likewise, the two ephemeral speciesHalophila ovalisand Zostera tasmanica had insufficient numbers of observations for variogram construction, so they werecombined and mapped jointly as ‘ephemeral species’ for a more qualitative description of where one or theother of these ephemeral, colonizing species may be found.

Modelling spatial dependence and mapping

Variograms represent changes in attribute similarity between sample locations as distance between thesamples increases. If there are no predictable spatial relationships or patterns, the variogram will be a flatline (e.g. randomly shuffled X,Y coordinates in Figure 3). The indicator variograms for every speciesshowed strong spatial dependence (e.g. Posidonia coriacea in Figure 3). The shape of the variogram ismodelled by fitting three parameters, namely the nugget (where the model crosses the y-axis), the sill(y-value where the graph plateaus), and the range (the distance on x-axis at which the sill occurs) (Figure 3,Table 4). The variograms were typically fitted with exponential models, indicating rapidly changingpatterns over short distances. The range can be interpreted as the maximum distance over which spatialdependence occurs. The variogram is modelled to a distance of half of the study area in order to reduceartefacts from comparing only data in the edges of the site. However, to achieve a good model for mappingrather than interpretive purposes, variograms that continue to increase past this distance can be fittedby using a very large range, as was the case for P. sinuosa in the east–west direction (Table 4,range ¼ 100 000 m). With the exception of the ephemeral species, the variograms of all seagrasses had abreak in slope at relatively short distances (range ¼ from 100 to 350m), and gradually increasing variability

Table 3. Video interpretation agreement between interpreters, and repeat observations for 60 video frames selected to represent allspecies identified: N ¼ number of frames assessed; Average=average number of interpretations in agreement; and Std:Dev: ¼ 1standard deviation around the average. (a) agreement among interpreters on species presence and absence. (b) agreement among

interpreters for only those frames in which at least one interpreter identified a species (species presence only, note N)

(a) Agreement on species presence/absence

1 Interpreter, 3 repetitions 3 Interpreters, 1 repetition

N Average (%) Std.Dev. (%) N Average (%) Std.Dev. (%)

Amphibolis antartica 54 93 20 60 94 19Amphibolis griffithii 54 95 12 60 97 9Posidonia australis 54 95 12 60 95 12Posidonia coriacea 54 89 16 60 86 17Posidonia sinuosa 54 95 12 60 87 16Halophila ovalis 54 99 5 60 98 8Zostera tasmanica 54 99 7 60 97 10

(b) Agreement on species

1 Interpreter, 3 repetitions 3 Interpreters, 1 repetition

N Average (%) Std.Dev (%) N Average (%) Std.Dev. (%)

Amphibolis antartica 5 33 0 0 n.a. n.a.Amphibolis griffithii 18 81 23 19 88 23Posidonia australis 8 46 25 9 52 18Posidonia coriacea 31 73 29 46 72 28Posidonia sinuosa 13 74 28 24 47 17Halophila ovalis 1 33 n.a. 4 58 17Zostera tasmanica 3 56 38 8 50 31

SEAGRASS DISTRIBUTION MAPPING 395

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 12: Probabilistic large-area mapping of seagrass species distributions

x

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

xxxxxxxx xxx xxxx x

xxxxxxxxxxxxxxxxxxxxxx

xxxxxxxxxxxxxxxxxx

xxx xxxxxxx

xxxγ

Distance (m)

0 500 1000 1500 2000 2500 30000.00

0.05

0.10

0.15

0.20

0.25

Variogram for Posidonia coriacea

Variogram model Variogram after X,Y coordinates randomly shuffled

Sem

ivar

ianc

e,

γ

Distance (m)

(B). Nested variogram models

0 500 1000 1500 2000 2500 30000.00

0.05

0.10

0.15

0.20

0.25

(A). Variograms and model used for interpolation

Structure (2)

Structure (1)

Nugget

Figure 3. Plots of (A) the Posidonia coriacea variogram (bullets) and final model used for interpolation (line), a variogram of the samedata with the location coordinates randomly shuffled (crosses), and (B) the three nested variogram models that were combined to

produce the final model. See Table 4 for model parameters.

Table 4. Semivariogram parameters used for species map interpolation. Numbers in parentheses refer to the number of nestedvariogram structures modelled, as illustrated in Figure 3(B)

Species Anisotropic? Nestedmodels?

Variogrammodel type

Nugget Sill Range (m)

Amphibolis No Yes 0.02(1) Exponential 0.1 350(2) Spherical 0.06 1800

Halophila ovalis +Zostera tasmanica

No No (1) Spherical 0.06 0.06 2800

Posidonia australis No Yes 0.01(1) Exponential 0.02 200(2) Exponential 0.02 2800

Posidonia coriacea No Yes 0.04(1) Exponential 0.05 100(2) Exponential 0.17 2800

Posidonia sinuosa Yes Yes 0.03north–south (1) Exponential 0.12 350

(2) Exponential 0.03 2500east–west (1) Exponential 0.12 800

(2) Exponential 0.03 100 000

K.W. HOLMES ET AL.396

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 13: Probabilistic large-area mapping of seagrass species distributions

to a second sill (range ¼ from 1800 to 2800m). In the case of P. sinuosa, the maximum distance showingspatial dependence recorded in the variograms is 2800m in the north–south direction, and essentially 800min the east–west direction. The ephemeral species increased in variance to approximately a range of 2800m,and could not be modelled as accurately as the other species.

Several models can be combined or nested to fit a more complex variogram, as was done for all theseagrasses except for the ephemeral species. Nested variogram models imply that the pattern beingmodelled is caused by multiple processes acting on separate spatial scales, with an overall additive effect. Inthe case of seagrass, there are plausible physical explanations for such nested patterns, such as thepossibility that patch size is locally related to patch age or growth of rhizome, over larger distances byhydrodynamic processes, and regionally by light availability which is related to water depth. Nested modelswere required for all species except the ephemerals.

Continuous, gridded maps of the probability of species occurrence across the study area were constructedusing indicator block kriging (Figure 4(A)). In contrast to manual methods of classifying sparse field datainto vegetation or habitat classes, these maps graphically depict as a continuum the inherent uncertainty inassigning a given location to one class or another. For species with fewer presence data (e.g. Posidoniaaustralis), all of the probabilities are lower than for a commonly occurring species (e.g. P. coriacea).Therefore, depending on the map-user’s application, the probabilities can be interpreted in terms ofabsolute values (e.g. what is the most common species found in the study area?) or as relative values (e.g.where is the most likely location to find a large patch of P. australis?).

From the continuous, gridded probability values, binary maps showing species presence/absence (Figure4(B)) and species assemblage maps were created at the following thresholds: 95%, 75%, 50% and 25%likelihood of presence (Figures 5 and 6). These maps represent all predicted species and combinations ofspecies covering at least 1% of the vegetated study area. At the 95% probability threshold, large areas withvegetation were not assigned to any specific species (see ‘unidentified seagrass’, Table 1). None of thespecies were mapped over these areas with high confidence, however, the presence of perennial seagrass ispresumed because it was mapped by the aerial photography classification. Using this method, the areas ofidentified seagrass species increase steadily as the probability threshold is lowered from 95% to the lowest‘confidence’ map at 25% probability of species presence (Figure 7). The incidence of less frequentlyobserved species also increases as the probability threshold is lowered, as is well illustrated by Posidoniaaustralis (Figures 5–7; Table 1).

The species mapped with the highest confidence (i.e. largest area at the 95% probability level) wasAmphibolis (9.5%; Table 1), which was identified most frequently from the video (Table 2). Posidoniasinuosa and P. coriacea were both mapped as present over approximately 3.5% of the area (Table 1),although there were nearly twice as many P. coriacea presences identified in the video (Table 2). The leastcommonly identified species (P. australis and the ephemeral species) were mapped virtually nowhere with95% confidence, owing to the low number of samples identified. The area covered by each species increasedas the threshold lowered from 95% to 25% confidence, but the proportion of the study area covered witheach species remained relatively consistent with the number of video frames counted as having a particularspecies present. That is, at the 25% confidence level, the area covered by each species decreased in thefollowing order: Amphibolis, P. coriacea, P. sinuosa, P. australis, and the ephemeral species. The ephemeralspecies and P. coriacea covered slightly less area than might have been predicted from the raw counts ofspecies presence.

The single species coverage at fixed probability thresholds remains closely coupled with the proportionsexpected from the raw video counts of seagrass presence; however, the patterns of mixed speciesassemblages provide insight into patterns of colonization or growth habit. Although only a smallpercentage of the area was mapped at the species level with 95% confidence, a large percentage ofP. coriacea and Amphibolis were mapped as a mixed assemblage (1% of total area, but 9% of Amphibolisand 21% of P. coriacea: Table 1, Figure 7). Only 3% of P. sinuosa overlapped with Amphibolis and none

SEAGRASS DISTRIBUTION MAPPING 397

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 14: Probabilistic large-area mapping of seagrass species distributions

Figure 4. Example of (A) a continuous (gridded) map of the probability of species presence, and (B) derived binary maps of speciespresence at the 95%, 75%, 50% and 25% probability thresholds.

K.W. HOLMES ET AL.398

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 15: Probabilistic large-area mapping of seagrass species distributions

Figure 5. Map of seagrass species and combinations of species with a (A) 95% and (B) 75% probability of occurrence.

SEAGRASS DISTRIBUTION MAPPING 399

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 16: Probabilistic large-area mapping of seagrass species distributions

Figure 6. Map of seagrass species and combinations of species with (A) a 50% and (B) a 25% probability of occurrence.

K.W. HOLMES ET AL.400

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 17: Probabilistic large-area mapping of seagrass species distributions

with P. coriacea. This pattern remained consistent at the lower probability thresholds. At 25% confidence,89% of the area mapped as Amphibolis and 81% of P. coriacea were in mixed assemblages, largely together(32.6% of perennial seagrasses), while only 47% of P. sinuosa was in a mixed assemblage, mainly withAmphibolis (8.7% of total seagrass), mixed Amphibolis and P. coriacea (5.2%), or with P. australis (5.2%).The ephemeral species, in contrast, were more similar to Amphibolis and P. coriacea distributions,exhibiting a large amount of overlap with other species (71% in mixed assemblages at a 25% probabilitythreshold), though in general they covered a fairly small area. These species probably cover a much largerarea, particularly in recently dredged and deeper waters (>10m), but because they were not dense canopy

Figure 7. Area (ha) over which the probability of seagrass species presence exceeded the threshold (95%, 75%, 50%, 25% probabilityof presence). The numbers printed above each bar refer to the percentage of the species area that fell within a mixed community. Total

seagrass cover in the study area was 1983 ha, determined using aerial photography classification.

SEAGRASS DISTRIBUTION MAPPING 401

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 18: Probabilistic large-area mapping of seagrass species distributions

perennials, their full extent could not be mapped from the aerial photography classification of seagrasspresence/absence.

Maps of kriging variance reflect the sample layout with much higher errors in areas with few or nosamples, particularly between transects and around the edges of the map (Figure 8). It is important toremember, however, that errors due to sampling bias or interpretation error in the video footage and aerialphotography analysis are not reflected in the kriging variance, and thus it does not provide a completerepresentation of the final map accuracy.

DISCUSSION

The mixture of species in Owen Anchorage is probably one of the most diverse for seagrass meadows intemperate waters, with up to seven species co-occurring in extensive meadows. The spatial pattern ofindividual species as components of seagrass meadows is substantially different owing to their specificbiological attributes, such as rates of clonal growth (Kendrick et al., 2005) and rapidity of rhizome maturity(Marba and Duarte, 1998), and is also affected during statistical analysis by the total number of presences

Figure 8. Kriging variance for the most abundantly sampled seagrass species (Posidonia coriacea) shown within boundaries of seagrasspatches mapped from aerial photograph classification.

K.W. HOLMES ET AL.402

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 19: Probabilistic large-area mapping of seagrass species distributions

identified and the layout of the samples. Each of the species studied here demonstrate different spatialpatterns of growth. Posidonia sinuosa rhizomes are long-lived and meadow-forming, and are commonlyfound in deeper water than most of the other species (Kirkman and Walker, 1989); Amphibolis tend to bemuch faster-growing, form smaller patches, and appear to persist in disturbed areas and higher energyenvironments (Cambridge, 1999); P. coriacea is described as a clumping species (Cambridge, 1999), whereclumps of seagrass are a minor component of shallow unvegetated sands and are common colonizers ofbare sediments, able to grow in areas with more vigorous water movement (Kirkman and Walker, 1989).Posidonia australis has similar growth habits to P. sinuosa, tends to form meadows in areas of accretingsand, and is often associated with patchy reefs within lagoons (Kirkman and Walker, 1989). Ephemeralspecies are interspersed as understorey and minor components of meadows with typically patchydistributions.

The characteristic species growth habits influence the patterns of distribution over large areas. Thesepatterns can be qualitatively interpreted from the experimental variograms constructed from the videodata. With the exception of the ephemeral species, all of the seagrasses were modelled with nestedvariograms, suggesting that additive processes acting at different spatial scales influence speciesdistributions. However, the distances over which the nested variograms were fitted were adjusted toproduce best-fit models of spatial dependence for mapping rather than for defining the spatial scalescommon among all of the species for exploring potential multi-scale relationships. Therefore, differentmathematical models and ranges were used for each species, and direct comparisons among the variogramparameters are not readily interpretable (Goovaerts, 1997). That said, several general patterns are evidentin the species variograms which relate to their distribution over a large area. All species were best fitted witha nugget effect and an exponential variogram model at the finest resolution, suggesting high variability overshort distances. For the ephemeral species, the nugget alone accounted for 48% of the total variance,indicating extreme variability over short distances, and is an indication that the sampling design did notprovide an adequate representation of ephemeral species distribution. The kriged maps model only half ofthe total variability in distribution of ephemeral species (Table 4). This is probably due to a combination offactors including the disparate growth habits of the two species combined to form the ephemeral speciesgroup, and potentially reflects that the full distribution of these species was not sampled, and thus theirnatural distributions are not accurately represented in the variogram or the kriged maps. The nugget effectsfor the other species explain from 11% to 22% of total variance. Species with fewer recorded presences tendto have a larger nugget effect, interpreted as unexplained short distance variability, and a noisy variogram(e.g. P. australis). The majority of the variance for all of the single taxa (Amphibolis, P. australis, P. sinuosa,and P. coriacea) is explained within approximately 500m, and gentle longer distance trends account for theremaining variance. All species showed increasing variance with distance to about 3 km. The one exceptionwas P. sinuosa. Meadows of P. sinuosa are elongated in the north–south direction. The effective sills forP. sinuosa are 2.5 km (north–south), and 0.8 km (east–west).

Maps for individual species were developed for probability of presence and then explored by assessing allspecies jointly at fixed probability thresholds. Although areas mapped with high probability values ofspecies presence provide high confidence of finding a particular species when in the field, areas mapped withlow probabilities provide insight into the process of colonization by individual species (i.e. incursion byinvasive species), and for targeting the most likely areas to find uncommon species. While areas with lowprobability of species presence would not normally be delineated in typical mapping exercises, they areparticularly useful to resource managers. This type of data is very flexible, especially compared to typicaldiscrete (polygon) maps of seagrass presence or species assemblages, and provides a base for ecologicalhypothesis formation and monitoring for a number of conservation activities. The probability thresholdmaps of multiple species simulate traditional vegetation assemblage maps, but provide a consistent level ofmapping confidence across the study area. For example, where P. coriacea and Amphibolis are representedas a mixed species complex in Figures 5 and 6, both species have the same probability of presence. Thus,

SEAGRASS DISTRIBUTION MAPPING 403

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 20: Probabilistic large-area mapping of seagrass species distributions

rather than providing a single map outcome, a digital library has been constructed which allows the usergreater flexibility for data query and site management.

The maps constructed at fixed probability thresholds reveal differences in associations among speciesover large areas. Although Posidonia sinuosa is not as commonly found in the towed video survey asAmphibolis or P. coriacea, it is mapped as present over a larger than expected proportion of the field area inthe 95% probability threshold maps (71 ha). Also, it occurs dominantly as a single species, rather than inmulti-species complexes at all probability thresholds. Patchily distributed species, like P. coriacea, arecommon in 95% confidence maps, but occur in both single species (67 ha) and mixed species meadows(17 ha). At the 25% probability threshold, the most common seagrass assemblage is a mix of P. coriaceawith Amphibolis, while P. sinuosa as a single species covered the second largest area. The areas mapped asmixed species including P. sinuosa are predominantly Amphibolis or P. australis (Kirkman and Walker,1989). In detailed field investigations, P. sinuosa has been found to commonly occur with P. australis alongembankments, but in less enclosed areas forms meadows alone with Amphibolis (Cambridge and Kuo,1979). While maps of mixed species at probability thresholds also show a large area populated by P. sinuosawith Amphibolis and P. coriacea, this is most likely due to the overwhelming presence of the other twoacross the study area at lower probability levels.

Kendrick et al. (2000) found P. sinuosa occupied the deeper edges of the sediment banks whereas themixed species assemblages of P. coriacea and Amphibolis were dominant in shallower areas. Thispartitioning of species distribution is both a product of seagrass loss, mostly of P. sinuosa, correlated withindustrial impacts on Parmelia Bank and the dynamic decadal-scale colonization of Success Bank by theseagrasses P. coriacea and A. griffithii documented in previous studies (Kendrick et al., 1999, 2000, 2002).

Traditional maps of vegetation communities provide no information about the spatial variation inmapping confidence (i.e. potential map error), giving the impression that any reported error is spatiallyinvariant. As shown by the map of kriging variance (Figure 8), which portrays the partial map error relatedto interpolation from the sample configuration, this is obviously not the case. The continuing presence of‘unidentified seagrass’ in the final map of the 25% probability threshold interval demonstrates thatpredicting species presence remains difficult in some portions of the study area. In traditional maps thisdiscrepancy would be concealed within the final map units. The probability levels of species occurrence andkriging variance can be used to optimize additional sampling (Defeo and Rueda, 2002; Pardo-Iguzquizaand Dowd, 2005). If the distribution of a particular species was desired, then areas mapped as having 50%probability of presence would require additional sampling; similarly for mapping mixed communities, theareas mapped as present at the 50% probability threshold which did not appear at the 75% thresholdshould be targeted. These are the regions mapped with the largest ambiguity, and where a limited numberof additional samples could greatly impact the results. At the extremes of the distribution many moresamples would be required to change the mapped result.

Previously published studies of seagrass species distributions have only provided approximatedistributions of species at points in the landscape (punctual kriging) based on composite quadrat ortransect seagrass density estimates (Fourqurean et al., 2001; Durako et al., 2002), or binary maps of speciespresence (Robbins and Bell, 2000). A probabilistic approach toward understanding the effect of waterquality on seagrass species composition has been used to explore future scenarios following land-use change(Fourqurean et al., 2003), illustrating the potential benefits of probabilistic mapping. Maps of individualspecies are more applicable to conservation goals, because some species are more sensitive to environmentalchange and thus can be used for detecting response to their environment (e.g. disturbance). Maps ofseagrass extent without species information, such as those developed from aerial photography, can only beused to monitor complete removal of vegetation. Although more general ecosystem information is providedby maps of seagrass communities, local extinction of individual species or change in contribution to thecommunity is not tracked. Thus, these maps may be misleading as to the actual impact of any disturbance.The species probability maps presented here form potentially the only large area (tens to hundreds of

K.W. HOLMES ET AL.404

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 21: Probabilistic large-area mapping of seagrass species distributions

kilometres) test of the seagrass species succession previously postulated for the Western Australiancoastline based on diver observations (Kirkman, 1985; Kirkman and Kuo, 1990).

Seagrass species mapping in the Owen Anchorage region combining detection from aerial photographyand mapped seagrass species distributions using indicator kriging provided a flexible output for multipleresource management and conservation applications. The development of seagrass presence maps from acombination of unsupervised and supervised classifications of aerial photographs is a standard approachfor mapping the presence of shallow-water benthos. The use of spatial statistics to develop probabilities ofspecies presence within these regions proved successful, and to our knowledge is the first application of thisapproach for mapping individual seagrass species. This methodology allows the exploration of individualspecies patterning as well as spatial emergence as related to the individual clonal biology of each species.This results in unique landscape features where P. sinuosa forms larger continuous and predominantlymonospecific meadows, whereas the more common Amphibolis and P. coriacea form multi-species patchymeadows. These mapped features suggest that the emergence of species patterns in seagrass landscapes isheavily influenced by differences in clonal growth among seagrass species.

ACKNOWLEDGEMENTS

We thank Cockburn Cement Limited as the sponsor of the 2004 seagrass species data collection and for the use ofpreviously assembled environmental datasets, and Oceanica Consulting Pty Ltd for their support in the field andprovision of the aerial photography. Many people participated in collection of species information through tow videoand scuba collection, with special thanks to Simon Grove, who was the primary video collector and interpreter.Funding for the authors was provided by the Cooperative Research Centre for Coastal Zone, Estuary and WaterwayManagement (Coastal CRC).

REFERENCES

Alberotanza L, Brando VE, Ravagnan C, Zandonella A. 1999. Hyperspectral aerial images. A valuable toolfor submerged vegetation recognition in the Orbetello Lagoons, Italy. International Journal of Remote Sensing 20:523–533.

Andrefouet S, Zubia M, Payri C. 2004. Mapping and biomass estimation of the invasive brown algae Turbinaria ornata(Turner) J. Agardh and Sargassum mangarevense (Grunow) Setchell on heterogeneous Tahitian coral reefs using4-meter resolution IKONOS satellite data. Coral Reefs 23: 26–38.

Balestri E, Cinelli F, Lardicci C. 2003. Spatial variation in Posidonia oceanica structural, morphological and dynamicfeatures in a northwestern Mediterranean coastal area: a multiscale analysis. Marine Ecology Progress Series 250:51–60.

Brown CJ, Cooper KM, Meadows WJ, Limpenny DS, Rees HL. 2002. Small-scale mapping of sea-bed assemblages inthe eastern English Channel using sidescan sonar and remote sampling techniques. Estuarine, Coastal and ShelfScience 54: 263–278.

Burrough PA, McDonnell RA. 1998. Principles of Geographical Information Systems. Oxford University Press:New York.

Cambridge ML. 1999. Growth strategies of Rottnest Island seagrasses. In The Seagrass Flora and Fauna of RottnestIsland, Western Australia, Walker DI, Wells FE (eds). Western Australian Museum: Perth, WA; 1–24.

Cambridge ML, Kuo J. 1979. Two new species of seagrasses from Australia, Posidonia sinuosa and P. angustifolia(Posidoniaceae). Aquatic Botany 6: 307–328.

Cambridge ML, Chiffings AW, Brittain C, Moore L, McComb AJ. 1986. The loss of seagrass in Cockburn Sound,Western Australia. II. Possible causes of seagrass decline. Aquatic Botany 24: 269–285.

Campey ML, Kendrick GA, Walker DI. 2002. Interannual and small-scale spatial variability in sexual reproduction ofthe seagrasses Posidonia coriacea and Heterozostera tasmanica, southwestern Australia. Aquatic Botany 74: 287–297.

Carruthers TJB, Dennison WC, Longstaff BJ, Waycott M, Abal EG, McKenzie LJ, Lee Long WJ. 2002. Seagrasshabitats of northeast Australia: models of key processes and controls. Bulletin of Marine Science 71: 1153–1169.

SEAGRASS DISTRIBUTION MAPPING 405

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 22: Probabilistic large-area mapping of seagrass species distributions

Castrignano A, Goovaerts P, Lulli L, Bragato G. 2000. A geostatistical approach to estimate probability of occurrenceof Tuber melanosporum in relation to some soil properties. Geoderma 98: 95–113.

Cochrane GR, Lafferty KD. 2002. Use of acoustic classification of sidescan sonar data for mapping benthic habitat inthe Northern Channel Islands, California. Continental Shelf Research 22: 683–690.

Cuevas-Jimenez A, Ardisson P-L, Condal AR. 2002. Mapping of shallow coral reefs by colour aerial photography.International Journal of Remote Sensing 23: 3697–3712.

Defeo O, Rueda M. 2002. Spatial structure, sampling design and abundance estimates in sandy beach macroinfauna:some warnings and new perspectives. Marine Biology 140: 1215–1225.

Deutsch CV, Journel AG. 1998. GSLIB: Geostatistical Software Library and User’s Guide. Oxford University Press:New York.

Dierssen HM, Zimmerman RC, Leathers RA, Downes TV, Davis CO. 2003. Ocean color remote sensing ofseagrass and bathymetry in the Bahamas Banks by high-resolution airborne imagery. Limnology and Oceanography48: 444–455.

Duarte CM. 1999. Seagrass ecology at the turn of the millennium: challenges for the new century. Aquatic Botany65: 7–20.

Duarte CM. 2002. The future of seagrass meadows. Environmental Conservation 29: 192–206.Durako MJ, Hall MO, Merello M. 2002. Patterns of change in the seagrass dominated Florida Bay hydroscape. In TheEverglades, Florida Bay, and Coral Reefs of the Florida Keys: An Ecosystem Sourcebook, Porter JW, Porter KG (eds).CRC Press: Washington, DC; 523–537.

Ferguson R, Korfmacher K. 1997. Remote sensing and GIS analysis of seagrass meadows in North Carolina, USA.Aquatic Botany 58: 241–258.

Fonseca MS. 1996. Scale dependence in the study of seagrass systems. In Seagrass Biology: Proceedings of anInternational Workshop, Kuo J, Phillips RC, Walker DI, Kirkman H (eds). Rottnest Island, Western Australia, 25–29January 1996, Faculty of Science, University of Western Australia, Nedlands, 95–104.

Fonseca MS, Whitfield PE, Kelly NM, Bell SS. 2002. Modeling seagrass landscape pattern and associated ecologicalattributes. Ecological Applications 12: 218–237.

Fourqurean JW, Willsie A, Rose CD, Rutten LM. 2001. Spatial and temporal pattern in seagrass communitycomposition and productivity in south Florida. Marine Biology 138: 341–354.

Fourqurean JW, Boyer JN, Durako MJ, Hefty LN, Peterson BJ. 2003. Forecasting responses of seagrass distributionsto changing water quality using monitoring data. Ecological Applications 13: 474–489.

Frederiksen M, Krause-Jensen D, Holmer M, Laursen JS. 2004. Spatial and temporal variation in eelgrass (Zosteramarina) landscapes: influence of physical setting. Aquatic Botany 78: 147–165.

Goodchild MF. 1994. Integrating GIS and remote sensing for vegetation analysis and modeling: methodological issues.Journal of Vegetation Science 5: 615–626.

Goovaerts P. 1997. Geostatistics for Natural Resources Evaluation. Oxford University Press: New York.Green EP, Short FT. 2003. World Atlas of Seagrasses. University of California Press: Berkeley, CA.Hemminga MA, Duarte CM. 2000. Seagrass Ecology. Cambridge University Press: Cambridge.Hewitt JE, Thrush SF, Cummings VJ, Turner SJ. 1998. The effect of changing sampling scales on our abilityto detect effects of large-scale processes on communities. Journal of Experimental Marine Biology and Ecology 227:251–264.

Jensen JR. 1986. Introductory Digital Image Processing: A Remote Sensing Perspective. Prentice-Hall: Englewood Cliffs,NJ.

Kelly NM, Fonseca M, Whitfield P. 2001. Predictive mapping for management and conservation of seagrass beds inNorth Carolina. Aquatic Conservation: Marine and Freshwater Ecosystems 11: 437–451.

Kendrick GA, Eckersley J, Walker DI. 1999. Landscape-scale changes in seagrass distribution over time: a case studyfrom Success Bank, Western Australia. Aquatic Botany 65: 293–309.

Kendrick GA, Hegge BJ, Wyllie A, Davidson A, Lord DA. 2000. Changes in seagrass cover on Success and ParmeliaBanks, Western Australia between 1965 and 1995. Estuarine, Coastal and Shelf Science 50: 341–353.

Kendrick GA, Aylward MJ, Hegge BJ, Cambridge ML, Hillman K, Wyllie A, Lord DA. 2002. Changes in seagrasscoverage in Cockburn Sound, Western Australia between 1967 and 1999. Aquatic Botany 73: 75–87.

Kendrick GA, Duarte CM, Marba N. 2005. Clonality in seagrasses, emergent properties 14 and seagrass landscapes.Marine Ecology Progress Series 290: 291–296.

Kirkman H. 1985. Community structure in seagrasses in south Western Australia. Aquatic Botany 21: 363–375.Kirkman H, Kirkman J. 2000. Long-term seagrass meadow monitoring near Perth, Western Australia. Aquatic Botany67: 319–332.

Kirkman H, Kuo J. 1990. Pattern and process in southern Western Australian seagrasses. Aquatic Botany 37:367–382.

K.W. HOLMES ET AL.406

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)

Page 23: Probabilistic large-area mapping of seagrass species distributions

Kirkman H, Walker DI. 1989. Regional studies } Western Australia seagrass. In Biology of Seagrasses: A Treatise onthe Biology of Seagrasses with Special Reference to the Australian Region, Larkum AWD, McComb AJ, Sheperd SA(eds). Elsevier: Amsterdam; 157–181.

Kuo J. 2005. A revision of the genus Heterozostera (Zosteraceae). Aquatic Botany 81: 97–140.Kruse FA, Lefkoff AB, Boardman JB, Heidebrecht KB, Shapiro AT, Barloon PJ, Goetz AFH. 1993. The SpectralImage Processing System (SIPS) } interactive visualization and analysis of imaging spectrometer data. RemoteSensing of Environment 44: 145–163.

Levin SA. 1992. The problem of pattern and scale in ecology. Ecology 73: 1943–1967.Marba N, Duarte CM. 1998. Rhizome elongation and seagrass clonal growth. Marine Ecology Progress Series174: 269–280.

McRea JE, Greene HG, O’Connell VM, Wakefield WW. 1999. Mapping marine habitats with high resolution sidescansonar. Oceanologica Acta 22: 679–686.

Miller J, Franklin J. 2002. Modeling the distribution of four vegetation alliances using generalized linear models andclassification trees with spatial dependence. Ecological Modelling 157: 227–247.

Mumby PJ, Edwards AJ. 2002. Mapping marine environments with IKONOS imagery: enhanced spatial resolution candeliver greater thematic accuracy. Remote Sensing of the Environment 82: 248–257.

Orth RJ, Kendrick GA, Marion SR. In press. Posidonia australis seed predation in seagrass habitats of Rottnest Island,Western Australia: patterns and predators. Marine Ecology Progress Series, Manuscript M6114.

Pardo-Iguzquiza E, Dowd PA. 2005. Multiple indicator cokriging with application to optimal sampling forenvironmental monitoring. Computers and Geosciences 31: 1–16.

Pasqualini V, Pergent-Martini C, Calabaut P, Pergent G. 1998. Mapping of Posidonia oceanica using aerialphotographs and side scan sonar: application off the island of Corsica (France). Estuarine, Coastal and Shelf Science47: 359–367.

Robbins BD, Bell SS. 2000. Dynamics of a subtidal seagrass landscape: seasonal and annual change in relation to waterdepth. Ecology 81: 1193–1205.

Schowengerdt RA. 1983. Techniques for Image Processing and Classification in Remote Sensing. Academic Press:New York.

Short FT, Neckles HA. 1999. The effects of global climate change on seagrasses. Aquatic Botany 63: 169–196.Short FT, Wyllie-Echeverria S. 1996. Natural and human induced disturbance of seagrasses. EnvironmentalConservation 23: 17–27.

Sintes T, Marba N, Duarte CM, Kendrick GA. 2005. Nonlinear processes in seagrass colonisation explained by simpleclonal growth rules. Oikos 108: 165–175.

Tou JT, Gonzalez RC. 1974. Pattern Recognition Principles. Addison-Wesley: Reading, MA.Vis C, Hudon C, Carignan R. 2003. An evaluation of approaches used to determine the distribution and biomass ofemergent and submerged aquatic macrophytes over large spatial scales. Aquatic Botany 77: 187–201.

Woodcock CE, Hayward J. 1992. Nested-hierarchical scene models and image segmentation. International Journal ofRemote Sensing 13: 3167–3187.

SEAGRASS DISTRIBUTION MAPPING 407

Copyright # 2006 John Wiley & Sons, Ltd. Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 385–407 (2007)