MAPPING THE SHALLOW MARINE BENTHIC HABITATS OF ROTTNEST ISLAND… · Rottnest Island, Western...

3rd EARSeL Workshop Remote Sensing of the Coastal Zone 7-9 June 2007, Bolzano, Italy Harvey, page 1

MAPPING THE SHALLOW MARINE BENTHIC HABITATS OF ROTTNEST ISLAND, WESTERN AUSTRALIA

Matthew J. Harvey1, Halina T. Kobryn1, Lynnath E. Beckley1, Thomas Heege2, Peter Hausknecht3 and Nicole Pinnel1

1. Murdoch University, School of Environmental Science, Perth, Australia; matt(at)harves.net 2. EOMAP, Gilching, Germany; heege(at)eomap.de 3. formerly HyVista Corporation, Perth, Australia; hausknecht(at)woodside.com.au

ABSTRACT The introduction of new, high resolution hyperspectral sensors has led to growing interest in the development of techniques to utilise data from these instruments for mapping the shallow marine environment. The increased spectral resolution of the hyperspectral sensors allows the use of the unique spectral signatures of the individual habitat components to identify these components within the image. Hyperspectral data also allows for the mapping of habitats in shallow areas that are inaccessible to other methods such as hydro-acoustic mapping. The coastal waters surrounding Rottnest Island, Western Australia, provide a unique opportunity to apply hyperspectral imaging techniques in a temperate environment because of the oligotrophic conditions maintained by the Leeuwin Current. The shallow marine benthic habitats of Rottnest Island Reserve have been mapped to a depth of ~15 m, using spectral signatures contained in a library created from in-situ measurements of the dominant habitat components. Three lines of HyMap hyperspectral data flown for the Rottnest Island Reserve in April 2004 were corrected for sunglint, atmospheric effects and the influence of the water column using the Modular Inversion and Processing System which requires no inputs from parameters measured in the field. A decision tree based classification scheme which utilises a range of spectral similarity measures was used to map the different habitat components identified in the bottom reflectance image and the results were validated in the field using SCUBA divers. The shallow subtidal habitats found around Rottnest Island are generally dominated by either bare sand, reef with large macroalgae, such as Ecklonia radiata and Sargas-sum spp., or a number of different seagrass species. These new hyperspectral imaging techniques provide a platform for the mapping of shallow marine benthic habitats over a broad area, at a scale that is relevant to marine planners and managers.

INTRODUCTION Australia’s Oceans Policy was released in 1998 to provide a framework for integrated and ecosystem-based planning and management for Australia’s vast marine territories (1). At the core of the policy is the development of regional marine plans for Australia’s entire exclusive economic zone. The primary goals of the regional marine plans include ensuring marine ecosystem health into the future, protection of marine biodiversity, promotion of diverse and sustainable marine industries and ensuring the estab-lishment of a National Representative System of Marine Protected Areas (1).

For a marine protected area to be classed as representative of the ecosystems they are designed to protect, they must “reasonably reflect the biotic diversity of the marine ecosystems from which they derive” (2). Therefore, in order to define representative areas, there is a basic need to meas-ure and map the biodiversity of Australia’s vast coastline (3,4).

There has been a significant amount of research that has examined methods both defining and measuring biodiversity (4-6) and the need to use biodiversity surrogates for practical applications is widely accepted (4,7-10). The aim of a biodiversity surrogate is that it both serves as an reliable indicator of general biodiversity and is readily measurable in the environment (6). A number of


studies in Australia, on the use of biodiversity surrogates, have concluded that marine habitats can be used as a biodiversity surrogate with a reasonable level of confidence (9,11,12).

Historically, the mapping of marine benthic habitats has been done using traditional field methods, which are both costly and labour intensive. More recently, passive remote sensing techniques have been used for mapping marine benthic habitats (13-16). Initially, most remote sensing of ma-rine benthic habitats was done using aerial photography or multispectral satellite data, such as Landsat 7 TM+, with varying results (17-20). In addition, the bulk of the work has been restricted to shallow coral reef environments and freshwater systems, with reasonably clear water. There has been very little work conducted in temperate waters, due to the poor water clarity of most temper-ate marine environments. In the coastal regions of south-western Australia we have a unique op-portunity to apply this technology in the clear oligotrophic waters, maintained by the presence of the southward flowing Leeuwin Current (21,22).

The use of HyMap hyperspectral data to map the shallow marine benthic habitats of Rottnest Island Reserve is the focus of the current research project. The overall aim of the project was to utilise Hy-Map imagery to create thematic classification maps of the marine benthic habitats of Rottnest Island Reserve. The project used a spectral reflectance library of the dominant marine benthic habitat com-ponents found at Rottnest Island to classify hyperspectral data. The results can be applied, within a GIS framework, to various planning and management scenarios relevant to the Rottnest Island Re-serve.

METHODS

Study site This study was carried out in the waters surrounding Rottnest Island, which form part of a Class A marine reserve. The island lies approximately 18 km offshore from Fremantle, Western Australia, at latitude 32°00’ S and longitude 115°30’ E (Figure 1) is oriented in a generally east-west direc-tion. The waters of the reserve extend approximately 800 m from the shoreline, encompass an area of 3,828 ha and vary in depth from exposed inter-tidal platforms to water >40 m deep. The majority of the islands marine habitats are exposed to high wave action during the typical winter storms, which generally approach from a south-westerly direction. The swell will refract around most of the island to impact the majority of the island coast, with the exception of the bays located at the eastern end of the island (23). The waters around Rottnest Island are typically oligotrophic because of the influence of the anomalous pole-ward flowing eastern boundary current known as the Leeuwin Current (24). The Leeuwin current transports warmer, low salinity, low nutrient water southward along the Western Australian coast from tropical water northwest of Australia (25,26).

The shallow benthic habitats of Rottnest Island support a large diversity of macroalgae, which are generally dominated by either the canopy forming species Ecklonia radiata or a variety of Sargas-sum species or by a diverse range of lower growing foliose and filamentous algae, often referred to as turfing algae assemblages (23). There are nine species of seagrass found at Rottnest Island, with the dominant species belonging to the genera Posidonia and Amphibolis, which can form ex-tensive meadows in the sheltered habitats of the island (Figure 2).

Hyperspectral data collection and pre-processing Three flight lines of HyMap hyperspectral data were flown over Rottnest Island in April 2004 by HyVista Corporation (Figure 3). Three corrections were applied to the data: sun-glitter, atmos-pheric and water column, using the Modular Inversion and Processing System (MIP) (27,28). Only results from the preliminary analysis of flight line one is presented in this paper.

MIP uses a physically based process to extract from the data information on the water constitu-ents, bathymetry, bottom cover type and bottom reflectance, with no external inputs. The architec-ture of the program combines a set of general and transferable computational methods in a chain, which connects the bio-physical parameters of the water column with the measured sensor radi-


ances. The Rottnest hyperspectral data was corrected using a number of generic bottom cover types: bare sediment, light and dark submerged aquatic vegetation. The signatures used as inputs were extracted directly from the image.

N

EW

S

115º15’ 115º45’ 116º00’115º30’

115º15’ 115º45’ 116º00’115º30’

31º4

5’32

º00’

32º1

5’32

º30’

Rottnest Island

Garden Island

Fremantle

0 10 20kilometres

31º45’32º00’

32º15’32º30’

Study Area

Western Australia

Perth

Shoalwater Islands Marine Park

Marmion Marine Park

N

EW

S

N

EW

S

115º15’ 115º45’ 116º00’115º30’

115º15’ 115º45’ 116º00’115º30’

31º4

5’32

º00’

32º1

5’32

º30’

Rottnest Island

Garden Island

Fremantle

0 10 20kilometres

0 10 20kilometres

0 10 20kilometres

31º45’32º00’

32º15’32º30’

Study Area

Western Australia

Study Area

Western Australia

Perth

Shoalwater Islands Marine Park

Marmion Marine Park

Figure 1: Map indicating the Rottnest Island study area in south-western Australia. Dashed lines represent boundaries of marine protected areas.

Figure 2: Dominant habitat components used as the basis of the classification scheme for marine benthic habitats of Rottnest Island Reserve, Western Australia.

Figure 3: Three flight lines of uncorrected HyMap hyperspectral data of Rottnest Island, Western Australia. Data was collected in April 2004.


A hyperspectral library of spectral reflectance signatures of the dominant habitat components for Rottnest Island was created from in-situ spectral measurements. The data were collected using an Ocean Optics USB 2000 spectrometer fitted with a 30 m fibre-optic cable with a diameter of 400 µm and a whole acceptance angle of 22.4°. The reflectance signatures were processed to a reso-lution of 1 nm and re-sampled to match the spectral response of the 2004 HyMap sensor.

Spectral separation analysis The use of spectral distance metrics has been used successfully in hyperspectral analysis as a method of matching reference spectra to those of an unknown target (29,30). There are a number of different spectral distance metrics that have been developed and applied to hyperspectral data. These include the spectral angle (SA), the spectral gradient angle, the spectral information divergence (SID) and two measures that combine SID and SA, the SID is multiplied by either the sine or tangent of the SA (30-32). The SA is one of the most commonly implemented metric in hyperspectral remote sensing and was used in the classifications in this study. The spectral angle discriminates between spectra by calculating the angle between the mean vectors of two spectra. The SA is invariant to the scalar multi-plication and calculates the distance between spectra based solely on their shape (29). This feature makes it particularly appropriate to an application in the marine environment where the highly variable underwater light field can create changes in illumination that will result in changes in the reflectance values, although the spectral shape of the signature has not changed.

The nature of hyperspectral data with its large numbers of continuous bands there is often a significant data redundancy due to high levels of correlation between bands (33). Consequently techniques that reduce the dimensionality of the data are often carried out as part of data analysis for classification (29,34). Many of these techniques focus on reducing the number of bands used to carry out and classifi-cation, a process which can often increase the accuracy of the classifications and reduce the likelihood of false positives resulting from algorithms used over-fitting the data. The aim of this data reduction is to determine the wavelengths in the data that maximise the discrimination between groups, while minimis-ing the discrimination within groups. By minimising the number of bands used you are not only able to increase classification accuracy but also reduce processing time for these often very large data sets.

The approach taken in this study was to maximise the discrimination between spectral signatures of the different dominant habitat components which form the basis for the marine habitat classifica-tion scheme used for Rottnest Island (Figure 4). Preliminary analysis of the spectral library data using this technique used the spectral angle to determine the degree of separation between differ-ent habitat components at each classification level. A genetic algorithm was implemented to iden-tify the wavelength combinations which would provide the greatest discrimination between these groups of habitat components for each classification level.

Figure 4: Conceptual diagram of the hierarchal classification scheme used to classify marine ben-thic habitats of Rottnest Island Reserve, Western Australia. Numbering denotes the level within the scheme, with each level providing more detail.


Marine benthic habitat classification and accuracy assessment The marine benthic habitat classification of Rottnest Island was carried out based on the hierarchal classification scheme previously outlined using three core data sets (Figure 4; Table 1). All proc-essing of the data was carried out on the data before any geo-rectification has been carried out in order to preserve as pure as possible spectral information within each pixel and all land pixels were masked out from analysis. Any final outputs from the classification process were projected to UTM, Zone 50 South using the GLT data supplied by HyMap.

Processing the data to level one was carried out using band thresholds on the MIP bottom cover-age data to isolate pixels which contain only bare sediment. This resulted in a classification image which separates vegetated pixels from those with 50% bare sand and provides the information to segment the data for processing to level 2.

Table 1: Summary of the data used and the classification method applied to each level in the clas-sification scheme.

Classification level Data used Classification method

1 • Percentage cover of bottom types generated as output from MIP processing of HyMap data

Simple threshold analysis of % cover of bare sand (≥50% & ≥100%)

2 & 3

• Bottom reflectance data generated as output of MIP processing

• R library of dominant habitat components and their mixtures

A decision tree classification based on the values of the spectral metrics calculated on optimal bands for each level determined by spectral separation analysis.

Classification of the image to level two uses the bottom reflectance, the library spectra at HyMap resolution and the results from the level one processing to exclude regions of 50% sand from analysis. The methodology adopted for processing the spectral image data to level 2 was based on unpublished work, which determined that the dominant spectral endmember in a mixed pixel will be identified using spectral metrics. The data was processed by calculating the spectral distance of each pixel from library spectra. The spectral distance metric and image bands used is based on the results of the spectral separation analysis. Each pixel is assigned a class based on the library spectra which it is most similar to, which defines the dominant habitat component for that pixel. This enables the image to be further segmented into broad habitat categories of algae or seagrass dominated pixels (Figure 4).

Ground validation data was collected from a boat by recording the benthic habitat type at numer-ous locations within the bounds of the hyperspectral image. Although in an ideal situation the sam-pling scheme used would cover all depths and cover types, the reality of working in the marine environment meant that data was collected at as many locations as it was possible to access, given the prevailing conditions and navigational hazards such as exposed reefs. A system which consisted of a GPS combined with depth sounder, fitted with a differential receiver, and a laptop to monitor the boats location and log ground validation data. The boats position was monitored via a real time link between the GPS and ArcView 3.2 using the DNR Garmin extension (35) which pro-vided real time locations of the boat and provided the ability to record validation points with the greatest accuracy possible. For each validation point, data such as the location, benthic habitat type, time, depth and estimated positional accuracy was stored in an ESRI shapefile format.

Ground validation data was analysed using ArcView 3.2 using a methodology that incorporates the inherent positional uncertainty in both the hyperspectral imagery and the ground validation data. For each ground validation data point an array of “satellite” points were created at distances of 3.5, 7.0 and 10.5 m respectively and at 45° intervals starting at 0° using the radiating lines extension for ArcView 3.x (36). These distances were determined by calculating the mean estimated positional


error of the GPS location of the validation data and adding one pixel width to account for image geo-registration accuracy.

Figure 5: The sampling point array designed to take the inherent uncertainty of both the geo-rectification of the hyperspectral imagery and locations of the ground validation data. The white centre point is the actual validation point and the black point the associated “satellite” points.

To determine the accuracy of the image classification each ground validation point was compared to the class assigned as a result of the classification process. A classification result was deemed to be correct if either the validation point itself or one of its associated points intercepted the correct class. The accuracy was assessed for each classification level and expressed as an overall accu-racy along with an associated Kappa statistic, which was calculated using the Kappa analysis ex-tension for ArcView 3.x (37). The Kappa statistic is often used in assessing the accuracy of classi-fication of remotely sensed data as it not only assesses the overall accuracy but accounts for chance agreement between categories (38).

RESULTS

Marine benthic habitat classification and accuracy assessment The hyperspectral data was classified to Level 1 using the bottom coverage data generated as an output of the MIP processing. The bottom coverage image was classified using a threshold to as-sign all pixels with at least 50% bare sediment coverage to a bare sand habitat type and all other pixels to a vegetated habitat. This represents a valid segmentation of the data because of the eco-logical make-up of the marine benthic habitats of Rottnest Island which generally consist of either bare sand or have some form of submerged aquatic vegetation covering most available substrate. The results of this process showed that of the 78.42% of the pixels that occur in water <15 m deep, 56.40% were vegetated and 22.03% were bare sand (Figure 6). Accuracy assessment of the clas-sification results in water <15 m determined that the overall accuracy was 92.1% with a Kappa statistic of 0.81 (p<0.00001).

Figure 6: The results of the classification to Level 1 of flight line one into vegetated pixels and sand pixels. The analysis results are only shown for water depths <15 m.

The image data was then classified to Level 2 to separate the vegetated pixels into those domi-nated by seagrass and those dominated by mixed algae (Figure 7). Classification to Level 2 was carried out using the corrected bottom reflectance data, with the pixel classified as bare sand ex-


cluded from the analysis. The data was classified using the spectral angle to determine the dis-tance of each pixel from the reference signatures which represented the two desired classes. The spectral angle was calculated using the reflectance values for only five wavelengths (469.3, 485.2, 500.1, 515.0 and 576.3 nm). Preliminary spectral separation analysis determined that this combi-nation provided the greatest separation between the two classes. The results showed that 62.6% of the 526,941 vegetated pixels are dominated by seagrass and 37.4% are dominated by mixed algae. The overall accuracy of the classification to Level 2 was 87.39% with a Kappa statistic of 0.81 (p<0.00001).

Figure 7: The results of the classification to Level 2 of flight line one which separates the vegetated regions into those dominated by seagrass and those dominated by mixed algae. The analysis re-sults are only shown for water depths <15 m.

Classification of the data to Level 3 separated the mixed algae category from Level 2 into low growing turfing algae and canopy forming brown algae (Figure 8). These two habitats exhibit dis-tinctly different spectral signatures and relate to a naturally observable distinction in the marine environment at Rottnest Island. Preliminary results from spectral separation analysis, using the genetic algorithm, determined that four wavelengths provided the best separation between these to habitat component types, those being 469.3, 485.2, 500.1 and 561.4 nm. The classification was produced using the spectral angle to determine the distance between each pixel and the reference spectra for each class and was calculated using only the four wavelengths. The results determined that 19.5% of the mixed algae habitat was dominated by the canopy forming brown algae, such as Ecklonia radiata, and the remaining 80.5% was turfing algae. The overall accuracy of the classifi-cation was 70.35%, with a Kappa statistic of 0.59 (p<0.00001).

Figure 8: The results of the classification to Level 3 of flight line one which separates the mixed algae regions into those dominated by canopy algae and those dominated by turf. The analysis results are only shown for water depths <15 m.


Figure 9: The typical reflectance spectra of the canopy algae and turf classes used for classifica-tion to Level 3.

CONCLUSIONS

The preliminary results of this study have shown very promising results for the application of hy-perspectral data for the classification of shallow marine benthic habitats in a temperate coastal environment. The overall classification accuracies obtained at Level 1 and 2 are very encouraging and provide a good basis for further separating these classes into their more specific components. Although the overall accuracy of the Level 3 classification was lower, it demonstrates the ability to identify the dominant components of these pixels at this level, using only basic spectral analysis. The work presented in this paper is the preliminary results of the spectral analysis and habitat classification of this data. The results discussed in this paper were obtained using HyMap data that was corrected for the influence of the water column using no target spectra from the location to fine tune the MIP model and the classifications are based on the most basic of the spectral separation techniques being developed. Previous work using the MIP correction process has indicated that the accuracy of the model outputs can be increased by fine tuning using the most representative spectra of the dominant bottom types in the image.

The hierarchal approach used to subset the data is similar to other proposed by Hochburg et al. (39) for the classification of coral reef environments. By taking this approach we were able to seg-ment the image, at each increase in classification level, to enable the use of only library spectra of habitat components that are typically found in those broad classes. This has the two pronged effect of both reducing the processing time of each step and reducing the possibility of misidentification of pixels with habitat components not likely to be dominant in the pixel. This is an important feature of the analysis when being applied to data from areas such as Rottnest Island where the marine habitats don’t often occur as homogenous patches at scales that will cover whole pixels. This means that with the exception of some bare sand areas there are very few pixels with only one habitat component present. This is especially true when the mixed algae class is considered with and excess of 300 species of macroalgae being found around the island. The seagrass habitats provide more homogenous habitat patches, but finer scale analysis to a genera level is hampered by the very minor spectral differences between the two groups at HyMap resolution. As a result of the typically heterogeneous nature of the shallow marine benthic habitats the classification of pix-els based on their dominant component was preferable to attempting to classify using an approach such a linear spectral un-mixing. The reason for this is that this method requires the use of repre-sentative spectra for all possible end members, which is in most cases an unrealistic goal.

The preliminary results of the spectral separation analysis are similar to other studies that have tested the discrimination between different benthic substrates reflectance signatures, namely, that the best discrimination is often obtained using a subset of the available wavelengths in higher resolution sensors (40). So far, only limited results from the spectral analysis are completed and validated, but with the implementation of a greater range of spectral metrics, a greater number of input library spectra and the use of derivative spectra we believe that the separation between the classes can be improved and classification to a higher level achieved.


There are a number of issues as to how to accurately validate the classification results using ground validation points. The first issue is the inherent uncertainty in the accuracy of the locations recorded for the validation point themselves and the second, being the uncertainty in the accuracy of the geo-rectification of the remotely sensed imagery, which when working with 3.5 m resolution data can have significant ramifications. This issue was addressed by taking the estimated posi-tional accuracy of the validation points into account and assuming that a pixels location could be one pixel out in any direction when assessing the accuracy of the classification results. This ap-proach is suitable when assessing the accuracy of the lower level classifications, where pixels generally occur as large patches, but at higher level classifications the results become very pixe-lated, and the interpretation of ground validation results can provide ambiguous results. The an-swer to this may require a finer scale approach to ground validation at a smaller scale than the whole image.

Overall, the approach taken in this study, to correct the hyperspectral data using only inputs de-rived from the data itself and to classify it using a hierarchical approach based on spectral metrics as a discrimination measure has proved to be successful. It also has scope for further refinement to improve the result and provide classifications to finer scales.

ACKNOWLEDGEMENTS We would like to acknowledge the financial support provided by Murdoch University and Rottnest Island Authority towards the acquisition of data and completion of the project. We would also like to thank the Department of Environment and Conservation, Western Australia for financial support towards completing fieldwork. We would also like to acknowledge to assistance provided by EOMAP and HyVista Corporation towards additional data processing and advice.

REFERENCES

1. Commonwealth of Australia, 1998. Australia's Oceans Policy, Volume 1 (Environment Austra-lia): 49pp.

2. ANZECC, 1998. Guidelines for establishing the national representative system of marine pro-tected areas (Environment Australia): 15pp.

3. Stevens T, 2002. Rigor and representativeness in marine protected area design. Coastal Management, 30: 237-248

4. Margules, C R, R L Pressey & P H Williams, 2002. Representing biodiversity: data procedures for identifying priority areas for conservation. Journal of Bioscience, 27: 309-326

5. Gray J S, 1997. Marine biodiversity: patterns, threats and conservation needs. Biodiversity and Conservation, 6: 153-175

6. Sarkar S & C R Margules, 2002. Operationalizing biodiversity for conservation planning. Jour-nal of Bioscience, 27: 299-308

7. Faith D P et al., 2001. Practical application of biodiversity surrogates and percentage targets for conservation in Papua New Guinea. Pacific Conservation Biology, 6: 289-303

8. Stevens T & R M Connolly, 2004. Testing the utility of abiotic surrogates for marine habitat mapping at scales relevant to management. Biological Conservation, 119: 351-362

9. Ward T J et al., 1999. Selecting marine reserves using habitats and species assembleges as surrogates for biological diversity. Ecological Applications, 9: 691-698

10. Banks S A & G A Skilleter, 2002. Mapping intertidal habitats and an evaluation of their conser-vation status in Queensland, Australia. Ocean and Coastal Management, 45: 485-509


11. Vanderklift M A, T J Ward & J C Phillips, 1998. Use of assemblages derived from different taxonomic levels to select areas for conserving marine biodiversity. Biological Conservation, 86: 307-315

12. O'Hara T D, 2001. Consistency of faunal and floral assemblages within temperate subtidal rocky reef habitats. Journal of Marine and Freshwater Research, 52: 853-863

13. Mumby P J et al., 1997. Coral reef habitat mapping: how much detail can remote sensing pro-vide? Marine Biology, 130: 193-202

14. Sotheran I S, R L Foster-Smith & J Davies, 1997. Mapping of marine benthic habitats using image processing techniques within a raster-based geographic information system. Estuarine, Coastal and Shelf Science. 44(Supplement A): 25-31

15. Louchard EM et al., 2003. Optical remote sensing of benthic habitats and bathymetry in coastal environments at Lee Stoocking Island, Bahamas: A comparative spectral classification approach. Limnology and Oceanography, 48: 511-521

16. Werdell P J & C S Roesler, 2003. Remote assessment of benthic substrate composition in shallow waters using multispectral reflectance. Limnology and Oceanography, 48: 557-567

17. Armstrong R A, 1993. Remote sensing of submerged vegetation canopies for biomass estima-tion. International Journal of Remote Sensing, 14: 621-627

18. Dustan P, E Dobson & G Nelson, 2001. Landsat Thematic Mapper: detection of shifts in com-munity composition of coral reefs. Conservation Biology, 15: 892-902

19. Naseer A & B G Hatcher, 2004. Inventory of the Maldives' coral reefs using morphometrics generated from Landsat ETM+ imagery. Coral Reefs, 23: 161-168

20. Purkis S J & R Pasterkamp, 2004. Integrating in situ reef-top reflectance spectra with Landsat TM imagery to aid shallow-tropical benthic habitat mapping. Coral Reefs, 23: 5-20

21. Pearce A, 1991. Eastern boundary currents of the southern hemisphere. Journal of the Royal Society of Western Australia, 74: 35-45

22. Hanson C E, C B Pattiaratchi & A M Waite, 2005. Seasonal production regimes off south-western Australia: influence of the Capes and Leeuwin Currents on phytoplankton dynamics. Marine and Freshwater Research, 56: 1011-1026

23. Wells F & D I Walker, Introduction to the marine environment of Rottnest Island, Western Aus-tralia, in The Marine Flora and Fauna of Rottnest Island, Western Australia, F. Wells, et al., Editors. 1993 (Western Australian Museum, Perth) p. 1-10.

24. Woo M, C B Pattiaratchi & W Schroeder, 2006. Summer surface circulation along the gas-coyne continential shelf, Western Australia. Continental Shelf Research, 26: 132-152

25. Cresswell G R, 1996. The Leeuwin Current near Rottnest Island, Western Australia. Journal of Marine and Freshwater Research, 47: 483-487

26. Cresswell G R & J L Peterson, 1993. The Leeuwin Current south of Western Australia. Journal of Marine and Freshwater Research, 44: 285-303

27. Heege T & J Fischer, 2004. Mapping of water constituents in Lake Constance using multispec-tral airbourne scanner data and a physically based processing scheme. Canadian Journal of Remote Sensing, 30: 77-86

28. Pinnel N, 2007. A method for mapping submerged macrophytes in lakes using hyperspectral remote sensing. PhD Dissertation, Technical University of Munich. 164pp.

29. Robila S A, 2005. Using spectral distances for speedup in hyperspectral image processing. International Journal of Remote Sensing, 26: 5629-5650


30. Du Y et al., 2003. A new hyperspectral discrimination measure for spectral similarity Algo-rithms and technologies for mutlispectral, hyperspectral and ultraspectral imagery IX. Proceed-ings of SPIE. 5093: 430-439

31. Du Y et al., 2004. New hyperspectral discrimination measure for spectral characteriztion. Opti-cal Engineering, 43: 1777-1786

32. Robila S A & A Gershman, 2005. Spectral matching accuracy in processing hyperspectral data. In: International Symposium on Signals, Circuits and Systems.

33. Landgrebe D, 2003. Signal theory methods in multispectral remote sensing (Wiley-Interscience) 508 pp.

34. Karpouzli E, T J Malthus & C J Place, 2004. Hyperspectral discrimination of coral reef benthic communities in the western Caribbean. Coral Reefs, 23: 141-151

35. Minnesota Department of Natural Resources, 2007. DNR Garmin extension for ArcView 3.x (Minnesota Department of Natural Resources)

36. Jenness J S, 2001. Radiating lines and points extension for ArcView 3.x (Jenness Enterprises)

37. Jenness J S & J J Wynne, 2006. Kappa analysis extension for ArcView 3.x (Jenness Enter-prises)

38. Congalton R G, 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 37: 35-46

39. Hochberg E J, M J Atkinson & S Andrefouet, 2003. Spectral reflectance of coral reef bottom-types worldwide and implications for coral reef remote sensing. Remote Sensing of Environ-ment, 85: 159-173

40. Hochberg E J & M J Atkinson, 2003. Capabilities of remote sensors to classify coral, algae, and sand as pure and mixed spectra. Remote Sensing of Environment, 85: 174-189

MAPPING THE SHALLOW MARINE BENTHIC HABITATS OF ROTTNEST ISLAND… · Rottnest Island, Western...

Documents

Transcript of MAPPING THE SHALLOW MARINE BENTHIC HABITATS OF ROTTNEST ISLAND… · Rottnest Island, Western...