in coastal regions and in small islands
Coastal management sourcebooks
Part 3 Habitat Classification and Mapping
|11||Mapping Coral Reefs and Macroalgae Part A|
Summary The applications of remote sensing for coral reef management include baseline cartographic mapping, mapping reef geomorphology and mapping reef habitats. Generally, more detailed mapping objectives require more sophisticated remote sensing techniques. Most satellite-mounted sensors are able to provide information on reef geomorphology and limited ecological information such as the location of corals, sand, algal and seagrass habitats with an accuracy ranging from 50–70%. The most cost-effective satellite sensors for habitat mapping are Landsat TM (for areas greater than one SPOT scene) and SPOT XS for areas within a single SPOT scene (i.e. < 60 km in any direction). Colour aerial photography can resolve slightly more detailed ecological information on reef habitats but, for general purpose mapping, satellite imagery is more effective (i.e. has slightly greater accuracy and uses up much less staff time). At low altitude, infra-red aerial photography can be used to estimate live-coral cover over shallow (< 1 m deep) reef flats. However, since the low altitude restricts the areal coverage of each photograph, this method is only appropriate for small areas.
The most accurate means of making detailed reef habitat maps appears to be use of airborne multi-spectral instruments such as CASI (Compact Airborne Spectrographic Imager). In the Caribbean, CASI can map assemblages of benthic species and substrata with an accuracy of > 80% (cf. < 37% for satellite sensors and 57% for colour aerial photography).
A wide variety of field survey techniques are available for assessing coral and macroalgal habitats. For broadscale habitat mapping (e.g. with satellite imagery), manta tow and visual belt transects are recommended and described. Rapid quantitative methods such as quadrats ,line intercept transects and underwater video are recommended for imagery with greater resolution like CASI.
Coral reefs generally exist in clear waters, which are highly suited to optical remote sensing. Their nutritional biology dictates that they usually live in regions of relatively clear water which transmit sufficient solar irradiance to allow their endosymbiotic algae (zooxanthellae) to photosynthesize (for overview of coral biology see Dubinsky 1990). This requirement has been capitalised upon by the remote sensing community and satellite investigations of coral reef structure date back to 1975, shortly after the launch of Landsat MSS (Smith et al. 1975).
Remote sensing has been touted to provide information on several parameters that are of importance to reef management. From a remote sensing point of view, coral reef boundaries are probably the easiest of these to map. This information may be used for routine planning requirements and locating the boundaries of management zoning schemes (e.g. Kenchington and Claasen 1988). The next level of sophistication is to distinguish principal geomorphological zones of the reef (e.g. reef flat , reef crest, spur and groove zone). For management purposes, such information may be used in a similar manner to that of ‘coral reef maps’. These maps may also have more sophisticated uses, which include the stratification of field sampling regimes and relating reef geomorphology to reef function and productivity (discussed by Fagerstrom 1987).
Maps of reef geomorphology have been made with considerable success (see forthcoming sections). However, mapping the ecological components of coral reefs appears to be considerably more difficult using conventional satellite-borne imagery (Bainbridge and Reichelt 1989). Ecological components may be defined in several ways, such as assemblages of coral species (e.g. Done 1983), assemblages of major reef-dwelling organisms, or assemblages of species and substrata (e.g. Sheppard et al. 1995). The choice depends on the ecological objective and type of reef. For example, coral species assemblages would be appropriate in places where coral cover was high (see Plate 8), but perhaps less appropriate where coral cover rarely exceeded (say) 20%. Irrespective of their foundation, maps with an ecological basis will be referred to here as ‘habitat maps’. Habitat maps have considerable potential in the management of reef biodiversity at ecosystem, community and (indirectly) at species scales. This has often been sought by protecting representative areas of habitat (McNeill 1994).
Perhaps the ‘holy grail’ of all coral reef remote sensing is the determination of live-coral cover. Whilst the issue of coral reef health is complex (see UNESCO 1995), one of the most widely accepted parameters to consider is live coral cover with a view to relating changes in cover to the processes responsible. As we shall see, however, optical remote sensing is severely limited at measuring this parameter and has only been successful in shallow reef flat environments.
Technically speaking, the term ‘macroalgae’ describes algae that are large enough to see by eye rather than a taxonomic grouping. Most macro algae can be categorised by life form into: (i) fleshy algae, (ii) calcareous algae (which have a calcium carbonate skeleton), (iii) turf algae (which form a mat < 1 cm high on the substratum), and (iv) crustose coralline algae (which form flat plates on the substratum and may cement coral fragments). Macroalgal species are divided among three large groups (green, brown and red algae), which are named according to the colour of their dominant photosynthetic and accessory pigments (Plate 8). a
Green algae (Chlorophyta) contain chlorophyll and are well represented in the tropics (Littler et al. 1989). Some of these algae (e.g. Enteromorpha spp.) favour stressful environments where nutrients are high and herbivory low. Others are calcified (e.g. Halimeda spp.) and contribute heavily to the sandy sediments of reef areas. Brown algae (Phaeophyta) predominantly contain the brown pigment, fucoxanthin, but colonies may range in colour from beige to almost black. In temperate seas, brown algae can form vast kelp forests (e.g. Macrocystis sp. in California) but their abundance and diversity in tropical seas is reduced. Some of the most common tropical genera include Sargassum and Turbinaria which are often associated with reef flats, and Lobophora which is fairly ubiquitous. The red algae (Rhodophyta) are the largest and most diverse group but arguably the least well understood (Littler et al. 1989). They contain large quantities of the pigment phycoerythrin, which can often resemble the pigmentation of brown algae (fucoxanthin). Red algae are extremely important reef-building organisms, which may form reef crests (e.g. Lithophyllum spp.) and large calcareous plates (Sporolithon spp.).
From a coastal management perspective, it would be useful to map algal-dominated habitats for biodiversity management and locating mariculture sites for harvesting commercial seaweeds such as Porphyra spp. and Euchema spp. Maps of algal biomass may also be useful in understanding coastal productivity and in managing algal mariculture. Algal habitats may also be correlated to the abundance of commercially significant gastropods such as the queen conch, Strombus gigas. While this remains to be proved, the topic of correlating marine resources to habitats is discussed in Chapter 18.
Structure of this chapter
This chapter sets out methods and anticipated outcomes of mapping coral reefs and macroalgae. It begins by evaluating the likely outcomes of using various satellite and airborne methods for reef and macroalgal assessment. This section is split into a general overview based on the literature and then a more detailed discussion of habitat mapping which draws mainly on results from the Turks and Caicos Islands. It also discusses the usefulness of various image processing methods. In several places, the accuracies of habitat maps are discussed with reference to other types of habitat such as seagrass and sand. This is because mapping of coral reefs and macroalgae is not an isolated process and is affected by the presence of neighbouring habitats.
The latter part of the chapter describes the field survey techniques used to describe coral and macroalgal systems. This is a vast subject area which is covered in several comprehensive texts. As such, the scope of this chapter is to give some general guidance and an overview of relevant literature.
Overview of descriptive resolution and sensors
The majority of studies in Table 11.1 provided geographic information on reef zones. These zones are primarily of a geomorphological, rather than an ecological nature and include fore reef, reef crest, algal rim, spur and groove, and so on. Any attempt to make detailed comparisons between the ability of different sensors is difficult because of the differences in reef terminology, study sites and objectives employed in each study. Generally speaking however, sensors with higher resolution will offer greater detail. A recent study of the reef zonation at Heron Island was conducted using Landsat TM (Ahmad and Neil 1994). The authors compared their results to those obtained by Jupp et al. (1986) who originally used Landsat MSS to map the island. TM revealed a similar number of principal reef zones to the MSS, but allowed the principal zones to be divided into a greater number of subzones. Reef zones we re mapped more precisely with the TM.
|Table 11.1 The descriptive resolution of various remote sensing methods for mapping coral reefs. Numbers denote referenced papers. Updated from Green et al. (1996).|
|Type of class||Landsat MSS||Landsat TM||SPOT XS||SPOT Pan||Merged TM/Pan||Airborne MSS||API|
|Coral reef (general)||30||22, 25, 30, 37,||37, 30||30||28, 30||29||14, 41|
|Reef zones (geomorphology)||2, 3, 4, 5, 13, 16, 17, 18, 19, 32||12, 37||2, 6, 11, 21, 37, 42||26||10,
24, 27, 33, 34
|Coral density||1, 39, 40, 43, 44||7, 8, 9|
|Coral reef habitats||43, 44||29||31|
|Coral colonies||10, 20, 36|
|35, 36, 38|
A few studies have described coral cover in terms of density (e.g. Zainal et al. 1993, Ahmad and Neil 1994). Measurements of density of this type are usually qualitative and are generally not supported by field data. For example, Ahmad and Neil (1994) include reef subzones described as, ‘medium coral head density’, ‘low coral head density’ and ‘high coral head density’. The vagueness of these terms reflects practical difficulties experienced in assigning numeric quantitative ranges to coral head density (which may require considerable fieldwork). This is not a criticism of the authors; it merely reflects a general limitation of remote sensing to provide information on reef structure and composition beyond a geomorphological level. This issue will be returned to later when discussing the overall usefulness of remote sensing products in an ecological context. A notable exception to this generalisation is seen in Bour et al. (1996). Coral density in New Caledonian reefs was categorised into low density (10–29% cover), medium density (30–70% cover) and high density (71–100% cover). These density bands were mapped using SPOT XS and the authors point out that these results may be useful for long-term monitoring of coral reefs.
Sheppard et al. (1995) achieved an overall map accuracy of 57% when using 1:10,000 colour aerial photography to map several coral reef, algal and seagrass habitats of Anguilla. In a study of similar habitats in the Turks and Caicos Islands, Mumby et al. (1998b) achieved greater accuracy (81%) using CASI. Satellite imagery of the same study site was found to be generally unsuccessful for mapping individual reefal habitats (accuracy < 37%). This synopsis of habitat mapping will be discussed in greater detail in the following section.
To date, the only successful remote sensing method for providing synoptic data on coral colonies and live coral cover is the use of low altitude aerial photography (see Catt and Hopley 1988, Thamrongnawasawat and Catt 1994). An altitude of 3000 feet (914 m) is generally considered optimal (Hopley and Catt 1989). Healthy corals strongly reflect near infra-red electromagnetic radiation. Therefore, these wavelengths are able to discriminate coral growth from coralline, turf and macro -algae. However, the major limitation of the technique is the poor depth of penetration (< 1 m) and therefore its use is largely confined to assessing the coral cover of emergent and extremely shallow reef flats at low tide (Hopley and Catt 1989). True colour photography will penetrate to approximately 15–20 m but does not offer such good discriminating power. The greatest resolution reported for low altitude aerial photography of reef flats discerned colonies of live massive and branching corals, dead corals and sand (Thamrongnawasawat and Hopley 1995). CASI is capable of sub-metre pixel sizes but the signal to noise ratio is likely to be low, especially for deeper reefs or habitats with low reflectance. The use of CASI for measuring shallow-water live coral cover is yet to be evaluated.
Coral species discrimination does not appear to be possible (Table 11.1) although this is barely surprising considering the spatial resolution of most sensors and the arguments made above.
Relatively little has been written explicitly about macroalgal remote sensing. In temperate waters, large phaeophytes have been mapped using a variety of sensors including Landsat TM for drifting Sargassum spp. (Tiefang et al. 1990) and CASI for regions of Ecklonia spp. in Perth, Australia (Jernakoff and Hick 1994). For tropical reef flats, aerial photography has been used to map Sargassum spp., Turbinaria spp. and other shallow-water algal assemblages (Manieri and Jaubert 1984, Catt and Hopley 1988, Mumby et al. 1995a). Zainal (1994) successfully used Landsat TM and a digital bathymetry model to map algal beds around Bahrain. Using this approach , Zainal achieved user accuracies of 86% for Sargassum spp. and 75% for algal turf. These accuracies were considerably higher than those reported by Sheppard et al. (1995) for mapping phaeophytes in Anguilla with aerial photography (38%). They were also greater than satellite-based results from Mumby et al. (1997) who mapped beds of the brown alga, Lobophora va riegata, with a user accuracy of 0% (SPOT Pan) to 41% (Landsat TM and SPOT XS). The reasons for this disparity are unclear but may be a consequence of the detail to which habitats we re defined in each study (studies by Sheppard et al. and Mumby et al. included a greater number of habitat classes and consequently tended to be less accurate). CASI was found to map this habitat with a user accuracy of 82% (Mumby et al. 1998b). Again, these results will be described further later.
With regard to assessing macroalgal biomass or the coverage of commercially important species, the literature seems extremely sparse. Meulstee et al. (1986) found an empirical relationship between the biomass of temperate, inter-tidal macrophytes and the density of colour in aerial photographs (error 10%). However, we are unaware of any further examples.
Descriptive resolution for mapping coral and macroalgal habitats
This section compares the ability of different satellite and airborne remote sensing methods for mapping reefal habitats. Image processing methods are assessed first using a case-study from the Turks and Caicos Islands. The descriptive resolution of each sensor is then assessed by extending the case-study from the TCI and describing a second case-study from the Caribbean which used 1:10,000 colour aerial photography to map the coastal zone of Anguilla (Sheppard et al. 1995).
Image processing methods
Creation of habitat maps from digital remote sensing usually involves either supervised or unsupervised classification of spectral bands (see Chapter 10). The addition of water column correction and contextual editing should improve map accuracies for submerged habitats (see Chapters 8 and 10 for implementation details). Surprisingly, water column correction has not been widely adopted as a pre-processing step for the remote sensing of tropical coastal waters which often satisfy the exponential attenuation requirements of the model. Reviewing relevant papers, Green et al. (1996) found only four studies out of forty-five (9%) had attempted water column correction. From a practical perspective, this prompted the question: What benefits, in terms of thematic map accuracy, accrue from water column correction and is the extra processing time cost-effective? A similar argument may be made for contextual editing. These issues are discussed below.
The following evaluation of these image processing techniques is derived from our case study in the Turks and Caicos Islands (Mumby et al. 1998a). Three hierarchical levels of habitat discrimination are given which include coral reef, algal and seagrass classes. For satellite comparisons, the fine, intermediate and coarse levels include 13, 8 and 4 habitats respectively. CASI imagery was confined to a smaller area so the fine, intermediate and coarse descriptive resolutions include 9, 6 and 4 habitats (see Table 11.2). Four levels of image processing were assessed and these are defined in Table 11.3.
|Table 11.2 Description and characteristics of benthic habitats determined from hierarchical classification of field data (except for Acropora palmata which was added later), showing mean percentage cover, densities and standing crop where appropriate. Class assignment is described for three levels of habitat discrimination: coarse (C), intermediate (I) and fine (F). Habitats present in CASI imagery are marked with an asterisk.|
|Table 11.3 Definition of image processing steps for the Turks and Caicos Islands case-study using satellite and CASI data.|
This section uses two measures of accuracy – overall accuracy and the Tau coefficient (see Chapter 4).
Supervised multispectral classification of reflectance data
Habitat map accuracies were lowest for this level of image processing (Figure 11.1a).
Figure 11.1 Effect of image processing methods on the accuracy of coral reef habitat maps. Results are shown for various satellite sensors and the Compact Airborne Spectrographic Imager, CASI. Overall accuracies and the Tau coefficient (with upper 95% confidence interval) are displayed for three levels of descriptive resolution (4, 8 and 13 habitat classes for the satellite imagery, and 4, 6 and 9 habitat classes for the more geographically restricted CASI imagery). Data are presented for supervised multispectral classification (a), the addition of water column correction (b), the addition of contextual editing (c) and the addition of both water column correction and contextual editing (d).
Water column correction
Water column correction of CASI imagery made a significant (P < 0.01) improvement to maps with fine habitat discrimination (Figure 11.1b). Inter-habitat similarity was high (Bray-Curtis Similarity, 60–80%) at this descriptive resolution and variable depth exerted a strong effect on accuracy, thus requiring water column correction. At coarse descriptive resolutions (coral, algae, sand, sea-grass), the habitats were sufficiently dissimilar to one another (Bray-Curtis Similarity, 10–15%) that depth invariant processing was not essential for habitat mapping (although it did make a minor improvement and may be more advantageous in other areas).
Map accuracy was significantly improved for Landsat TM at coarse and intermediate descriptive resolutions. However, water column correction was not significantly beneficial for sensors which produced a single depth invariant band (SPOT XS and Landsat MSS). Supervised classification of a single band is limited because the statistical separation of habitat spectra is confined to one dimension. Therefore, whilst depth-invariant processing of SPOT XS and Landsat MSS data may have reduced the effects of variable depth, it did so at the expense of the number of spectral bands available to the classifier (i.e. dimensions of the discriminant function). Landsat TM and CASI were amenable to depth-invariant processing because three and six depth-invariant bands (respectively) were available for supervised classification. We suggest that where only a single depth-invariant bottom-index can be created, the benefit of accounting for variable depth is out-weighed by the need to input several bands to the classifier – even if the component bands exhibit depth effects. This conclusion is undoubtedly site specific and may not hold in areas where variation in bathymetry is much greater.
To ensure that contextual editing does not create bias or misleading improvements to map accuracy, the decision rules must be applicable throughout an image and not confined to the regions most familiar to the interpreter. The simplest of these for coral reefs is the presence or absence of coral and seagrass habitats on the forereef (Figure 11.2). A full list of contextual rules is given in Table 11.4. with reference to Figure 11.3 (Plate 12). Spectral confusion between these habitats was greatest in Landsat MSS and SPOT Pan, which had the poorest spatial and spectral resolutions respectively. Consequently, thematic maps from these sensors showed the greatest improvements in accuracy (Figure 11.1c). The capability of Landsat TM was improved for coarse-level habitat mapping but more detailed maps were unaffected. Accuracies derived from SPOT XS were not significantly affected by contextual editing although small gains were found (3–4%). Similarly, overall accuracies of CASI images were not significantly affected by contextual editing.
|Table 11.4 Contextual rules used during post-classification editing. Misclassified habitats were re-coded to the habitats shown. Locations are illustrated in Figure 11.3 (Plate 12).|
Water column correction and contextual editing
When combined, depth compensation and contextual editing made a significant improvement upon basic image processing (Figure 11.1d and 11.1a). The combined approach was collectively more accurate than the singular implementation of water column correction or contextual editing (although the improvement was not always significant). The relative importance of depth compensation and contextual editing varied between sensors. Habitat maps from Landsat TM and CASI benefited from depth-invariant processing whereas Landsat MSS was more amenable to contextual editing. Accuracies from SPOT XS showed the greatest improvement when depth compensation and contextual editing were used together. This resulted from a change in inter-habitat misclassification, which was brought about by water column correction. Prior to this processing step, sand was a major source of confusion between habitats. Water column correction, which was optimised for sand habitats, improved the mapping of sand but confusion between coral and seagrass spectra increased (presumably because of the dependency on a single depth-invariant band). The change of classification error to corals/seagrass permitted subsequent contextual editing to make a significant improvement to accuracy.
Refinements to basic image processing may be considered cost-effective if the additional investments in time are justified by improved map accuracy. Since thematic map accuracy is finite, continued processing will inevitably reach a stage where the ‘accuracy pay-off’ declines and processing becomes progressively ineffective. The daily accruement of accuracy from water column correction and contextual editing remains high (Table 11.5) and on the premise set out above, it seems that both processing steps are cost-effective.
|Table 11.5 Summary of processing steps for habitat mapping showing the total implementation time, the % gain in accuracy per day’s processing effort and the methods which significantly improve upon basic image processing (denoted ). % accuracy gains are averaged for all satellite images and expressed for coarse (C), intermediate (I) and fine (F) descriptive resolution. Note: the 3 days required to conduct multispectral classification has been added to each processing step - i.e. water column correction took an additional 1/2 day to implement so the total processing time would be 3.5 days. Figures in the table assume that the user is familiar with the processing methods and does not have to learn them from scratch (if that is the case, refer to processing times in Chapter 19).|
Other image processing methods
We have discussed the use of water column correction and contextual editing, both of which improve the classification accuracy of mapping coral and macroalgal habitats. There are at least two other processing methods which have been employed in conjunction with reef mapping.
Bour and co-authors have described two alternative methods for mapping reefs in New Caledonia with SPOT XS imagery. Both involve plotting bands 1 and 2 against one another to examine the spread of data. Ordinarily, the bands are quite highly correlated which limits their potential for spectrally discriminating habitats. The first method de-correlates the bands by rotating their axes until they align with the main variability in the data (de Vel and Bour 1990). A texture layer was also created and entered into the supervised classifier with the rotated bands. The authors found that the approach improved on standard supervised classification of SPOT data but the accuracies were not compared.
The second method resulted in the creation of ‘coral index’ for mapping coral density (Bour et al. 1996). Regions of the original bi-plot of SPOT bands are related to features on the reef by reference to known points on the imagery. Conceptually, this process is the reverse of running a supervised classification on a 2-band image and then looking at the signatures plotted in 2-band feature space (which is usually done statistically by the classification algorithm, see Chapter 10). Bour et al. (1996) found that an ellipse may be drawn on this bi-plot which defines coral reefs of varying density. A mathematical transformation is then carried out to align band 2 with the direction of the ellipse. This stretches the range of values for coral density along a single axis which becomes a coral index. The range of values are then divided into three sections which constitute sparse, medium and dense corals.
Conceptually, this process is similar to water column correction. Rather than creating a depth-invariant index, the transformation aligns axes to a gradient in coral density. However, Bour and co-authors’ method does not take account of variable depth and has been found to be of limited use in anything but shallow water (Peddle et al. 1995). In addition, the degree to which a ‘coral index’ improves thematic map accuracy is yet to be demonstrated.
Recently, Peddle et al. (1995) have attempted to apply the method of Spectral Mixture Analysis (SMA) to coral reefs in Fiji. The method was originally developed for terrestrial systems and aims to address sub-pixel variability (mixels). The seascape is divided into a few distinct components – in this case, deep water and coral reef. Spectral characteristics of pure deep water and homogeneous coral reef are then described in each of several spectral bands. The radiance of each pixel in the image is then assumed to be divisible into its constituent fractions i.e. the amounts contributed by deep water and coral. In this case, Peddle et al. (1995) used extensive field data to translate the radiance at a given depth into its constituent ‘deep water likeness’ and ‘coral likeness’. Once the model is developed, coral area can be predicted for each pixel provided that the depth is known. As such, the method is limited to areas where good bathymetry data exist. Another possible problem is the use of a ‘pure’ coral signature that is derived for shallow (or emergent) coral. For straight forward biological considerations, such as photo-adaptation and species zonation with depth, it does not necessarily follow that shallow-water coral reefs will resemble those of deep water. However, the authors concede that the method is still under development and that it will need to incorporate a greater number of component habitats (e.g. sand, algae) to realise its potential.
The descriptive resolution of satellite imagery
This section assumes that water column correction is conducted on multispectral imagery and that contextual editing is conducted on all images. Even though the focus of this chapter is coral reefs and macroalgae, seagrass habitats will also be mentioned. This is because maps of coral reefs and algae are often intrinsically linked to seagrass (e.g. due to misclassification errors and contextual editing). Seagrass habitats are treated fully in the following chapter.
Coarse descriptive resolution
A pronounced and usually significant drop in accuracy was found consistently between coarse, intermediate and fine habitat discrimination. For mapping at coarse descriptive resolution (i.e. four habitat classes; sand, coral, macroalgae, seagrass), Landsat TM was significantly more accurate than other satellite sensors (overall accuracy 73%; Figure 11.4, Plate 12). SPOT XS also achieved a relatively high overall accuracy (67%). The accuracy of merged Landsat TM/SPOT Pan was not significantly different from SPOT XS (Z = 1.82, P > 0.05). The maps derived from SPOT Pan and Landsat MSS had an accuracy of < 60% (Figure 11.5).
Coral and sand habitats were generally more accurately distinguished than macro algal and seagrass habitats (Table 11.6). This is because macroalgal and seagrass habitats were spectrally and spatially confused with one another, which is not unusual (e.g. Kirkman et al. 1988) and has several causes. Whilst the photosynthetic pigments in seagrass and algae (e.g. chlorophyll, phycoerythrin and fucoxanthin) have different albedo characteristics, satellite spectral bands are generally unsuitable for distinguishing them (see Maritorena et al. 1994). This is because the range of wavelengths that penetrate water is small (< 580 nm) and may not encompass the characteristic reflectance minima and maxima for a particular pigment. For example, whilst most photo-synthetic pigments show reflectance minima below 450 nm, the maxima (‘red edges’) lie between 670 – 700 nm, which is beyond the range of water-penetrating irradiance. Where distinguishing minima and maxima exist within the water-penetrating spectrum, satellite bands may be too broad to distinguish them. For example, the reflectance minima for both green and brown algae are below 500 nm and their maxima are 550 nm and 575 nm respectively (Maritorena et al. 1994). SPOT XS band1 cannot differentiate these maxima because it detects radiance within the range 500–590 nm. Landsat TM can, in principle, distinguish these maxima because band 1 is sensitive to 450–520 nm and band 2 detects within the range 520–600 nm. However, given sensor noise and light attenuation problems, precise discrimination is unlikely.
|Table 11.6 User accuracies of habitat classes for all sensors and at two levels of descriptive resolution (coarse - habitats in bold – and fine). Note: details of habitat types do not apply directly to API categories that were described by Sheppard et al. (1995); however, categories of Sheppard and co-authors are broadly analogous to those described in the table. The most accurate satellite sensors for each habitat are shaded to facilitate comparison with airborne remote sensing.|
Accuracy of sensor (%)
|Living and dead stands of Acropora palmata||52||90|
|Microdictyon marinum (77%), Sargassum spp. (4%), medium soft coral density (5 m-2 ) and rubble (10%)||0||0||69||19||13|
|Bare substratum (40%),low soft coral density (3 m-2), Microdictyon marinum (30%), Lobophora variegata (12%)||32||44||57||54||32||81|
|Bare substratum (80%),medium soft coral density (5 m-2)||4||34||44||39||10||48||80|
|Bare substratum (60%),high soft coral density (8 m-2), Lobophora variegata (14%),high live coral cover (18%) of which ~ 9% is Montastraea spp.||33||18||36||51||47||66||83|
|Lobophora variegata (76%) and branching red/brown algae (9%)||13||41||31||41||0||38||82|
|Amphiroa spp. (40%), sand (30%),encrusting sponge (17%), sparse Thalassia testudinum and calcareous green algae||11||25||24||75||7|
|Sand and occasional branching red algae (< 6%)||11||45||64||46||50||73||75|
|Thalassia testudinum of low standing crop (5 g. m-2) and Batophora spp. (33%)||100||8||14||22||14|
|Thalassia testudinum of low standing crop (5 g. m-2) and sand||49||50||35||73||36|
|Medium–dense colonies of calcareous algae – principally Halimeda spp. (25 m-2); Thalassia testudinum of low standing crop (~10 g. m-2)||6||3||5||8||0|
|Dense colonies of calcareous algae – principally Penicillus spp. (55 m-2) and Halimeda spp. (100 m-2); Thalassia testudinum of medium standing crop (~80 g. m-2)||0||12||8||0||0||68||77|
|Thalassia testudinum and Syringodium filiforme of 5–80 g. m-2 standing crop||6||15||37||2||15||40||72|
|Thalassia testudinum and Syringodium filiforme of 80–280 g. m-2 standing crop||48||40||34||56||46||100||93|
Coral habitats also possess a high cover of macroalgae and the corals themselves contain photosynthetic pigment-bearing zooxanthellae. Whilst they may be spectrally confused with seagrass and some algal habitats, coral reefs may be spatially distinguished. This is because their location (context) within the reef landscape is usually confined to the seaward margin of the coastal zone (i.e. fringing reef) where wave exposure is moderate to high. In this study, contextual discrimination of seagrass and algal habitats was more difficult because gradients of exposure were generally less obvious and it was not so easy to predict the location of habitats.
Intermediate and fine descriptive resolution
Overall, the difference in accuracy between intermediate and fine descriptive resolution was considerably larger than the variation between sensors for either level of habitat discrimination. In practical terms, if the objective is to map more detail than coral reef, macroalgae, sand, seagrass, then the accuracy of a habitat map is more sensitive to the choice of descriptive resolution than the choice of sensor. It should be borne in mind, however, that overall accuracies for both higher descriptive resolutions were low; intermediate (8 habitats) 38–52%, fine (13 habitats) 21–37% and showed marked variability (Table 11.6). We conclude that satellite imagery is not well suited to detailed mapping of benthic habitats. This conclusion is in agreement with Bainbridge and Reichelt (1989) who concluded that satellite imagery is more appropriate for studying reef geomorphology than reef biology. The combined spatial and spectral resolutions of satellite sensors were not capable of reliably distinguishing many habitats that had a high inter-habitat similarity (Bray-Curtis Similarity, 60–80%). This was borne out by the high variability in accuracy associated with individual habitat classes. The poor separability of spectra rendered the supervised classification unable to assign pixels to appropriate habitat classes and resulted in large and variable allocation errors. It must be borne in mind, however, that mapping with fine descriptive resolution is an ambitious objective for any remote sensing method – some reef habitats look similar even to the field surveyor underwater!
Satellite imagery versus aerial photography
This section is separated into 1) a comparison of processed satellite imagery against aerial photography and 2) the visual interpretation of satellite imagery and aerial photography. The latter procedure is necessary if digital analysis of imagery is not possible (e.g. due to inadequate computing facilities). In this case, prints of an image can be used as a surrogate for aerial photography. Habitat maps may be created by tracing polygons on to an acetate overlay. Where field data are not available, delineation and identification of polygons is a subjective process which is guided by the colour, contrast, texture and context of areas in the imagery. Comparisons are made between the authors’ satellite data of the Caicos Bank and aerial photography of Anguilla which was interpreted by Sheppard et al. (1995). The habitats mapped in each area are moderately similar (Table 11.7), allowing broad comparisons to be made.
|Table 11.7 Habitat types from the Caicos Bank and Anguilla (fine descriptive resolution)|
Note: The accuracies reported for satellite imagery in this section are greater than those reported earlier because fewer habitat classes were used in the comparison with aerial photography.
1. Processed satellite imagery versus aerial photography
On the basis of the results from Sheppard et al. (1995), satellite sensors compared favourably to aerial photography for coarse and intermediate levels of habitat discrimination. Aerial photography was found to be inferior to Landsat TM and similar to merged Landsat TM/SPOT Pan, SPOT XS and SPOT Pan (Figure 11.6). This may appear surprising given the superior spatial resolution of aerial photography. However, digital satellite sensors have better spectral resolution and at coarse and intermediate descriptive resolutions, the habitats were sufficiently dissimilar to one another (Bray-Curtis Similarity, 10–15% and 30–50% respectively) that a crude discrimination of their spectra was possible. This conclusion did not apply to maps from Landsat MSS which were less accurate than those from aerial photography. For fine descriptive resolution, aerial photography was more accurate than all satellite sensors (Figure 11.6, Table 11.6).
Figure 11.6 Comparison of satellite sensors, aerial photography and airborne multispectral (CASI) data for mapping marine habitats at three levels of descriptive resolution using supervised classification. Data for aerial photography were recalculated from Sheppard et al. (1995). The upper part of each bar represents the overall accuracy, and the lower (solid) part, the Tau coefficient and its upper 95% confidence limit. Water column correction was conducted on multispectral imagery. Contextual editing was carried in all cases.
Note that accuracies for satellite sensors are higher than described elsewhere (Figures 11.1, 11.4 and 11.5). This results from the exclusion of rarer (and less accurately mapped) categories that were not present within the area mapped using CASI and thus could not be used in this comparison. Inclusion of these categories would have biased the comparison in favour of the airborne sensors.
2. Visual interpretation of satellite imagery versus aerial photography
Depth compensation was carried out for all satellite imagery except SPOT Pan which only has a single spectral band. The results of visual interpretation presented in Figure 11.7 reflect the visual identification of habitat types from satellite images rather than the delineation of habitat maps which can involve additional errors in boundary demarcation. As such, the comparison with aerial photographs is not entirely fair because the accuracies reported for aerial photographs incorporate these additional errors which result from habitat delineation. For further details on visual interpretation, refer to Chapter 10.
Figure 11.7 Comparison of satellite sensors and aerial photography for mapping marine habitats at three levels of descriptive resolution using visual interpretation. Data for aerial photography were recalculated from Sheppard et al. (1995). The upper part of each bar represents the overall accuracy, and the lower (solid) part, the Tau coefficient and its upper 95% confidence limit. Water column correction was conducted on multispectral imagery. Contextual editing was carried in all cases.
The existence of boundary errors notwithstanding, for coarse and intermediate levels of habitat discrimination, aerial photography (Sheppard et al. 1995) was found to be slightly poorer than all forms of satellite imagery and significantly less accurate than SPOT XS, which had the greatest visual contrast. By contrast, aerial photography was superior at mapping habitats with fine detail, in which case it was significantly better than all satellite image types.
In conclusion, if new imagery is to be acquired for visual interpretation and the choice includes aerial photography and satellite imagery, it is best to acquire a (processed) SPOT XS image for coarse and intermediate habitat mapping and aerial photography for fine habitat mapping.
CASI imagery versus satellite imagery and aerial photography
For all three levels of descriptive resolution, CASI imagery gave significantly more accurate results than satellite sensors and aerial photography (Figure 11.6, P < 0.001). The accuracy with which individual habitats we re mapped was more consistent than that found for aerial photography or satellite sensors (Table 11.6) and even fine habitat discrimination was possible with an accuracy of 81% (almost double that achieved with any satellite; Figure 11.8, Plate 13). CASI has the advantage of offering tremendous flexibility to the user. In this case, four narrow spectral bands were set to penetrate water, which increased the likelihood of distinguishing habitat spectra (band settings; 402.5–421.8 nm,453.4-469.2 nm, 531.1–543.5 nm, 571.9–584.3 nm). For other band settings refer to Table 5.10 of Chapter 5.
Whilst CASI was found to provide significantly greater accuracies than aerial photography, the comparison was not entirely balanced. Although the habitat categories were comparable, Sheppard et al. (1995) mapped a much larger area than that tested for CASI (14,600 ha versus 100 ha). Until carefully controlled comparisons can be conducted, it is perhaps safest to conclude that for comparable areas, CASI would be at least as good as aerial photography.
As a practical tool for mapping at high resolution, the relative merits of CASI and aerial photography require closer inspection. Whereas photographs offer greater spatial resolution than CASI, making use of this resolution is not straightforward. It is highly unlikely that a photo-interpreter would delineate features smaller than several metres because to do so would be too time consuming. This statement is borne out in the trace illustrated by Sheppard et al. (1995) in which the minimum polygon size was probably several metres or more. In contrast, polygon size does not constrain a digital classification of pixels. Thus, the spatial resolution of CASI may, in effect, be finer than the practical resolution of aerial photography. In addition, the delineation of habitats is likely to be both faster and more objective. It would be interesting to make an explicit evaluation of this issue – i.e. how does efficiency of each method vary with area covered? Furthermore, it would be useful to compare the effectiveness of digital remote sensing (i.e. CASI) and thematic classification of photographs which have been digitised using a scanner. Scanned aerial photographs have been used to great effect for mapping small areas (Thamrongnawasawat and Hopley 1995). A similar argument may be made for the new generation of digital cameras that are capable of taking images in a few broad spectral bands. Intuitively, however, CASI would be expected to fare better because of its greater spectral resolution (up to 21 spectral bands available to distinguish habitats).
Conclusions regarding coral and macroalgal habitat mapping
If new imagery is required for a site, the most cost-effective solution depends on the mapping objectives, required accuracy, size of the area, climate of the area (e.g. persistence of cloud cover), and the availability of technical expertise and equipment. These issues are discussed further in Part 5 but a few simple rules emerge here :
For maximal gain in accuracy per unit processing time, multispectral imagery should be subjected to water column correction and post-classification contextual editing.
For areas greater than 60 km in any direction (the size of a single SPOT scene), Landsat TM is likely to be the most cost-effective option. While being approximately US $1,700 per scene more expensive than SPOT XS, it covers nine times the area of SPOT and offers greater accuracy. If more than one SPOT scene is required to cover an area, SPOT will be more expensive than Landsat TM.
If image processing is not possible and maps must be made by visual interpretation, it is likely to be cheaper to purchase a pre-processed satellite image than mount a new campaign of aerial photography. This statement makes the assumption that imagery is optimised for visual interpretation (i.e. that the spectral contrast has been stretched and (preferably) that depth-invariant processing has been carried out).
CASI is significantly (P < 0.01) more accurate than satellite imagery for all levels of reef habitat mapping.
Satellite sensors cannot be used for mapping reefs with fine habitat discrimination if an accuracy exceeding 37% is required.