Can. J. Remote Sensing, Vol. 29, No. 2, pp. 187–200, 2003

A comparison of Landsat ETM+, SPOT HRV, Ikonos, ASTER, and airborne MASTER data for coral reef habitat mapping in South Pacific islands Patrick Capolsini, Serge Andréfouët, Cédric Rion, and Claude Payri Abstract. The performances of the Landsat-7 ETM+, ASTER, SPOT HRV, and Ikonos satellite sensors and the airborne MASTER (MODIS–ASTER simulator) were compared for coral reef habitat mapping in South Pacific reefs. This unique image data set provided different spatial resolution (4 m for Ikonos to 30 m for Landsat-7 ETM+), spectral resolution (two visible bands for SPOT-HRV to five visible bands for MASTER) and digitization (8–16 bits). We focused on two islands (Tahiti and Moorea, French Polynesia) with barrier and fringing structures representative of reefs of South Pacific volcanic islands. Five levels of benthic habitat complexity were defined (with three, four, five, seven, and nine classes). Using a supervised maximum likelihood algorithm, the comparisons suggested several trends in sensor performances. Overall accuracies of Landsat-7 ETM+ compared well with sensors with higher spatial (Ikonos) or spectral (MASTER) resolution for low or moderate habitat complexity mapping. For high-complexity mapping (nine classes), Ikonos performed best, suggesting that high spatial resolution is important. For low- and moderate-complexity mapping, MASTER performed best, suggesting that spectral resolution and digitization seem more critical. However, these trends must be discussed cautiously in the light of various factors before any generalization can be made. These factors include issues in reconciling–scaling ground-truth data at multiple spatial and thematic scale, reefs specificities, and environmental conditions during image acquisition. Résumé. Les performances des capteurs satellites Landsat-7 ETM+, ASTER, SPOT-HRV et Ikonos ainsi que le capteur aéroporté MASTER (instrument de simulation de MODIS–ASTER) ont été comparées pour la cartographie des habitats coralliens des récifs du Pacifique Sud. Cet ensemble unique de données fournit différentes résolutions spatiales (de 4 m pour Ikonos à 30 m pour Landsat ETM+), résolutions spectrales (de deux bandes visibles pour SPOT-HRV à cinq pour MASTER) et niveau de numérisation des données (entre 8 et 16 bits). Notre étude porte sur deux îles (Tahiti et Moorea) dont les structures récifales barrières et frangeantes sont représentatives de celles rencontrées dans les îles volcaniques du Pacifique Sud. Cinq niveaux de complexité d’habitats ont été définis, avec 3, 4, 5, 7 et 9 classes respectivement. Les performances des classifications (algorithme du maximum de vraisemblance) montrent plusieurs tendances. Pour les classifications en 3–7 classes, les résultats de Landsat-7 ETM+ sont comparables à ceux des capteurs à meilleure résolution spatiale (Ikonos) ou spectrale (MASTER). Ikonos fournit de meilleurs résultats pour 9 classes, ce qui suggère que la résolution spatiale est un facteur clé pour des classifications benthiques complexes. La résolution spectrale et le niveau de numérisation semblent plus critiques pour 3–7 classes car MASTER est le plus performant. Cependant, ces résultats doivent être interprétés avec prudence à la lumière de plusieurs facteurs avant de pouvoir être généralisés. Il faut tenir compte de la manière dont les classes benthiques sont agrégées thématiquement et spatialement, des particularités des récifs étudiés et enfin des conditions environnementales lors de la prise d’images.

Introduction 200 The enhanced thematic mapper plus (ETM+) sensor on board the Landsat-7 spacecraft provides the capability to achieve high-resolution assessment of coral reefs at a global scale. Indeed, the long-term acquisition plan (LTAP) of the Landsat-7 mission coordinated the multiple and exhaustive coverage of coral reefs worldwide since April 1999 (Arvidson et al., 2001). The resulting data set is unique in quantity and quality (low cloud cover) (Gasch et al., 2000). The archive provides substantial input for decadal-scale change detection analysis (Andréfouët et al., 2001; Palandro et al., 2003) and mapping. The latter application is expected to provide results © 2003 CASI

similar to those from the thematic mapper (TM) on board Landsat-5. For shallow marine habitat mapping in the Caribbean or Indo-Pacific reefs, TM was able to recognize typically four to six habitats including sedimentary areas,

Received 30 January 2002. Accepted 22 November 2002. P. Capolsini1 and C. Payri. Laboratoire Terre-Océan, Université de la Polynésie française, Tahiti, French Polynesia. S. Andréfouët. Institute for Marine Remote Sensing, College of Marine Science, University of South Florida, FL 33701, U.S.A. C. Rion. Université Paul Sabatier, Toulouse, France. 1

Corresponding author (e-mail: [email protected]). 187

Vol. 29, No. 2, April/avril 2003

seagrass beds, coral, and algal zones with a reasonable overall accuracy of -70% (Ahmad and Neil, 1994; Armstrong, 1993; Mumby et al., 1997). Mumby et al. (1997) conducted an intersensor comparative study in Turks and Caicos reefs (located southeast of the Bahamas). This study compared four satellite sensors (Landsat multispectral scanner (MSS), Landsat TM, SPOT XS, and SPOT Pan) and the high-resolution digital airborne sensor CASI (compact airborne spectrographic imager). All sensors provided very different spatial (from 0.5 to 80 m) and spectral (from 1 to 21 bands) resolutions. Reef habitats were categorized into coarse (four classes), intermediate (nine classes), and fine (13 classes) detail. The TM instrument obtained superior results for coarse habitat mapping and exhibited accuracy similar to that of SPOT XS and SPOT Pan for intermediate habitat detail. Landsat MSS was consistently the least accurate sensor, and CASI maps were significantly more accurate than satellite sensors at any level of detail. Our first objective is to increase the geographic coverage of Landsat evaluation for reef mapping. We focus on reef sites of the Pacific Ocean to balance the Caribbean results provided by Mumby et al. (1997). We believe that many comparative studies are required to account for the numerous local reef geomorphology and habitat zonations encountered worldwide. Turks and Caicos lagoons and reefs are representative of Caribbean structures but differ from reefs in other regions, such as the Indian Ocean or Indo-Pacific. Lagoons dominated by seagrass beds are common in the Caribbean but do not exist in any of the 118 islands of French Polynesia (South Pacific). Conversely, large barrier and fringing reefs common in South Pacific volcanic islands are not encountered in the Bahamas. Our second objective is to compare sensor performances for coral reef habitat mapping. The Turks and Caicos study was based on “old” satellite sensors such as Landsat TM and Landsat MSS. Comparative evaluation of satellite sensors should also include the most recent sensors such as Landsat ETM+, Ikonos, and the advanced spaceborne thermal emission and reflection radiometer (ASTER) (Abrams, 2000). In this study, we aim to compare image classification results obtained from old (SPOT high resolution visible imaging system (HRV)) and recent (ETM+, ASTER, Ikonos) space sensors. We add to this set of satellite sensors one airborne instrument, the MODIS–ASTER simulator (MASTER) (MODIS, moderate resolution imaging spectroradiometer) (Hook et al., 2001). ETM+ provides more coral reef images than any other sensor at global scale (Gasch et al., 2000), but it may be interesting to include other sources of information for a specific study. Additional data sources prove useful in avoiding persistent cloud cover or completing a time series for a change detection analysis. They should also be examined when data are available at minimal cost (e.g., ASTER). A joint U.S.–Japan program produced the ASTER instrument on board the Terra satellite launched in December 1999. Ikonos is the first commercial satellite offering 4 m spatial resolution in multispectral mode since late 1999. Despite the high costs of images, Ikonos with its high-resolution products is perceived as a future key player 188

in coral reef mapping. Assessment of its potential is urgently required. The recent Pacific Rim II (PACRIM II) campaigns (Tapley et al., PACRIM 2 Scientific Objectives, unpublished report) provided the opportunity to acquire hyperspectral data over coral reef areas using MASTER. Our third objective is to discuss the issues of spectral and spatial resolution and digitization. Addition of the MASTER instrument is key to this objective. Since MASTER data provide spatial resolution (20 m) comparable to that of ETM+, ASTER, and HRV, it further provides enhanced spectral resolution for underwater targets (eight bands in the range 400– 700 nm as opposed to two or three bands for space sensors). MASTER also provides enhanced radiometric sensitivity (16 bits digitization versus 11 bits for Ikonos and 8 bits for ETM+, ASTER, and HRV). This unique image data set allows us to provide a premier comparison of the performances of various recent sensors for coral reef habitat mapping in South Pacific islands. Although a few authors (Mumby et al., 1997; Hochberg and Atkinson, 2000; Lubin et al., 2001; Andréfouët et al., 2001; Mumby and Edwards, 2002) have already discussed the spectral and spatial resolution issues, bit depth comparisons have not been performed extensively. We discuss the influence, in coral reef habitat classification, of the spatial and spectral resolution and digitization. We expose and discuss the trends observed at two sites for different levels of habitat complexity and make conclusions concerning the relative influence of spectral–spatial resolution and digitization.

Material and methods Study areas Fieldwork on coral reef systems is generally of an expeditionary type. Constraints on boats, diver availability, weather conditions, and water conditions (visibility, waves, current) often limit the feasibility of extended ground truth. Therefore, we narrowed the fieldwork to two heterogeneous reefs. High heterogeneity allows visiting a large variety of reef habitats and limits the problem of spatial autocorrelation for accuracy assessment. This study takes place on Tahiti and Moorea, two volcanic islands of the Society Archipelago, French Polynesia. Both islands are rimed by reef structures (fringing reefs, barrier reefs, channels, and lagoons) typical of Pacific Ocean islands, with the exception of atolls, which are low-altitude carbonate islands (Guilcher, 1988). The main zone of interest, hereafter called Taapuna Reef (Figure 1), is located on the west coast of Tahiti Island, at 17°35′S and 149°37′W. It is a shallow barrier reef system, bounded by the oceanic crest on the west, a deep channel on the east, and the Taapuna pass on the south. There is no real fringing reef (i.e., in contact with land), though typical fringing communities are encountered at the edge of the channel. As a control site to confirm the trends observed on Taapuna Reef, we included a subset of Moorea Island (17°29′S, 149°53′W). This zone, hereafter called Tiahura Reef (Figure 1), is located on the northwestern part of Moorea. It © 2003 CASI

Canadian Journal of Remote Sensing / Journal canadien de télédétection

includes both barrier and fringing reefs, separated by a channel 2–5 m deep. Tiahura Reef was previously mapped using SPOT HRV data using a possibilistic multisource fusion classifier (Andréfouët et al., 2000). More details are provided hereafter on Tiahura habitats, but the presence of a real fringing reef on Tiahura is the main difference between Taapuna and Tiahura.

of the bands useful for underwater targets are compared in Figure 2. In the ASTER images, two bands used for stereoimaging (bands 3N and 3B) have similar spectral characteristics. ETM+, Ikonos, and ASTER images have been acquired in austral fall, and MASTER and HRV images in austral winter. Image acquisition took place during a short time interval (March 2000 – March 2001), with the exception of the SPOT HRV image acquired in 1986 and available only for Tiahura Reef. Conversely, Ikonos data exist only for Taapuna Reef.

Image data The main characteristics of the five types of images (Ikonos, ETM+, ASTER, MASTER, and HRV) used in this study are summarized in Table 1. The relative spectral responses (RSR)

Figure 1. Location of Taapuna and Tiahura reefs on Tahiti and Moorea islands, respectively, in French Polynesia. Table 1. Main characteristics of the images for Taapuna and Tiahura reefs.

Sensor

Date of acquisition

Spatial resolution (m)

SPOT HRV

11 Aug. 1986

20

3

8

Landsat-7 ETM+

26 April 2000

30

4

8

ASTER

21 Mar. 2001

15

3

8

MASTER

4 and 5 Aug. 2000

20

8

16

Ikonos

19 Mar. 2000

4

4

11

No. of spectral bandsa

Digitization (bits)

Weather conditions Light east wind (10–15 km/h); low south swell (1 m) Light east wind (20 km/h); low south swell (1 m) Light northeast wind (10 km/h); heavy north swell (2.5 m) Moderate east wind (30–40 km/h); moderate south swell (1.5–2.0 m) Strong northeast wind (40–50 km/h); moderate south swell (1 m)

Note: Weather conditions provided by Meteo-France at Tahiti Airport, a few kilometres to the north of Taapuna Reef and 20 km east of Tiahura Reef. a Visible to near infrared (VIS/NIR). © 2003 CASI

189

Vol. 29, No. 2, April/avril 2003

Figure 2. Relative spectral response (RSR) of the five sensors (MASTER, SPOT HRV, ASTER, Landsat-7 ETM+, and Ikonos) considered in this study. We used the first eight bands of the MASTER sensors. The SPOT HRV RSR is for SPOT 1, HRV1.

MASTER data were acquired for both reefs from a DC-8 flying at an altitude of 8000 m. Habitat classification scheme In this study, we provide a multisensor comparison for reef habitat mapping for different habitat complexity levels. There are many issues in this exercise. We first consider different levels of benthic habitat complexity. A typical simple level would be a sand–rubble–

190

constructed three-classes scheme, where the constructed class includes any bottom with dead or live coral covered or not by algae. Complexity can be added when considering benthic covers. For instance, constructed can be split into live coral, dead coral with encrusting algae, and dead coral with fleshy algae. Further, architectural growth forms can increase the complexity, splitting live coral into branching live coral and massive live coral. There are virtually endless possibilities of increasing this hierarchy of habitats by including as many in situ variables as desired. It is thematic driven. © 2003 CASI

Canadian Journal of Remote Sensing / Journal canadien de télédétection

If we consider only one sensor, there are no major difficulties in considering different level of habitats within the same hierarchy. For mapping purposes, however, habitat definition must depend on the spatial resolution of the image considered (Andréfouët and Claereboudt, 2000). This conditions the scale at which ground-truth data are collected and the hierarchy itself, since the quantitative–qualitative criteria (e.g., percent coral cover) used to define the classes are different. This difference is due to different levels of intrapixel homogeneity. For instance, it is very possible to find clusters of Ikonos pixels (at 4 m resolution) covered by >50% coral. This threshold can be used to define a generic coral class, but it is unlikely that a Landsat pixel (at 30 m resolution) will be entirely covered by >50% coral. At best, 10–20% coral cover can be expected for Landsat pixels for the types of reefs we have considered here. Therefore, a coral class should be defined with >15% coral cover for the 15–30 m resolution, but with >50% coral cover for Ikonos resolution. Keeping the same 50% threshold at a different scale will result in the absence of a coral class at the lowest resolution. This may be “politically incorrect” in coral reef environments (Mumby and Harborne, 1999). Thematically, however, it is important to highlight areas of highest coral cover at any scale, even if this cover is low. Habitat classes within a given hierarchy of habitats suitable for one sensor (e.g., Ikonos) can always exist, generically, in a different hierarchy suitable for another sensor (e.g., Landsat). However, the quantitative–qualitative definition of classes and ground-truth scale may differ. Reconciling different scales of perception is critical for many remote sensing applications based on multisensor, multiresolution data. There are no clear techniques to achieve this goal, since this is highly thematic dependent (e.g., Kalkhan and Stohlgren, 2000). In earlier coral reef studies, scaling of habitats according to the spatial resolution of the sensors was not accounted for. In Mumby et al. (1997), the same hierarchy of habitats was considered for every sensor, with spatial resolution from 1 to 80 m, implicitly assuming that detailed habitats surveyed and characterized at 1 m resolution also existed at 80 m resolution. In theory, it is possible to build different hierarchies for different resolutions by using different sets of variables (e.g., growth forms of corals for Ikonos, but not for Landsat) as envisioned by Andréfouët and Claereboudt (2000). The drawback is that sensor performance and classification results are compared with

different sets of classes. Unfortunately, the comparison may then be meaningless, but differences may be narrowed by avoiding variables that are too specific. Here, four sensors have relatively similar spatial resolution: ASTER (15 m), SPOT HRV (20 m), MASTER (20 m), and Landsat ETM+ (30 m). Ikonos is unique with its 4 m high resolution. We could have discarded this sensor, but we suspected it would permit detection of the level of habitat complexity for which spatial resolution is critical. We addressed the scaling challenges described earlier in the following way. First, we assumed that habitats present on the reefs were similar in the 15–30 m spatial resolution range. This is a reasonable assumption. Thus, the same hierarchy of habitats can be used to compare Landsat, ASTER, SPOT, and MASTER classification results. Second, to allow inclusion of 4 m Ikonos data in the comparison, we avoided habitat descriptors that were too specific such as remarkable species (coral, algae) or architectural variables (coral growth forms, size of rubbles, rugosity, etc.). Third, we used similar thresholds in coral and algal cover when defining the classes for different spatial resolution and used semiquantitative descriptors instead of quantitative ones. With these minimal assumptions and precautions, it is possible to compare multiresolution classification schemes. The typology of habitats for Taapuna Reef at Ikonos resolution came from 26 training sites. The sites were selected according to visual interpretation of colour and texture of the Ikonos image. For Landsat, ASTER, SPOT, and MASTER, the typology of habitats was built around the same “seed” pixels. We used almost the same environmental variables for the Tiahura Reef (Andréfouët et al., 2000) (Table 2). The only difference was that brown algae cover was explicitly accounted for in Taapuna Reef. For Tiahura, it was implicitly assumed that brown algae covered the dead coral structures (Andréfouët et al., 2000). This addition resulted in a slight modification of the initial Tiahura classification scheme (Table 3), with a new “algal crest” and the fusion of the classes sand and isolated patches on sand substratum (Andréfouët et al., 2000). In the field, at each site we noted the environmental variables on a semiquantitative scale (Table 2). A similarity matrix between sites was built using the Bray–Curtis similarity measurement (Legendre and Legendre, 1984). The sites were then hierarchically clustered using the group-average method to

Table 2. Environmental descriptors and codification used to categorize coral reef habitats on Taapuna barrier reef (Tahiti Island). Cover (%)

Dead and live coral structures

Code

Sand

Coral

Dead coral

Algae

Rubble

Pavement

Size (cm)

Height (cm)

1 2 3 4 5 6 7

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–5 5–15 15–25 25–50 50–75 75–90 90–100

0–10 10–40 40–100 100–150 >150

0–10 10–40 40–100 100–150 >150

Note: A qualitative variable, not presented here, was included to account for remarkable features (e.g., fringing-like coral forms of growth). © 2003 CASI

191

Vol. 29, No. 2, April/avril 2003 Table 3. Classification scheme of Tiahura Reef (Moorea Island) adapted from Andréfouët et al. (2000). Cover (%) Class

Dominant type of bottom

Terrigenous sand Dense patches fringing Fringing Sand Dead coral

Sand with rubble and mud Coral or terrigenous sand Coral or terrigenous sand Coral sand Heterogeneous (sand, rubble, pavement) Heterogeneous (sand, rubble, pavement) Heterogeneous (sand, rubble, pavement) Pavement, cemented dead coral (covered by brown or red algae)

Living coral Heterogeneous Algal crest

produce a dendrogram. We defined five levels of complexity by cutting the dendrogram for different thresholds of similarity (Figure 3). Image processing, classification, and accuracy assessment We applied a first-order atmospheric correction (dark object, i.e., deep water, correction) to every image. This linear correction does not change the results of a statistical classification. Analytically, however, if one wants to compute accurate above-water radiance or reflectance, deep water correction often results in an overcorrection in shallow water, since it also removes some water column contribution. In terms of bathymetric corrections, since our areas of interest excluded depths greater than 2 m, no water depth correction was necessary. The 26 sites at Taapuna Reef were used to train the image classifier. We added training pixels by photointerpretation around these sites to improve the radiometric definition of the classes. The exact location of all field sites was determined using a hand-held global positioning system (GPS) and by taking advantage of the precision in geo-location offered by the Ikonos products. Tests conducted on nearby remarkable sites (including a geodetic observatory located at Université de la Polynésie française) proved that Ikonos geodetic precision was excellent and conforms to the specifications (±12 m). Accordingly, a control site was generally a cluster of 4 × 4 or 5 × 5 pixels. In April and September 2001, we surveyed 20 transects at the Tiahura reef for training in addition to the transects described in Andréfouët et al. (2000). These new transects were 50–250 m long, and water clarity imposed a practical width of -20–40 m. If Taapuna Reef was classified in one block, Tiahura Reef was a priori segmented in two zones, fringing and barrier, to avoid some of the misclassifications observed in Andréfouët et al. (2000) This avoided the confusion between dark terrigenous sand at the boundary between land and fringing reef and constructed bottom with high coral cover at the oceanic edge of

192

Dead coral

Living coral

Location

<5 >15 <15 <5 >25

<5 >5 <5 <5 5–15

Fringing Fringing Fringing Fringing, barrier Barrier

<25

>25

Barrier

30

30

Barrier

<10

<10

Barrier

the barrier reef. For Tiahura images, we defined only eight classes. All images are classified using the maximum likelihood algorithm with equal probabilities of the classes implemented in the Environment for Visualizing Images (ENVI®) toolbox. To compare the accuracy of maps obtained from images with different spatial resolution, two strategies can be applied: (1) The number of control sites is the same for every image, but this may imply a different area of interest from one image to another. This is the approach followed by Mumby et al. (1997). If N control sites are required (and N should be greater than 50 (Hay, 1979)), one needs a larger zone to obtain N independent censuses in a Landsat image at 30 m resolution than in an Ikonos image at 4 m resolution. With this strategy, however, the comparison may be irrelevant because the area investigated is not the same from one image to another (Stehman, 1997). (2) The area of interest is exactly the same for every image. With this option, the number of control sites Ni may need to be different from one image i to another because of the different spatial resolutions. Since testing the difference in accuracy between two partitions Pi and Pj does not require Ni = Nj (Stehman, 1997), we favoured this strategy. In November 2001, 2680 control sites at the Taapuna Reef were sampled using a random stratified sampling scheme and chosen based on the Ikonos image. These 2680 sites are scattered throughout the area of interest and were sampled by the same team of snorkelers. Each site was assigned to one of the classes for each level of habitat complexity. A coarse Landsat, MASTER, or ASTER pixel could include several Ikonos control sites. To reconcile the scales, Ikonos control pixels were spatially and thematically averaged in the following way: a coarse pixel (which may include one or several Ikonos control pixels) was assigned to the dominant

© 2003 CASI

Canadian Journal of Remote Sensing / Journal canadien de télédétection

Figure 3. Hierarchical clustering of 26 Taapuna Reef training sites and definition of the different levels of habitat complexity. At a coarse level, three classes can be defined for a dissimilarity threshold of about 65%. Increasing this value provides more complex typologies. A1, dead structures with algae on sandy bottom; A2, sand; A3, dead structures without algae on sandy bottom; B1, fringing-like dead structures with algae on sandy bottom; B2, dead structures with algae on sandy bottom; B3, sand; B4, dead structures without algae on sandy bottom; B5, fringing-like dead structures without algae on sandy bottom; Hetero., heterogeneous.

Ikonos control class. If no class was dominant, we visually checked the image and made a decision based on the signature and proportion of all the Ikonos pixels within the coarse pixel. We ensured that no obvious discrepancies occurred by comparing the mean radiance of training pixels with the radiance of each control pixel. This allowed us to identify dark coarse pixels assigned to bright class, or vice versa. These artefacts were explained by the fact that an Ikonos control pixel could fall on a pocket of sand (bright) within networks of coral constructions (dark) or on isolated coral heads (dark) in the middle of sand (bright). Based on these evaluations, we limited

© 2003 CASI

the number of control points for Landsat, MASTER, and ASTER to 53, 100, and 167, respectively, on Taapuna Reef. For Tiahura Reef, no Ikonos image was available to help with accurate georeferencing of ground-truth data. Thus, we collected ground-truth data along transects. They are easier to locate than individual pixels because they frequently cross remarkable areas (sand patches, etc.). In addition to the 11 transects described in Andréfouët et al. (2000), in April and September 2001 we surveyed 12 more transects for control. Transect starting points were randomly selected along the coast or along the algal crest. On Tiahura Reef, we obtained 1327,

193

Vol. 29, No. 2, April/avril 2003

Figure 6. Overall accuracy and tau coefficients for Tiahura Reef classifications in eight classes using SPOT HRV, MASTER, Landsat-7 ETM+, and ASTER sensors and a maximum likelihood classifier. Error bars are 95% confidence intervals. Table 4. Results of Z tests comparing the differences in tau (τ) coefficients for Taapuna Reef for each combination of sensors and each level of habitat complexity (in three to nine classes).

Nine classes Ikonos MASTER ETM+

Figure 5. Overall accuracy (a) and tau (b) coefficients for Taapuna Reef classifications in nine, seven, five, four, and three classes (Cl.) for Ikonos, MASTER, Landsat-7 ETM+, and ASTER sensors using a maximum likelihood classifier. Error bars are 95% confidence intervals.

1042, 885, and 548 control pixels for ASTER, SPOT, MASTER, and Landsat data, respectively. Accuracy of the classifications is quantified using confusion matrices and tau (τ) and overall accuracy (P) measurements (Stehman, 1997). We present both measurements for the purpose of comparison with previously published studies. P is simply the proportion of control pixels correctly classified, and τ measures the improvement of a classification over a random assignment of pixels (Ma and Redmond, 1995). Variance and confidence interval (95%) of the accuracy measurements were estimated as specified in Ma and Redmond (1995), assuming a normal distribution. While comparing two classifications with accuracies Pi (or τi) and Pj (or τj), the Z test was used to check the null hypothesis H0 (Pi = Pj) and the alternative hypothesis H1 (Pi ≠ Pj) (Ma and Redmond, 1995; Stehman, 1997).

Seven classes Ikonos MASTER ETM+ Five classes Ikonos MASTER ETM+ Four classes Ikonos MASTER ETM+ Three classes Ikonos MASTER ETM+

MASTER

ETM+

ASTER

4.633

3.556 0.030

5.261 0.714 0.560

0.357

1.690 1.607

3.169 2.306 0.076

1.129

0.159 0.785

2.184 2.335 0.996

2.324

0.186 1.116

2.400 3.428 1.452

1.677

0.383 1.197

1.510 2.311 0.461

Note: Values in bold indicate rejection of the null hypothesis (95% significance level) of equality of the accuracies, i.e., significant difference between sensors.

Results and discussion For Taapuna Reef, analysis of the dendrogram suggests five levels of habitat complexity, with three, four, five, seven, and nine classes (Figure 3). Landsat maps for each level of

194

© 2003 CASI