article in press and monitor shallow coral reefs worldwide, to characterize inter-reef structural...
TRANSCRIPT
ARTICLE IN PRESS
1
2
3
4
5
6
7891011121314151617181920
21
2223242526272829303132333435
36
373839
40
41
www.elsevier.com/locate/rse
Remote Sensing of Environment xx (2003) xxx–xxx
ED PROOF
Multi-site evaluation of IKONOS data for classification of
tropical coral reef environments
Serge Andrefoueta,*, Philip Kramerb, Damaris Torres-Pullizac, Karen E. Joyced,Eric J. Hochberge, Rodrigo Garza-Perezf, Peter J. Mumbyg, Bernhard Rieglh,
Hiroya Yamanoi, William H. Whitej, Mayalen Zubiak, John C. Brockc,Stuart R. Phinnd, Abdulla Naseerl, Bruce G. Hatcherl, Frank E. Muller-Kargera
a Institute for Marine Remote Sensing, College of Marine Science, University of South Florida, 140 7th Avenue S., St. Petersburg, FL 33701, USAbRosenstiel School of Marine and Atmospheric Science, University of Miami, Miami, FL, USA
cCenter for Coastal and Regional Marine Studies, United States Geological Survey, St. Petersburg, FL, USAdBiophysical Remote Sensing Group, Department of Geographical Sciences and Planning, University of Queensland, St. Lucia, Australia
eHawaii Institute of Marine Biology, University of Hawaii, Honolulu, Kaneohe, USAfCoral Reef Ecosystems Ecology Laboratory, Marine Resources Department, CINVESTAV-I.P.N. Unidad Merida, Merida, Mexico
gMarine Spatial Ecology Laboratory, University of Exeter, Exeter, UKhOceanographic Center, National Coral Reef Institute, Nova Southeastern University, Miami, FL, USA
iSocial and Environmental Systems Division, National Institute for Environmental Studies, Onogawa, Tsukuba, Ibaraki, JapanjDepartment of Marine Science and Coastal Management, The University of Newcastle, Newcastle upon Tyne, UK
kLaboratoire Terre-Oceans, Universite de la Polynesie Francaise, Tahiti, French PolynesialDepartment of Biology, Dalhousie University, Halifax, Nova Scotia, Canada
TReceived 3 June 2002; received in revised form 5 December 2002; accepted 22 April 2003NCORRECAbstract
Ten IKONOS images of different coral reef sites distributed around the world were processed to assess the potential of 4-m resolution
multispectral data for coral reef habitat mapping. Complexity of reef environments, established by field observation, ranged from 3 to 15
classes of benthic habitats containing various combinations of sediments, carbonate pavement, seagrass, algae, and corals in different
geomorphologic zones (forereef, lagoon, patch reef, reef flats). Processing included corrections for sea surface roughness and bathymetry,
unsupervised or supervised classification, and accuracy assessment based on ground-truth data. IKONOS classification results were
compared with classified Landsat 7 imagery for simple to moderate complexity of reef habitats (5–11 classes). For both sensors, overall
accuracies of the classifications show a general linear trend of decreasing accuracy with increasing habitat complexity. The IKONOS sensor
performed better, with a 15–20% improvement in accuracy compared to Landsat. For IKONOS, overall accuracy was 77% for 4–5 classes,
71% for 7–8 classes, 65% in 9–11 classes, and 53% for more than 13 classes. The Landsat classification accuracy was systematically lower,
with an average of 56% for 5–10 classes. Within this general trend, inter-site comparisons and specificities demonstrate the benefits of
different approaches. Pre-segmentation of the different geomorphologic zones and depth correction provided different advantages in different
environments. Our results help guide scientists and managers in applying IKONOS-class data for coral reef mapping applications.
D 2003 Elsevier Inc. All rights reserved.
U Keywords: Landsat; Bathymetric correction; Glint; Accuracy; Habitat mapping; Seagrass42
43
44
45
1. Introduction
Remote sensing provides an effective way to observe
and monitor shallow coral reefs worldwide, to characterize
46
47
48
49
0034-4257/$ - see front matter D 2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2003.04.005
* Corresponding author. Tel.: +1-727-553-3987; fax: +1-727-553-1103.
E-mail address: [email protected] (S. Andrefouet).
inter-reef structural differences, and to map intra-reef
habitat diversity and zonations, assess bathymetric varia-
tions, design survey protocols, conduct biogeochemical
budgets, and map beta-diversity (Andrefouet, Claereboudt,
Matsakis, Pages, & Dufour, 2001; Andrefouet, Muller-
Karger, Hochberg, Hu, & Carder, 2001; Andrefouet &
Payri, 2000; Capolsini, Andrefouet, Rion, & Payri, 2003;
Hochberg & Atkinson, 2000; Jupp et al., 1985; Liceaga-
RSE-05965
T
ARTICLE IN PRESS
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx2
UNCORREC
Correa & Euan-Avila, 2002; Mumby, 2001; Mumby &
Edwards, 2002; Mumby, Green, Clark, & Edwards, 1998;
Palandro, Andrefouet, Dustan, & Muller-Karger, 2003;
Purkis, Kenter, Oikonomou, & Robinson, 2002; Roelfsma,
Phinn, & Dennisson, 2002). The recent increase of remote
sensing applications targeting reef environments (Andre-
fouet, Muller-Karger, et al., 2001) reflects the growing
concern about drastic and negative changes occurring on
reefs over the past three decades due to anthropogenic (e.g.
pollution, fishing, and coastal development) or natural (e.g.
global warming) stresses.
The satellite data most commonly used since the mid-
1980s for direct observation of coral reefs have been
medium spatial resolution digital images, i.e. a spatial
resolution of 10–30 m. This includes data delivered by
the Indian Remote Sensing Satellite C (IRS-C), Satellite
pour l’Observation de la Terre (SPOT) 1–4 High Resolu-
tion Visible (HRV), Landsat 5 Thematic Mapper (TM), and
more recently by SPOT 4–5, Landsat 7 Enhanced The-
matic Mapper Plus (ETM+), and Advanced Spaceborne
Thermal Emission and Reflection Radiometer (ASTER)
sensors. Conversely, ‘‘high resolution’’ images are those
with a spatial resolution greater than 10 m such as those
provided by IKONOS or Quickbird (1–4 m). This study
aims to clarify the potential of high spatial resolution
IKONOS data for reef mapping worldwide.
In their review presenting the use of remote sensing for
coastal tropical assessment, Green, Mumby, Edwards, and
Clark (1996) pointed out the difficulty of comparing
different reef assessments due to lack of consistency in
classification schemes, in in situ data collection and image
processing methods, and in accuracy assessment protocols.
SPOT HRV and Landsat TM data have been used most
frequently because of their availability starting in the mid-
1980s. The work of various independent investigators
worldwide has now helped define the potential of medium
spatial resolution data for reef applications (e.g. Ahmad &
Neil, 1994; Andrefouet, Muller-Karger, et al., 2001; Mat-
sunaga & Kayanne, 1997; Purkis et al., 2002; Yamano &
Tamura, in press). It is now clear that for geomorphology
and habitat-scale applications, SPOT and Landsat data are
adequate for simple complexity mapping (3–6 classes), but
for more complex targets (7–13 classes) they are limited
by their spatial and spectral resolution and likely by their
digitization rate (8 bits) (Capolsini et al., 2003; Hochberg
& Atkinson, 2003; Mumby, Green, et al., 1998; Mumby &
Edwards, 2002).
The 1999 launch of the commercial IKONOS satellite,
operated by Space Imaging (SI), provides for the first
commercial space sensor with 11 bit, high-spatial resolu-
tion (4 m), calibrated data in four wide spectral bands that
are potentially useful for coral reef studies. The spectral
bands closely match the first four bands of the ETM+
sensor (Thome, 2001). Despite the quick attenuation of red
radiance in water, the near-infrared (NIR) band is poten-
tially useful for very shallow water targets (Menges, Hill,
ED PROOF
& Ahmad, 1998) and low tide conditions when benthos is
exposed. The IKONOS satellite data has generated great
interest among coral reef researchers that were previously
limited to the use of aerial color photographs (Andrefouet
et al., 2002; Palandro et al., 2003) or costly digital airborne
multispectral or hyperspectral data (Mumby, Green,
Edwards, & Clark, 1999) for very high resolution work.
To the best of our knowledge, four independent peer-
reviewed studies have already addressed the potential of
IKONOS for reef habitat mapping and there are probably
many more investigations in progress, judging by the
number of coral scenes acquired and archived by SI.
Mumby and Edwards (2002) compared IKONOS classifi-
cation results of Caribbean (Turks and Caicos) coastal
areas between different spaceborne and airborne sensor
data. They concluded that IKONOS data were unable to
discriminate habitats in a complex (13 classes) classifica-
tion scheme, but that they were adequate for moderate to
simple complexity mapping (9–5 classes). They found
IKONOS provides an acceptable accuracy (64–74% over-
all accuracy), although similar to that obtained with a
medium resolution sensor like the Landsat TM. Capolsini
et al. (2003) applied a similar multi-sensor comparative
approach for South Pacific (Tahiti) reefs and reached
similar conclusions (66–86% overall accuracy for 7–3
classes). Maeder et al. (2002) also mapped a Caribbean
reef (Roatan Island, Honduras) with good results (>85%
accuracy) using a simple (five classes, including a deep
water class) classification scheme. Finally, assuming linear
mixing and using in situ spectral reflectance measure-
ments, Hochberg and Atkinson (2003) simulated IKONOS
classification for various sea floors made of different
proportions of algae, coral, and sand. Without actually
processing any IKONOS data, they suggested that coral
dominated habitat could not be accurately separated from
algae dominated habitat using IKONOS data. Using sim-
ulated IKONOS data, coral cover was overestimated and
algae cover underestimated on their test-site of Kanehoe
Bay, HI. Capolsini et al. confirmed this prediction using
real images of Tahiti reefs, where true coral habitats
(coral>60%) were poorly assessed (9.54% users’ accuracy)
in a simple classification scheme (ruble, sand, algae,
coral).
The number of IKONOS evaluations for reefs is already
quite impressive and, more importantly, each seems to
provide results consistent with the others. However, be-
cause of the variety of sites and methods considered, we
felt that some of the Green et al. (1996) remarks would
still be valid if many independent studies were conducted
without some coordination to prevent too much heteroge-
neity in terms of classification scheme (habitat descrip-
tion), classification algorithms, and accuracy assessment
protocols. Therefore, beginning in 2000, the University of
South Florida (USF) worked with the NASA Scientific
Data Purchase (SDP) program (Stennis Space Center) to
task IKONOS for acquisition of scenes of representative
ARTICLE IN PRESS
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 3
reefs around the world. We worked with international
remote sensing and coral biology/ecology/geology scien-
tists to process and evaluate the images on their most
intensively studied research sites where ground-truth data,
local reef expertise, and other satellite or airborne remote
sensing data were available. Most of these investigators
agreed to contribute to the present study, making this effort
certainly the largest international cooperation for remote
sensing of reefs.
This paper presents a synthesis of the results in terms of
classification and mapping of coral reef habitats using
IKONOS data, as well as a comparison with ETM+
performances for selected sites. Each investigator that
contributed to this study will likely publish their own
detailed results in the future on different subjects since
each of them has a specific interest in the IKONOS data
(e.g. change detection). Despite the initial wish to have
consistent methodology rigorously applied throughout the
mapping exercise, local specificities and expertise, and cost
of field work lead to slightly different means of data
processing, ground-truthing strategy, or evaluation of
results. Nevertheless, the bulk of work provides clear
trends and lessons discussed hereafter.
T237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
UNCORREC2. Material and methods
2.1. Reef-sites and image data-set
Ten different sites were considered for this comparative
study (Fig. 1). The sites represent the primary biogeograph-
ic coral regions of the world according to Veron (1995),
and includes bank reefs, fringing reefs, barrier reefs, and
atolls.
Table 1 summarizes the IKONOS and Landsat 7 data
used for this study. We avoided images with obvious
water quality issues (e.g. suspended sediments). Only
scenes with clear waters were used, thus minimizing
the effects of variable water optical properties for this
comparative study. Until early 2001, the IKONOS images
delivered to NASA SDP were systematically resampled
with cubic convolution (CC), but analysis of the same
image delivered with CC and nearest neighbor (NN)
resampling showed that the textural information was
significantly degraded using CC in different coral reef
shallow floors (SA, unpublished data). Thus, we system-
atically requested NN resampling after April 2001. IKO-
NOS data were geocorrected in UTM WGS-84, as Master
Standard (MS) or Original Standard (OS) products (Dial,
Bowen, Gerlach, Grodecki, & Oleszczuk, in press), but
never as Precision products since no ground control
points were provided to Space Imaging for our study
sites. Landsat 7 ETM+ data (Table 1) were ordered
through the Eros Data Center in either Geotiff or HDF
format, with NN resampling and UTM WGS-84. The
Landsat image of Shiraho Reef, Japan was provided by
ED PROOF
the National Space Development Agency of Japan, in
Fast-L7 format, NN resampling.
2.2. Habitats: similarities and differences between sites
One of the challenges of this comparative study is to
reconcile different habitat classification schemes. Indeed, the
10 sites are characterized by different bottom features, and
the methods used by investigators in each region addressed
their problems in particular ways. The various investigators
provided classification schemes ranging from simple (4–5
classes for Shiraho and Glovers), to moderately complex (7–
8 classes for Arue, Addu, Biscayne Dubai, and Andros), and
then to very complex (13–15 classes for Mayotte, Heron,
and Boca Paila) (Table 2). The very complex sites were also
described in less detail by hierarchical simplification. Heron
Reef has been described using 13, 7, and 5 habitats classes.
Boca Paila has been described using 15 and 8 habitats.
Mayotte has been split from 1 general site with 14 general
classes into 2 sub-sites with 10 specific classes each, pro-
viding better thematic description for each sub-site.
To compare classification results throughout this range of
habitat complexity, it is desirable to put the various schemes
into the same hierarchical framework. For the four Atlantic/
Caribbean sites (Biscayne in USA, Andros in Bahamas,
Glovers in Belize, and Boca Paila in Mexico), we can refer
to the scheme provided by Mumby and Harborne (1999)
(Fig. 2). This is a multi-level hierarchical model with geo-
morphologic and benthic components. This scheme was
applied to each site but it was necessary to adjust the
thresholds in benthic cover and depth (Fig. 3). For instance,
Andros coral cover was high (5–15%) in most of the
geomorphological strata. Thus, most of the Andros reef
should be classified as ‘‘coral’’ since according to Mumby
and Harborne (1999) ‘‘coral’’ classes are characterized by a
cover >1%. However, a coral label for the entire Andros reef
would be misleading. Similarly, the notion of ‘‘deep’’ or
‘‘shallow’’ lagoon floors differs between sites (Figs. 2 and 3).
Extrapolation of the Mumby and Harborne (1999)
geomorphology/benthic hierarchical framework to non-Ca-
ribbean sites is possible in theory. For Shiraho (Japan) and
Dubai, the geomorphology of the site is simple: shallow
(0–3 m) lagoon for Shiraho and gentle slope for Dubai.
Therefore, since there is only one geomorphologic unit,
there is no specific reference to geomorphology in the
classification scheme and only benthic features were used
(Table 2). Conversely, for Addu, Arue, Heron, and espe-
cially Mayotte, reference to geomorphological strata is
required (Table 2). Mayotte is by far the most complex
site. Quod, Bigot, Dutrieux, Maggiorani, and Savelli.
(1995) pointed out the richness of this site, characterized
by two contrasted barrier-reefs separated by the deep
(f 70 m) Longogori pass. The northern barrier reef
(Pamandzi Reef) is characterized by extensive seagrass
and algal beds and internal spur and grooves systems,
while the southern reef (Ajangoua Reef) is free of seagrass
UNCORRECTED PROOF
ARTICLE IN PRESS
270
271
272
273
274
275
276
277
Fig. 1. Location of the study sites. The map indicates the main coral biogeographic areas according to Veron (1995). Our sites represent the Arabian Gulf,
Indian Ocean, Indo–Pacific, Pacific, and Caribbean biogeographic zones. Each site is presented using a RGB color composite based on the red, green, and blue
bands of the IKONOS sensor. Size (in km) of the image showing the most characteristic zone is indicated, though the actual processed area may be wider.
Includes material Space Imagingn.
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx4
and richer in small coral heads, with presence of enclosed
lagoons. Our own in situ survey of Mayotte Reef in
December 2000 highlighted nearly 30 different habitats,
the main difference with Quod et al. was the extensive
areas of dead corals, consequences of a 1998 coral
bleaching event. For mapping purposes, we considered as
a starting point only 14 broader classes for Mayotte (Table
2). Arue reef in Tahiti Island (French Polynesia) is repre-
UNCORRECTED PROOF
AR
TIC
LE
IN P
RE
SS
t1.1 Table 1
Site information and characteristics of the IKONOS and Landsat imagest1.2
Site and Latitudeb Longitudeb IKONOS Landsat 7 ETM+ Ground-truth Depth (m) Referencest1.3investigatora
Acquisition date Product Resampling Path/row Acquisition date Formatt1.4
Addu AN, SA, BGH � 0.6149 73.1201 14 March 2000 MS CC 145/60 20 December 2000 HDF March 2002 0–15 Stoddart (1964)t1.5Andros PK 24.5833 � 77.7833 12 March 2001 OS NN 13/43 26 March 2000 EDC-Geotif 2000/2001 0–30 Kramer, Kramer,
and Ginsburg (1998)t1.6Arue SA, MZ � 17.5750 � 149.6000 11 March 2000 OS CC 53/72 6 June 2000 Earthsat-Fast June 2000 0–12 Frouin and Hutchings (2001)t1.7Biscayne DTP, JB, SA 25.3500 � 80.2166 18 March 2001 OS CC 15/42 5 February 2000 EDC-HDF 2001/2002 2–12 Jaap (1984)t1.8Boca Paila RG 19.9166 � 87.5000 20 July 2000 MS CC 19/46 24 July 1999 EDC-Geotif 1999/2000 0–30 Kramer and Kramer (2000)ct1.9Dubai BR 24.9400 54.8900 2 May 2001 MS NN 160/42 N/A N/A Fall 1995 0–9 Riegl (1999)t1.10Glovers PJM, WHW 16.8166 � 87.8000 12 April 2001 OS NN 18/48 8 November 2000 EDC-Geotif 1999–2001 0–18 McClanahan and
Muthiga (1998)t1.11Heron SA, KEJ, SRP � 23.5258 151.8900 7 May 2001 MS NN 90/77 14 November 1999 EDC-HDF 2001/2002 0–15 Rogers (1997),
Smith et al. (1998)t1.1291/77 18 September 1999 EDC-HDFt1.13
Mayotte SA � 12.8933 45.2180 31 August 2000 OS NN 161/69 30 August 2000 EDC-Geotif Dec. 2000 0–15 Quod et al. (1995)t1.14Shiraho HY 24.3166 124.2333 28 March 2002 OS NN 115/43 23 February 2002 Fast-L7A EROSd 1999 0–3 Kayanne et al. (2002),
Harii and Kayanne
(submitted for publication)t1.15
OS: Original Standard, MS: Master Standard, CC: cubic convolution, NN: nearest neighbor convolution, N/A: not available.t1.16a Investigators: initials from author list.t1.17b Lower left corner of the IKONOS scene. Negative latitude for South, negative longitude for West.t1.18c Describe Yucatan reefs.t1.19d Provided by National Space Development Agency of Japan.t1.20
S.Andrefo
uet
etal./Rem
ote
Sensin
gofEnviro
nmentxx
(2003)xxx–
xxx5
UNCORRECTED PROOF
ARTICLE IN PRESS
278
279
280
281
282
283
284
285
286
287
288
289
t2.1 Table 2
Classification scheme for the 10 sites (Boca Paila 15 classes and Mayotte 14 classes are not shown)t2.2
Site Class label Site Class labelt2.3
Addu Sand/rubble, backreef Andros Coral (Acropora sp.) crestt2.4Classes: 8 Sand, backreef Classes: 8 Dense coral (Montastraea sp.) on forereeft2.5
Sand, lagoon floor Gorgonian plain on forereeft2.6Coral/algae, patch reef Dense gorgonian on escarpmentt2.7Algae, crest Algae on low relief spur and groovest2.8Coral, forereef Pavement on forereeft2.9Seagrass, backreef Deep (>12 m) sand on forereeft2.10Coral, backreef Shallow sand/ruble on forereeft2.11
Arue Dense brown algae (>80%), crest Biscayne Dense patchest2.12Classes: 8 High density coral heads on heterogeneous pavement Classes: 8 Diffuse patchest2.13
Low density coral heads on sandy floor Deep (>3 m) dense seagrass on lagoon floort2.14Dense brown algae (>50%) on reef flat, heterogeneous floor Deep (>3 m) sparse seagrass on lagoon floort2.15Moderate brown algae (15–25%) on reef flat,
heterogeneous floor
Shallow ( < 3 m) dense seagrass on lagoon floort2.16
Sparse brown algae ( < 5%) on reef flat, heterogenous floor Shallow ( < 3 m) moderate seagrass on lagoon floort2.17Sand and ruble (>90%) Shallow ( < 3 m) sparse seagrass on lagoon floort2.18Deep (>10 m) lagoon floor Deep (>3 m) sand on lagoon floort2.19
Boca Paila Dense coral (>25%) Dubai Dense coralt2.20Classes: 7 Moderate coral ( < 25%) Classes: 8 Sparse coralt2.21
Seagrass/algae Seagrasst2.22Heterogeneous lagoon floor Shallow algaet2.23Pavement Deep algaet2.24Shallow sand, lagoon floor Pavementt2.25Deep sand, forereef Shallow sandt2.26
Deep sandt2.27Glovers Forereef Heron Branching and massive corals on forereeft2.28Classes: 5 Brown algae Classes: 5 Multi-growth forms corals (>25%) on reef flatt2.29
Seagrass/Lobophora sp. Heterogeneous (coral < 25%) reef flatt2.30Seagrass Heterogeneous reef flat, sand bottom
(10%< rock–algae–coral < 25%)t2.31Sand Sand (>90%)t2.32
Heron Branching corals (>75%) on forereef and crests Mayotte Forereeft2.33Classes: 13 Multi-growth forms, dense corals (>50%) on reef flat Classes: 14 Dense coral margins on spur-and-grooves,
enclosed lagoon and passt2.34Multi-growth forms, moderate corals (25–50%) on reef flat Dense coral heads (>25%) on heterogeneous floor, reef flatt2.35Coral pavement (>75%) on crest and reef flats Sparse coral heads ( < 10%) on sand floor, reef flatt2.36Coral head, forereef Dense Thalassodendron sp. seagrass
and Padina sp. algae (>80%)t2.37Heterogeneous reef flat (coral < 10%, sand, rocks,
fleshy algae, coralline)
Dense Thalassodendron sp. seagrass (>80%)t2.38
Dead coral (>80%) coated by encrusting coralline Diffuse seagrass ( < 25%) on sandt2.39Dead coral (>80%) covered by fleshy algae Mixed seagrass and algaet2.40Sand with rocks ( < 15%) covered by fleshy algae Mixed seagrass beds, with corals, rocks, ruble, algaet2.41Sand with scattered dead coral heads ( < 15%),
coated by encrusting coralline
Brown algae on reef flat and crestst2.42
Sand with dense dead coral heads (>50%),
covered by fleshy algae
Coralline mountst2.43
Coral sand and mud (>90%) Pavement and rublet2.44Pavement and ruble (>90%) Deep sand channels on forereeft2.45
Shallow sand on reef flatt2.46Shiraho Coralt2.47Classes: 4 Seagrass/algaet2.48
Pavementt2.49Sandt2.50
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx6
sentative of the Tahitian reefs under moderate influence of
the harbor and industrial activities (Frouin & Hutchings,
2001). This is a morphologically complex reef, with a
fringing reef, deep channels, a large barrier reef and three
large patch reefs (Fig. 2). Two wide passes connect the
channels with the ocean on each side of the reef complex.
Heron Reef is a large reef platform of the southern Great
Barrier Reef (Australia) and one of the most studied reef
sites in the world, including using remote sensing. Here,
we considered only the western side (Fig. 2) where most
of the ground-truth data were collected in 2001 and 2002.
It is also the most heterogeneous area since the eastern
ORRECTED PROOF
ARTICLE IN PRESS
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
Fig. 2. Top: Habitat classification scheme as proposed by Mumby and Harborne (1999) for mapping Caribbean coral reefs. Both geomorphologic and benthic
keys may define a habitat. For instance, three habitats encountered on dense patch reefs are highlighted: a coral zone with presence of massive colonies of
Montastraea sp., an algal zone dominated by the brown algae Lobophora sp., and a bare substrate zone densely covered by gorgonians. Depending on the
precision of the classification scheme, a practitioner can provide details only at geomorphological level, or at benthic levels, or a combination of both. Bottom:
for Glovers Reef, the final classification scheme is a simple five-class scheme, with a forereef (geomorphology) class and four generic benthic classes (brown
algae, seagrass, mixed seagrass, and sand) not related to a specific geomorphologic zone (e.g. sand includes sandy areas in any geomorphological zones).
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 7
UNCside of the bank is dominated by a large shallow sandy
lagoon (Smith, Frankel, & Jell, 1998).
Table 2 summarizes the different classes, without details
on the main benthic species, which are available in the
references provided for most of the sites (Table 1). These
references describe in detail the benthic community structure
present on each site. Species-level description is the final
level of the hierarchy of habitats. Clearly, dominant coral,
algae, or seagrass species are generally different from one
region (or one site) to another. For instance, Mayotte has
extensive Thalassodendron sp. seagrass beds, while Thalas-
sia sp. and Syringodium sp. dominate Biscayne and Halo-
phila sp. and Halodule sp. are frequent in Dubai. Andros and
Boca Paila coral crests are dominated by Acropora palmata,
but behind Shiraho’s crest Montipora spp. and Heliopora
coerulea are dominant. Brown algae Sargassum sp. and
Turbinaria sp. are dominant on Arue patch reef, while
Lobophora sp. extensively colonized Glovers patch reefs.
2.3. IKONOS image processing
IKONOS data were processed differently depending on
image quality and characteristics of the site. Main stages of
processing included surface roughness correction, depth
correction, classification, and accuracy assessment.
2.3.1. Surface roughness correction
A large fraction of the images acquired via NASA SDP
suffer from sea surface effects due to wind-generated wave
patterns and associated sun glint. As of May 2002, nearly
50% of the 40 images (not all of them processed here)
delivered to USF suffered significantly from this problem
UNCORRECTED PROOF
ARTICLE IN PRESS
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
Fig. 3. Example of contrasted habitat classification schemes for two Caribbean sites: Biscayne Bay, FL (USA) and Andros island (Bahamas). Coloured sections
highlight the zones present on each reef. For Biscayne, the study area includes lagoon floor and patch reefs. For Andros, the area of interest includes the oceanic
side of the coral reef system with the forereef, crests, and spur-and-grooves. In both sites, benthic classes are related to a geomorphologic zone, highlighted with
similar color. For instance, for Biscayne, a class of dense seagrass has been defined both for the shallow lagoon floor (blue) and the deep lagoon (green). Sand
is related on the lagoon floor (yellow), without separation between deep and shallow. Diffuse gorgonians are related to diffuse patch reefs (purple). For Andros,
the same bi-component geomorphology-benthos hierarchy occur, e.g. dense gorgonians occur only on high relief escarpment (purple); forereef (green) is split
into four classes: sand (yellow), diffuse gorgonian (light purple), coral Montastraea (dark green), and bare bedrock (grey). Andros and Biscayne both comprise
eight classes of habitats, but with minimum overlap between the two schemes. This illustrates the variety of habitats classification scheme that can occur in the
same region (see also Fig. 2 for Glovers).
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx8
(Fig. 4). For this study, a sea surface roughness correction
was required for Heron and Biscayne. Dubai and Glovers
images were also corrected, but the wave patterns were not
that significant and the original images were used. Details
of the algorithm for sea surface correction are provided in
Hochberg, Andrefouet, and Tyler (in press). Briefly, the
process consists of analyzing the sea surface assuming that
the water is virtually opaque in the NIR band (Siegel,
Wang, Maritorena, & Robinson, 2000) and that the relative
amount of radiance reflected upward at the sea surface is
solely a function of geometry, independent of wavelength.
This means that pixels with glint contribution in NIR
bands also have similar glint contribution in total upward
radiance in visible bands. Identifying the pixels with
maximum and minimum radiances in the NIR enables
estimation of the percentage of glint contribution in each
pixel, which is then corrected to absolute radiance in the
visible bands.
2.3.2. Depth correction or depth-invariant indices
Lyzenga (1981) proposed a method to eliminate the
effects of water column attenuation on bottom radiances
ECTED PROOF
ARTICLE IN PRESS
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
Fig. 4. Pre-processing of IKONOS images. Top: Example of surface correction aimed at removing wave and glint effects that occur on almost 50% of the 40
images provided by NASA SDP to USF. The level of details available in deep areas is drastically improved after correction. Bottom: An example of false color
composite made with three depth-invariant bottom indices, one from each pair of IKONOS bands 1–3. The backreef appears in pink dividing the outer forereef
from the lagoon. This technique limits misclassification due to depth between shallow dark objects and deep bright objects. Surface correction and depth-
invariant indices were not systematically applied, depending on image quality and site characteristics (see Table 3).
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 9
UNCORR(or reflectances). Basically, Lyzenga showed that pixels of
the same bottom-type located at various unknown depths
appear along a line in the bidimensional histogram of two
log-transformed visible bands. The slope of this line is the
ratio of diffuse attenuation of the two bands. Repeating
this for different bottom types at variable depth results in a
series of parallel lines, one for each bottom type. Projec-
tion of these lines onto an axis perpendicular to their
common direction results in an unitless depth-invariant
bottom-index where all pixels from a given bottom-type
receive the same index-value regardless of its depth. Two
visible bands (or one 2D histogram) provide one index.
Three bands can provide three depth-invariant indices by
permutation. The main drawback of this method is that
index values cannot be related to radiance or reflectance
measurements. Also, in some cases, application of Lyzen-
ga’s method is problematic because the same bottom type
may not occur over a wide range of depths, thus biasing
the accurate estimation of the ratio of diffuse attenuation
(Maritorena, 1996). This is the case for Biscayne and Arue
in this project. Nevertheless, it has been proven that
applying similar empirical techniques and classifying the
resulting index images instead of the initial images could
significantly increase the accuracies of the maps (Mumby,
Clark, Green, & Edwards, 1998). Here, the technique has
been applied to Glovers (Fig. 4), Heron, Andros, and Boca
Paila for two reasons: (1) sites presented a significant
depth range and (2) deep sand channels and shallow
lagoon sand pools were adequate to train and apply the
method.
2.3.3. Unsupervised and supervised classification
The regional investigators applied two strategies. For
Glovers and Shiraho, they first conducted an unsupervised
classification and then assigned the different segments to
a given benthic category according to expert knowledge
and ground-truth data. For Andros, Heron, Mayotte,
Dubai, Arue, Biscayne, Boca Paila, ground-truthed poly-
gons in each class were used to train a supervised
maximum likelihood classifier. For Mayotte, the two
contrasted barrier-reefs were processed simultaneously
and then separately. For Glovers only, a contextual
decision rule was applied. It re-classified any lagoonal
pixel classified spectrally into ‘‘forereef’’ to the correct
T
ARTICLE IN PRESS
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
t3.1t3.2
t3.3
t3.4t3.5t3.6t3.7t3.8
t3.9t3.10t3.11t3.12t3.13t3.14t3.15t3.16t3.17t3.18t3.19
t3.20t3.21t3.22t3.23t3.24t3.25
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx10
C
category. Depending on the depth of the site or the tide
conditions, three or four IKONOS spectral bands were
used (Table 3).
2.3.4. Accuracy assessment
It was not possible to survey any of the reefs simul-
taneously with the acquisition of the images (Table 1)
even though some sites were visited several times in
consecutive years (Table 1). The time gap between
ground observations and image acquisition is generally
between a few months and a year. The exceptions were
Dubai, where large-scale video tracks (Riegl, Korrubel, &
Martin, 2001) were collected in 1995 and the image
acquired in 2001, and Shiraho, with a 3-year gap
(Kayanne, Harii, Ide, & Akimoto, 2002). Mumby and
Edwards (2002) also reported a gap of 5 years. They
stated that habitat delineation was unlikely to evolve
significantly during this time frame. This may actually
be questionable depending on the history of the site and
the type of perturbations that may have occurred (hurri-
canes, coral bleaching, etc.), but for our data set we have
considered this statement valid, except for Dubai. Formal
accuracy assessments were summarized in confusion ma-
trices for each site with generally more than 100 inde-
pendent control points for the total partition, with the
exception of Mayotte (1230 points along 25 transects of
variable length) and Heron (5 transects provided 93
points, in addition to 72 isolated independent points
reef-wide) (Fig. 5). The Dubai classification has been
compared with continuous shipborne video surveys that
highlight the rate of change on the reefs after the 1998
UNCORRETable 3
Image processing parameters and classification results
Site Surface
correction
Depth
correction
IKONOS
bands
Nb co
point
Addu N N 4 400
Andros N Y 3 150
Arue N N 4 200
Biscayne Y N 3 123
Boca
Paila
N Y 3 150
Dubai N N 3 N/A
Glovers N Y 3 150
Heron Y Y 3 165a
Mayotte N N 4 1230c
Shiraho N N 4 104
N/A: Not computed.a Partly independent point, partly clusters (see text for details).b Using image from 18/09/99 (low tide) and 14/11/99 (high tide), respectivelyc Clusters only (see text for details).d Ajangoua barrier reef (no seagrass).e Pamandzi barrier reef (seagrass).
ED PROOF
bleaching event. Dubai reef areas that may have changed
after 1998 and that could bias the accuracy assessment
were removed. The assessment was systematic, along a
grid placed over the image. Observed and computed
classes at each grid-point were used to build the confu-
sion matrix, for a total of 1086 points. The accuracy
metric that we have considered in this study is the overall
accuracy (Stehman, 1997), i.e. the proportion of control
points correctly classified.
2.4. Landsat 7 image processing
Where both IKONOS and Landsat 7/ETM+ data were
available (Table 1), ETM+ data were processed similarly
to the IKONOS data, with the exception of sea surface
correction that was never applied. We did not try to
classify ETM+ data for high complexity classification
(>10 classes), but only for moderate complexity (5–10
classes). The same ground-truth point/transects were used
for accuracy assessments for both IKONOS and Landsat
classifications. Of particular interest is the tandem IKO-
NOS/ETM+ for Mayotte, acquired only 1 day apart in
August 2000 in similar low-tide conditions. For Heron,
we considered two ETM+ images, one each in high and
low tide condition, and we used three and four bands,
respectively. For Andros, we do not provide an overall
accuracy because the area considered for Landsat map-
ping included the shallow lagoon and forereef, whereas
for IKONOS, only the forereef was processed. Moreover,
Landsat accuracy assessment was not done independently
but was based on original training polygons. Further, we
ntrol
s
Classes Overall accuracy
IKONOS (%)
Overall accuracy
Landsat 7 (%)
8 66 56
8 74 N/A
8 70 52
8 84 56
15 45 N/A
7 74 53
8 71 N/A
11 51 42
5 77 71
13 42 N/A
7 61 N/A
5 78 66/61b
14 61 N/A
10d 73 56
10e 68 50
4 81 63
.
RECTED PROOF
ARTICLE IN PRESS
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
Fig. 5. Example of validation of IKONOS classification using in situ large scale transects on Heron Reef. The classifications of three sites (Dubai, Mayotte, and
Heron Reef) have been controlled using this technique. Here, on the south of Heron, the transition between five benthic classes can be compared with in situ
observation where the percent cover of different substrates has been estimated visually in 20� 20 m units. Despite the difference in resolution (4 vs. 20 m), we
note the good agreement between in situ data and classification, with accurate definition of transition zones, homogeneous zones, and heterogeneous patchy
zones (data from Joyce, Phinn, Roelfsema, Neil, & Dennison, 2002).
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 11
Rcompare IKONOS and Landsat only in qualitative terms
for this site.
O 468469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
UNC3. Results and discussion
Examples of classification accuracy and final maps are
provided in Figs. 5 and 6. Classification accuracies for
each site obtained with IKONOS and Landsat are provided
in Table 3. Fig. 7 presents the pooled overall accuracies vs.
habitat complexity achieved with IKONOS. We also added
the results available for Roatan (Maeder et al., 2002),
Turks and Caicos (Mumby & Edwards, 2002), and
Punaauia (Capolsini et al., 2003). For Dubai, image
classification in eight classes (Table 2) identified as algae
several areas previously dominated by corals in 1995.
Generally, the four different classes of dense corals iden-
tified by video were accurately recognized but only as a
broad coral class (Table 2). A posteriori ground-truthing
showed that the coral areas apparently misclassified as
algae really changed after the 1998 bleaching event (Riegl,
1999), suggesting that IKONOS could be used for moni-
toring changes in benthic communities for similar sites if a
reference had been available. This is confirmed by Palan-
dro et al. (2003) who combined aerial photographs and one
IKONOS image of Carysfort Reef (Florida) to quantify the
rate of coral loss in the last 20 years.
The general trend in Fig. 7 is a linear decrease of
accuracy with increasing complexity. Accuracies range
from an average of 77% for 4–5 classes to 71% for 7–
8 classes, 65% in 9–11 classes, and 53% for more than 13
classes. There is no obvious bias that could be explained
by the skills of the investigators. For instance, in 13–14
classes, both best and worst results are provided by the
same investigator (SA). At 8–10 classes, all results are
very consistent. The variations are the natural consequen-
ces of the nature of the site and the way that the images
have been processed (Table 3). We did not compute the
variance of each accuracy because the sampling schemes
for accuracy assessment were not the same and in some
cases the rigorous conditions of application would likely
have been violated (Foody, 2002; Stehman, 1997; Steh-
ORRECTED PROOF
ARTICLE IN PRESS
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
Fig. 6. Examples of classification maps for Arue, Andros, Biscayne, and Glovers. For the three first sites overall accuracy was >70%, while for Glovers the
11-class scheme led to a poor 50% accuracy, prompting the implementation of a more simple five-class scheme (Fig. 2).
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx12
UNCman, 1999). For reference, Mumby and Edwards (2002) or
Capolsini et al. (2003) computed the 95% confidence
interval for different accuracy metrics (Tau and overall
accuracy). Confidence intervals were in the range F 5–
10% for several hundred independent control points,
depending on the classification schemes (3–13 classes)
and the number of control points per class. Our initial
intention was to apply more systematic processing using
the same classification schemes, but this proved to be quite
difficult or impossible because of the many local specific-
ities and constraints.
The linear decreasing trend in Fig. 7 could be used to
estimate a priori the accuracy to be expected for a given site
using the methods described here. Fig. 7 compares the
results achieved by IKONOS and Landsat. Landsat also
provides a linear trend, but at a level 15–20% lower than
IKONOS throughout the range of habitat complexity.
Depending on the range of depth, type of geomorphology
and habitats, and water clarity at the time of acquisition, the
final accuracy will be modulated in the range highlighted
here. For instance, for IKONOS, for simple four-class
mapping, 68–81% overall accuracy can be expected accord-
ing to Turks and Caicos (T&C), Shiraho, and Punaauia.
Highest accuracies are obviously obtained for the shallow
fringing and barrier reefs (Shiraho and Punaauia), while
T&C comprised habitats between 0 and 25 m depth. Fig. 7
also can aid in deciding the level to which habitat complex-
ity can be addressed for a given level of accuracy for both
sensors. Using the data presented in Fig. 7, two relations
may be derived: overall accuracy (Y) vs. number of classes
(X) was Y=� 3.90X + 86.38 (r2 = 0.63) for Landsat and
Y=� 2.78X + 91.69 for IKONOS (r2 = 0.82). If an 80%
accuracy is required for scientific or management applica-
tions, only four to five classes can be used using IKONOS.
RECTED PROOF
ARTICLE IN PRESS
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
20
30
40
50
60
70
80
90
2 3 4 5 6 7 8 9 10 11 12 13 14 15
Roatan
Punaauia
Turks and Caicos
Heron
Andros
Arue
Biscayne
Glovers
Mayotte
Shiraho
Boca Paila
Addu
Dubai
20
30
40
50
60
70
80
90
2 5 643 987 10 11 12 13 14 15
Arue
Biscayne
Mayotte (Ajangoua)
Mayotte (Pamandzi)
Shiraho
Boca Paila
Addu
Heron (3 bandes)
Heron (4 bandes)
Punaauia
Turks and Caicos
Glovers
Number of habitat classes
Number of habitat classes
Ove
rall
accu
racy
(%
)O
vera
ll ac
cura
cy (
%)
Fig. 7. Relation between habitat classification scheme complexity (number of classes) and overall accuracy for the IKONOS sensor (top) and Landsat ETM+
sensor (bottom). Roatan data are from Maeder et al. (2002). Punaauia data are from Capolsini et al. (2003). Turks and Caicos data are from Mumby and
Edwards (2002), using textural information for IKONOS processing and Landsat 5 TM data.
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 13
UNCORAround the 70% threshold, most of sites with less than 10
classes are included, but there are exceptions.
The accuracies reported here are for the overall parti-
tion. Individual habitats that may be of importance for
particular applications do not necessarily have the same
accuracy, for better or worse. For most of the sites, we do
not have enough control points to discuss accuracy within
each individual class (Congalton & Green, 1999), but
some observations are worth mentioning. For instance,
sand is always very well classified (e.g. 90–92% user’s
accuracy for deep and shallow sand in Dubai). Conversely,
with the exception of Dubai (59% user’s accuracy), true
coral areas are generally poorly classified when algae-
dominated habitats also occur on the same site at the same
depth in a patchy fashion, confirming the predictions of
Hochberg and Atkinson (2003) with 43.8% user’s accura-
cy. High accuracy for coral zones seems possible on
several sites (e.g. Andros and Heron) but this is because
the image processing has been stratified by geomorpho-
logic zone (e.g. Andros with 97% user’s accuracy for
forereef-high relief Montastraea annularis zones) or be-
cause there is no real competition with other classes (e.g.
Heron does not have shallow dense seagrass beds spec-
trally similar to deep corals). High accuracy also occurs for
forereef geomorphologic zones, which are implicitly coral-
rich (95% user’s accuracy in Glovers).
Comparison of the accuracy of IKONOS and Landsat
assessments confirms previously published results obtained
with moderate resolution sensors (Landsat TM, SPOT-
HRV, ASTER). An exception was Chauvaud, Bouchon,
and Maniere (1998) and Chauvaud, Bouchon, and Maniere
(2001), whose accuracies are certainly optimistically biased
(>90% for more than 25 classes using SPOT-HRV). Land-
sat accuracies also decreased linearly with increasing
habitat complexity. It is adequate for simple habitat com-
plexity mapping, and the results for seven to eight classes
are consistent, within a low range of accuracy (50–56%)
(Table 3). For Heron, lower complexity (five classes)
provided acceptable accuracy (>60%), especially with the
image acquired at low tide since this resolved the shallow
heterogeneous hard-bottom and coral zones along the
reef rim (Table 3). For the Andros forereef where ETM+
T
ARTICLE IN PRESS
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx14
UNCORREC
accuracy could not be computed without optimistic bias,
IKONOS better discriminated shallow water ( < 5 m)
habitats which are often small and patchy. The texture of
the habitat seems exploitable using the IKONOS sensor at
4-m resolution for this site (see below further discussion
on texture). In deeper (>5 m) forereef zones, habitats are
more spatially continuous (e.g. high relief Montastraea
zone visible in Fig. 6) and both sensors provide similar
patterns. Narrow geomorphological features such as terra-
ces and steps are more evident on IKONOS. The decision
to use IKONOS or Landsat (or any other medium resolu-
tion sensor) is constrained generally by consideration
of cost-effectiveness (Mumby et al., 1999). IKONOS
is clearly not the best option for covering large areas
(Mumby & Edwards, 2002). For small areas of high
interest (research sites or marine protected areas), our
results show the level of improvement in mapping accu-
racy that can be expected for science or management
applications. This also demonstrates the potential of con-
ducting multi-sensor reef mapping by coupling Landsat
broad (but accurate) 4/5-class mapping for large reef
stretch (e.g. Florida Keys, Bahamas, Great Barrier Reef,
etc.) with more precise (and still accurate) IKONOS
mapping for specific (patchy) areas.
Higher accuracies could be potentially achieved using
textural measurements as suggested by Mumby and
Edwards (2002). For coarse habitat mapping (four classes)
on Turks and Caicos, overall accuracy increased from 68%
to 75% when textural neo-channels (3� 3 window vari-
ance) were combined with depth-invariant indices in the
classification process. However, this is likely to be site
(patchiness, depth) and image quality dependent. Indeed,
tests on Mayotte show that despite contrasted textural
signatures for different bottom-types, the texture did not
improve the overall accuracy. For Mayotte, the fact that we
could also use the NIR bands throughout most of the reef
system likely explains this lack of improvement. Mumby
and Edwards also noted that texture did not improve
classification results when using 10-band multispectral
airborne data. Thus, the extra spectral information seems
more important than texture. We suggest that more sys-
tematic tests are required to fully quantify and qualify the
benefits of using textural signatures in various conditions.
Another potential way to improve accuracy is to use
contextual knowledge to modify the classifications a poste-
riori (Mumby, Clark, et al., 1998). The benefit of contextual
knowledge can also be inferred from our results. Indeed,
higher accuracies were achieved when the reefs were a
priori segmented into main geomorphologic or contrasted
zones. Glovers was processed as a whole (including fore-
reef, lagoon, and rim) and reached 77% accuracy in five
very broad classes (Table 2), even after some contextual
editing. Conversely, Andros reached 74% in eight very
specific classes (Table 2), but was initially pre-segmented
to process only the forereef, thus avoiding confusion with
classes present on the backreef (e.g. seagrass, small patch
ED PROOF
reefs). The backreef processed alone would likely provide
results similar to Biscayne (lagoon floor with seagrass/sand,
patch reefs), where eight classes were mapped with 84%
accuracy. Finally, splitting the Mayotte reef system in 2
barrier reefs with 10 very specific habitats each yielded
better results than a global classification in 14 thematically
broader classes. This strongly suggests that the practitioner
should carefully pre-segment the image and then process by
zones to optimize results. This can be performed empirical-
ly, by visual interpretation, or in a formal way. Andrefouet,
Roux, Chancerelle, and Bonneville (2000) formalized with a
fuzzy membership function the ‘‘distance to the shore’’
factor for the classification of SPOT images to avoid
misclassification between fringing mud and coral crests.
Suzuki, Matsakis, Andrefouet, and Desachy (2001) integrat-
ed directions (‘‘perpendicular’’, ‘‘parallel’’) when classify-
ing reef flats in atolls using Landsat imagery. However,
these are complex processes and for simplicity, in most of
the cases, pre-segmentation is made relatively easy by the
high resolution of the IKONOS images where fronts and
boundaries can be accurately delineated by simple visual
interpretation (e.g. Andros). Eventually, similar habitats in
different zones can be merged depending on the desired
classification scheme (e.g. algae on forereef merged with
algae on patch reefs).
It has been suggested that the depth correction tech-
nique should be systematically applied (Green, Mumby,
Edwards, & Clark, 2000) because of its simplicity and
potential benefit. Here, we had difficulties to do so at each
site. Biscayne, for instance, does not have the same
habitats at different depths and the Lyzenga (1981) method
cannot be applied. In this case, slightly different techniques
(e.g. ratios, Maritorena, 1996; Paredes and Spero, 1983;
Polcyn, Brown, & Sattinger, 1970) or ancillary field data
(bathymetry) are required to overcome the bathymetric
challenge. The patterns of misclassifications (patches clas-
sified as dense seagrass) on Biscayne suggest that a depth
correction could be useful, even though without bathymet-
ric correction, the overall accuracy appears quite good
(84%). For Andros, tests showed that the maximum
distance between habitats was achieved using two depth-
invariant indices (using bands 1–2 and bands 2–3) and the
original band 3. A similar conclusion arose for Boca Paila
where only the depth-invariant index based on bands 1 and
2 was useful, in conjunction with other unprocessed bands.
In general, there is no doubt that bathymetric correction at
the scale of habitats should enhance overall accuracy and
avoid misclassifications. However, many of these misclas-
sifications may happen simply because very different
zones are considered simultaneously in the classification
process. In a Caribbean reef, considering lagoon and back
reef communities (e.g. seagrass) with forereef coral com-
munities in the same process typically leads to problems.
Therefore, we suggest that future work should examine the
relative benefits of contextual knowledge and depth com-
pensation to achieve high accuracy. Indeed, this is proba-
ARTICLE IN PRESS
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx 15
bly site-dependent. Mumby, Clark, et al. (1998) showed
that depth correction is more critical than contextual
editing for a Turks and Caicos site, but in this case, the
morphology of the site was simple.
T
726
727
728
729
730
731
732733734735736737738739740741742743744745746747748749750751752753754755
EC
4. Conclusion
This is the first international coordinated effort to assess
the potential of a high resolution spaceborne sensor for coral
reef studies. This is also the first consistent compilation of
coral reef applications from commercially available data,
made available via the NASA SDP program. This work
informs scientists and managers on the type of accuracy
they can expect using IKONOS and Landsat 7 ETM+ data
for a coral reef mapping applications. However, our objec-
tives were partially met. We failed to adequately conduct
similar processing on a variety of sites as we initially
planned. We quickly realized that this is inherent to the
diversity of sites we have selected and to the different image
qualities dependent on environmental conditions (wind,
tide, water quality). Nevertheless, the processing of 10 sites,
completed by 3 independent studies, clearly highlight the
potential of IKONOS for coral reef habitat mapping in
general, using standard methods for reef habitat mapping.
Results can be improved for each site since more sophisti-
cated techniques may work better for a given site and for a
given image. For instance, to achieve high accuracy, we
recommend to pre-segment a reef into different zones if
possible. The comparative approach also suggests which
algorithms need to be improved or developed in the future
(bathymetric correction, contextual edition, and texture) for
a given type of reefs.
R 756757758759760761 OR5. Uncited referenceMaritorena, Morel, & Gentili, 1994
762763764765766767768769770771772773774775776777778779780781782783UNCAcknowledgements
Obviously, this study would not have been possible
without the spirit and crew of the NASA Scientific Data
Purchase program at Stennis Space Center under the
successive responsibility of Fritz Pollicelli and Troy
Frisbie. We greatly acknowledge their help in the tasking
requests and their support of our research since 1999.
Andrew Mattee and Michael Satter were our contacts at
Space Imaging. This research was supported by NASA
grants NAG5-10908 to SA and NAG-3446 to FMK and
Kendall Carder. Many individuals helped gather field data
on the various sites. We are indebted to Chris Roelfsema,
Bill Dennison, Fabienne Bourdelin, Claude Payri, Bernard
Thomassin, Michel Pichon, Hajime Kayanne, Saki Harii,
Yoshiyuki Tanaka, Ernesto Arias-Gonzales, Don Hickey,
Nancy Dewitt, Tonya Clayton, David Palandro, Chuanmin
Hu, and the staffs of Heron Island Research Station,
Biscayne National Park, and Service des Peches et de
l’Environment Marin de Mayotte. We are grateful to GIS-
LAGMAY (Mayotte) and the World Bank, via Andy
Hooten, for their support for travel and field work in
Heron Isl. Finally, thanks to Adam Lewis (Great Barrier
Reef Marine Park Authority) who provided the Landsat
images of Heron Reef. This is IMaRS contribution 052.
ED PROOFReferences
Ahmad, W., & Neil, D. T. (1994). An evaluation of Landsat Thematic
Mapper (TM) digital data for discriminating coral reef zonation:
Heron Reef (GBR). International Journal of Remote Sensing, 15,
2583–2597.
Andrefouet, S., Berkelmans, R., Odriozola, L., Done, T., Oliver, J., &
Muller-Karger, F. E. (2002). Choosing the appropriate spatial resolution
for monitoring coral bleaching events using remote sensing. Coral
Reefs, 21, 147–154.
Andrefouet, S., Claereboudt, M., Matsakis, P., Pages, J., & Dufour, P.
(2001). Typology of atolls rims in Tuamotu archipelago (French Poly-
nesia) at landscape scale using SPOT-HRV images. International Jour-
nal of Remote Sensing, 22, 987–1004.
Andrefouet, S., Muller-Karger, F., Hochberg, E., Hu, C., & Carder, K.
(2001). Change detection in shallow coral reef environments using Land-
sat 7 ETM+ data. Remote Sensing of Environment, 79, 150–162.
Andrefouet, S., & Payri, C. (2001). Scaling-up carbon and carbonate me-
tabolism in coral reefs using in situ and remote sensing data. Coral
Reefs, 19, 259–269.
Andrefouet, S., Roux, L., Chancerelle, Y., & Bonneville, A. (2000). A
fuzzy possibilistic scheme of study for objects with indeterminate boun-
daries: Application to French Polynesian reefscapes. IEEE Transactions
on Geoscience and Remote Sensing, 38, 257–270.
Capolsini, P., Andrefouet, S., Rion, C., & Payri, C. (2003). A comparison of
Landsat ETM+, SPOT HRV, IKONOS, ASTER and airborne MASTER
data for coral reef habitat mapping in South Pacific islands. Canadian
Journal of Remote Sensing, 29, 187–200.
Chauvaud, S., Bouchon, C., & Maniere, R. (1998). Remote sensing tech-
niques adapted to high resolution mapping of tropical coastal marine
ecosystems (coral reefs, seagrass beds and mangrove). International
Journal of Remote Sensing, 19, 3625–3639.
Chauvaud, S., Bouchon, C., & Maniere, R. (2001). Cartographie des bioce-
noses marines de Guadeloupe a partir des donnees SPOT (recifs coral-
liens, phanerogammes marines, mangroves). Oceanologica Acta, 24,
S3–S16.
Congalton, R., & Green, K. (1999). Assessing the accuracy of remotely
sensed data: Principles and practices. New York: Lewis Publishers.
Dial, G., Bowen, H., Gerlach, F., Grodecki, J., & Oleszczuk, R. (2003).
IKONOS satellites, imagery, and products. Remote Sensing of Environ-
ment (in press).
Foody, G. M. (2002). Status of land cover classification accuracy assess-
ment. Remote Sensing of Environment, 80, 185–201.
Frouin, P., & Hutchings, P. A. (2001). Macrobenthic communities in a
tropical lagoon (Tahiti, French Polynesia, central Pacific). Coral Reefs,
19, 277–286.
Green, E. P., Mumby, P. J., Edwards, A. J., & Clark, C. D. (1996). A review
of remote sensing for the assessment and management of tropical coast-
al resources. Coastal Management, 24, 1–40.
Green, E. P., Mumby, P. J., Edwards, A. J., & Clark, C. D. (2000). Remote
sensing handbook for tropical coastal management. Paris: UNESCO.
Harii, S., & Kayanne, H. (submitted for publication). Larval dispersal,
recruitment, and adult distribution of the brooding stony octocoral He-
liopora coerulea on Ishigaki island, southwest Japan. Coral Reefs.
T
ARTICLE IN PRESS
784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850
851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917
S. Andrefouet et al. / Remote Sensing of Environment xx (2003) xxx–xxx16
UNCORREC
Hochberg, E., & Atkinson, M. (2000). Spectral discrimination of coral reef
benthic communities. Coral Reefs, 19, 164–171.
Hochberg, E., & Atkinson, M. (2003). Capabilities of remote sensors to
classify coral, algae and sand as pure and mixed spectra. Remote
Sensing of Environment, 85, 174–189.
Hochberg, E. J., Andrefouet, S., & Tyler, M. R. (2003). Sea surface cor-
rection of high spatial resolution IKONOS images to improve bottom
mapping in near-shore environments. IEEE Transactions on Geociences
and Remote Sensing (in press).
Jaap, W. C. (1984). The ecology of the South Florida coral reefs: A com-
munity profile. Report Fish Wildlife Service OBS-82/08.
Joyce, K. E., Phinn, S. R., Roelfsema, C., Neil, D. T., & Dennison, W. C.
(2002). Mapping the southern Great Barrier Reef using Landsat ETM
and the Reef Check classification scheme. Proceedings of 11th Austral-
asian Remote Sensing and Photogrammetry Conference. Brisbane:
Causal Publications, CDROM.
Jupp, D. L. B., Mayo, K., Kuchler, D. A., VanClaasen, D., Kenchington,
R. A., & Guerin, P. R. (1985). Remote sensing for planning and manag-
ing the Great Barrier Reef of Australia. Photogrammetria, 40, 21–42.
Kayanne, H., Harii, S., Ide, Y., & Akimoto, F. (2002). Recovery of coral reef
populations after the 1998 bleaching on Shiraho Reef in the southern
Ryukyus, NW Pacific. Marine Ecology Progress Series, 239, 93–103.
Kramer, P. A., & Kramer, P. R. (2000). Ecological status of the Meso-
American Barrier Reef System. Impacts of Hurricane Mitch and 1998
coral bleaching. Report to the World Bank, RSMAS, University of
Miami, 73 pp.
Kramer, P. A., Kramer, P. R., & Ginsburg, R. N. (1998). Assessment of
coral reef health, Andros Barrier Reef, Bahamas. Progress Report,
RSMAS-University of Miami.
Liceaga-Correa, M. A., & Euan-Avila, J. I. (2002). Assessment of coral reef
bathymetric mapping using visible LANDSAT TM data. International
Journal of Remote Sensing, 23, 3–14.
Lyzenga, D. R. (1981). Remote sensing of bottom reflectance and water
attenuation parameters in shallow water using aircraft and Landsat data.
International Journal of Remote Sensing, 2, 71–82.
Maeder, J., Narumalani, S., Rundquist, D., Perl, R., Schalles, J., Hutchins,
K., & Keck, J. (2002). Classifying and mapping general coral-reef
structure using IKONOS data. Photogrammetric Engineering and Re-
mote Sensing, 68, 1297–1305.
Maritorena, S. (1996). Remote sensing of the water attenuation in coral
reefs: A case study in French Polynesia. International Journal of Re-
mote Sensing, 17, 155–166.
Maritorena, S., Morel, A., & Gentili, B. (1994). Diffuse reflectance of
oceanic shallow waters: Influence of water depth and bottom albedo.
Limnology and Oceanography, 39, 1689–1703.
Matsunaga, T., & Kayanne, H. (1997). Observation of coral reefs on Ishi-
gaki island, Japan, using Landsat TM images and aerial photographs.
Proceedings of 4th Int. Conf. on Remote Sensing for Marine and Coast-
al Environments ( pp. 657–666), Orlando, FL.
McClanahan, T., &Muthiga, N. (1998). An ecological shift in a remote coral
atoll of Belize over 25 years.Environmental Conservation, 25, 122–130.
Menges, C. H., Hill, G. J. E., & Ahmad, W. (1998). Landsat TM data
and potential feeding grounds for threatened marine species turtle in
northern Australia. International Journal of Remote Sensing, 19,
1207–1221.
Mumby, P. J. (2001). Beta and habitat diversity in marine systems: A new
approach to measurement, scaling and interpretation. Oecologia, 128,
274–280.
Mumby, P., Clark, C. D., Green, E. P., & Edwards, A. J. (1998). Benefits of
water column correction and contextual editing for mapping coral reefs.
International Journal of Remote Sensing, 19, 203–210.
Mumby, P. J., & Edwards, A. J. (2002). Mapping marine environments
with IKONOS imagery: Enhanced spatial resolution can deliver greater
thematic accuracy. Remote Sensing Environment, 82, 248–257.
Mumby, P., Green, E. P., Clark, C. D., & Edwards, A. J. (1998). Digital
analysis of multispectral airborne imagery of coral reefs. Coral Reefs,
17, 59–69.
ED PROOF
Mumby, P. J., Green, E. P., Edwards, A. J., & Clark, C. D. (1999). The cost-
effectiveness of remote sensing for tropical coastal resources assessment
and management. Journal of Environmental Management, 3, 157–166.
Mumby, P. J., & Harborne, A. R. (1999). Development of a systematic
classification scheme of marine habitats to facilitate regional manage-
ment and mapping of Caribbean coral reefs. Biological Conservation,
88, 155–163.
Palandro, D., Andrefouet, S., Dustan, P., & Muller-Karger, F. E. (2003).
Change detection in coral reef communities using the IKONOS sen-
sor and historic aerial photographs. International Journal of Remote
Sensing, 24, 873–878.
Paredes, J. M., & Spero, R. E. (1983). Water depth mapping from passive
remote sensing data under a generalized ratio assumption. Applied Op-
tics, 22, 1134–1135.
Polcyn, F. C., Brown, W. L., & Sattinger, I. J. (1970). The measurement of
water depth by remote sensing techniques. Report 8973-26-F, Willow
Run Lab., University of Michigan, Ann Arbor.
Purkis, S., Kenter, J. A. M., Oikonomou, E. K., & Robinson, I. S. (2002).
High resolution ground verification, cluster analysis and optical model
of reef substrate coverage on LANDSAT TM imagery (Red Sea, Egypt).
International Journal of Remote Sensing, 23, 1677–1698.
Quod, J. P., Bigot, L., Dutrieux, E., Maggiorani, J. M., & Savelli, A.
(1995). La reserve de la Passe en S (Ile de Mayotte). Expertise
biologique et cartographie des peuplements benthiques. Report AR-
VAM/IARE/SCE DES PECHES, Sainte Clotilde, La Reunion, 31
pp. + appendices.
Riegl, B. (1999). Corals in a non-reef setting in the Arabian Gulf (Dubai,
UAE): Fauna and community structure in response to recurring mass
mortality. Coral Reefs, 18, 63–73.
Riegl, B., Korrubel, J. L., & Martin, C. (2001). Mapping and monitoring of
coral communities and their spatial patterns using a surface based video
method from a vessel. Bulletin of Marine Science, 69, 869–880.
Roelfsma, C., Phinn, S., & Dennisson, C. W. (2002). Spatial distribution of
benthic microalgae on coral reefs determined by remote sensing. Coral
Reefs, 21, 264–274.
Rogers, R. W. (1997). Brown algae on Heron Reef flat, Great Barrier Reef,
Australia: Spatial, seasonal and secular variation in cover. Botanica
Marina, 40, 113–117.
Siegel, D. A., Wang, M., Maritorena, S., & Robinson, W. (2000). Atmos-
pheric correction of satellite ocean color imagery: The black pixel as-
sumption. Applied Optics, 39(21), 3582–3591.
Smith, B. T., Frankel, E., & Jell, J. S. (1998). Lagoonal sedimentation and
reef development on Heron Reef, southern Great Barrier Reef province.
In G. Camoin, & P. Davies (Eds.), Reefs and carbonate platforms in
the Pacific and Indian Oceans ( pp. 281–294). Blackwell Science.
Stehman, S. V. (1997). Selecting and interpreting measures of thematic
classification accuracy. Remote Sensing of Environment, 62, 77–89.
Stehman, S. V. (1999). Basic probability sampling designs for thematic
map accuracy assessment. International Journal of Remote Sensing,
20, 2423–2441.
Stoddart, D. R. (1966). Reef studies at Addu atoll, Maldive Islands: Pre-
liminary results of an expedition to Addu atoll in 1964. Atoll Research
Bulletin, 116, 1–122.
Suzuki, H., Matsakis, P., Andrefouet, S., & Desachy, J. (2001). Satellite
image classification using expert structural knowledge: A method based
on fuzzy partition computation and simulated annealing. Proceedings of
Annual Conference of the International Association for Mathematical
Geology (CDROM) Cancun, Mexico.
Thome, K. (2001). Absolute radiometric calibration of Landsat 7 ETM+
using the reflectance-based method. Remote Sensing of Environment,
78, 27–38.
Veron, J. E. N. (1995). Corals in space and time: The biogeography and
evolution of the Scleractinia. Comstock/Cornell.
Yamano, H., & Tamura, M. (in press). Can satellite sensors detect coral
reef bleaching? A feasibility study using radiative transfer models in
air and water. Proceedings of 9th Int. Coral reef Symposium, Bali,
Indonesia.