International Archives of the Photogrammetry, 
Remote Sensing and 
  
32 Analysis of classification results 
Ten (10) different band combinations were subjected for 
classification for 5 coarse types of shallow benthic cover (coral, 
seagrass, sand, algae, reef rock). In addition, we have also 
applied the classification procedure for cumulative number of 
bands in consecutive order of increasing wavelength to assess 
the effect of systematic increase in band number. 
Evaluation of accuracy results indicate that the five band- 
combination (see Table 2) involving bands 1 of IKONOS and 
ASTER, bands 2 of Spot and Landsat, and band 4 of IKONOS 
yields the best result (84.5% ). However, this achievement is 
still comparable (P<0.5) to other five band combinations 
spanning different band ranges, which are almost identical at 
83%. On the opposite end, the classification results from 
spectrally similar bands from the four sensors produced poorer 
results (34%, 30% and 28% respectively, in the order of 
increasing band range), the worst being those bands located in 
the near infrared (NIR: 780-900 nm) range. It is also observable 
that higher spatial resolution still commands considerable 
improvements in overall accuracy except in the case of Landsat 
ETM+ where it is significantly better that SPOT XS and 
ASTER VNIR because of the presence of a visible blue band 
(band centered at 483.2nm). Figure 5 shows the relationship 
between accuracy and increase in the number of bands 
corresponding to wavelength. 
Table 2. Band combinations, classification performance, and 
- 
depth estimation accuracy. 
    
   
    
    
    
  
  
   
  
    
  
Classification 
Accuracy (96) 
Users | Makers Overall 
Depth 
estimation 
rms Error* 
    
   
   
   
Channel number 
   
  
Legend: I-Ikonos, L-Landsat ETM, S-SPOT XS, A-ASTER. 
Number defines channel setting. *Normalized (ratio of rms 
error to actual depth). 
for all classification results, the 
In terms of thematic accuracy, 
areas while there are 
method delivers best accuracy for sandy 
common difficulties encountered for the seagrass and coral 
assification of seagrass beds may be 
attributed to patchy configuration of the meadows, which 
cannot be accommodated by the input signal coming from the 
lower-resolution satellite. The problem for coral 
misclassification arises from confusion in distinguishing them 
from macroalgae classes. This misinterpretation is ascribed to 
the presence of symbiotic algae zooxanthella covering the coral 
itself, which is spectrally similar to macroalgal species such as 
Sargassum sp. and Lobophora sp. (Hedley and Mumby, 2002). 
Overall, classification performed with the spectral unmixing 
classes. The miscl 
Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
    
procedure outlined above is generally better than conventional 
methods (Mumby and Edwards, 2002) applied to individual 
datasets where accuracies are reported to be lower by more than 
5% for coarse habitat mapping. 
100 900 
mz] Band range 
—e— Classification accuracy 800 
_ —A— Depth estimate rms error t E 
d 700 = 
3 600 5 
2 > 
tl S 
5007 
400 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
1237475678 9 19 11:12 13 14 
Number of bands 
Figure 5. Plot of classification accuracy and depth estimate 
RMS error with increasing number of bands of longer 
wavelength. 
3.3 Result of bathymetry estimation 
Figure 6 illustrates the output bathymetry map for Fukido area 
using the processed multisource image. It can be seen the model 
provides a rich topographic detail of a complicated reef system. 
The presence of the small sand cay areas are well-depicted as 
well as the abrupt increase in depth at the interface of the reef 
crest and outer reef flat. Shallow water depths in seagrass and 
seaweeds are also within realistic range. This is a common 
pitfall of  bathymetry mapping found in conventional 
approaches like band ratio or regression lines where depth of 
“darker” bottom cover such as seagrass beds and corals are 
overestimated. 
A 
cree 
    
   
     
     
Figure 6. Result of bathymetry estimation (superimposed on 
IKONOS true-color imager) for Fukido River Mouth area using 
the fused satellite imagery. 
    
With reference to Figure 5, there appears to be direct 
relationship between accuracy of depth estimates and the 
classification correctness. The normalized rms error is reduced 
to 0.19 when all 14 bands are used and achieves most 
enhancement with use of the 10" band (Ikonos band 3). The 
1000 
pe 
in 
di 
al 
Actual Denth (m 
3.4 
Ba: 
a 
inc 
img 
lim 
spa 
and 
Des 
hab 
on | 
inv: 
sea 
chai 
ano; 
cou 
atteı 
and 
Als 
para 
will 
data 
We 
misr 
tech 
anot 
bath 
prec: 
In tl 
reco; 
at di