2004 
  
rated 
—. 
Cn 
— 
Inc.) 
nents 
small 
a 25- 
ing a 
f area 
re no 
tide 
(1) is 
4,) at 
actual 
i an 
:d tide 
(8) 
w tide 
onship 
single 
1 image 
to band 
(9) 
|. Sensor 
metry 
ed the 
nce and 
ological 
signal. 
on their 
ectra to 
yplied to 
(10) 
(b) 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV. Part B7. Istanbul 2004 
where 2 is the scattering angle between incidence (6!,4,) and 
reflection direction (6,4) and expressed as 
cos 4 = cos0, cos0! + sin0' sin0! cos(d, —$4,) . Note that the 
directional parameters (0,0!) have been primed to indicate 
their adjusted value underwater due to refraction effects. 
The G and F define the morphological characteristics of the 
benthic cover that are three-dimensional, specifically corals and 
seagrass. The BRDF for sand is assumed to be Lambertian. 
2.6 Fusion of multisource imagery with varying resolution 
To fuse images, we employ multiresolution decomposition 
algorithm (Gross and Schott, 1998; Piella, 2004). This method 
exploits the fact that the reflectance from the high resolution 
image bears a linear relationship with its equivalent composite 
of image pixels of lower resolution. In order to determine the 
proper relationship, we degraded the higher-resolution satellite 
data to correspond with the pixel size of lower-resolution 
satellite data. We then increase the pixel size of the lower 
resolution image but now weighted according the regressed 
relationship for the nearest band. In this way, the detail of 
objects captured in the higher resolution image is preserved 
while retaining spectral integrity. 
Going through all of these procedures, the intermediate product 
at this stage could now be imagined to be the reflectance 14- 
band image free from inherent effects of the atmosphere, water 
surface conditions and object morphology. 
2.7 Simultaneous fractional cover, classification and depth 
estimation 
Classification of benthic cover is based on the evaluation of 
spectral unmixing results and radiative transfer model after the 
necessary image corrections are accomplished and image fusion 
is attained (Paringit and Nadaoka 2003). The proportions, /; of 
the different benthic cover types are initially deconvolved by a 
non-negative iterative least squares solution with sum-to-one 
constraint (NILSSTOC). Using the estimated proportions for 
the n number of benthic cover, the approximate reflectance is 
computed by radiative transfer model (RTM): 
R(b) =R, (b)exp(-2k,d,) + [=lexp(-2k,4, JS /R (b) (11) 
where R,(b)is the reflectance at a nearby deep area (outside 
the reef). The attenuation coefficient k, also varies only for 
each band. The depth, d. initially given, varies with each sensor 
by the difference d. sd n.) . The total rms (root mean 
Square) error, R^ between the approximate reflectance R(b) 
and the actual image data R(b) value for m number of bands 
are then evaluated. We use a downhill simplex method to 
iteratively vary d , repeat NILLSTOC and RTM, that will lead 
0 a minimal and stable R’, The final product of this step 
therefore will be a set of f, and the estimated average depth d . 
In the classification, benthic cover is assigned according to its 
ecological significance (Edinger and Risk, 2000). This scheme 
I5 adopted because sometimes it is not necessary to assign 
benthic cover with the largest / for a given pixel. Biological 
999 
researchers often regard the presence of a certain important 
habitat as the pertinent cover even if only occurring at a fraction 
physically. 
2.8 Verification and accuracy assessment 
In order to check the consistency of the merged datasets, we 
compared the reflectance spectra of pure benthic cover obtained 
from the image against the spectral data taken from the field. 
We also evaluated the relative differences and/or similarities 
between in-situ reflectance transect and its transect 
representation in equivalent location in the image. We also 
analyzed the classification and bathymetry estimates based on 
confusion matrices and statistical measures of errors 
respectively. 
3. RESULTS AND ANALYSIS 
3.1 Spectral consistency of merged datasets 
Processed data show that there is strong correspondence 
between the image reflectance and the measured field 
reflectance (Figure 3). The additional bands augmented 
appreciably in the recovery of the spectral curves by defining 
spans of abrupt change in high and low absorption points. 
Errors seem to be lower on shorter wavelength ranges (9%) 
especially from targets with naturally high reflectance on the 
VIS range. 
  
         
  
  
  
60 { 
© Seagrass 
D Coral 
50 A Sand £| 
X Algae | 
‘3 O Reef Rock i 
e 40 Sand ^ A 
9 — — Coral A^ 
8 30 - Seagrass A A fb + 
= Algae 
= Reefrocks 
20 
O 9 | 
10 OQ 9 | 
0 == J 
400 500 600 700 800 
Wavelength (nm) 
Figure 3. Spectral signatures of typical reef cover types (lines) 
superimposed with the equivalent reflectance values (marks) 
obtained from each inclusive band of the satellite sensors used. 
As shown in Figure 4, there appears to be a very strong 
correlation of reflectance values between image and in-situ 
transects. The strength of the retrievals are significantly reduced 
(P>0.1) for longer wavelengths particularly the NIR bands. 
       
  
  
  
    
     
  
   
0.26 0.2 
* Reflectance . Reflectance 
. . 
0.24 Ideal line Ideal line 
*. 
30.22 $0.18 
S e e = 
= 2 
3 0.2 ^ Z 
S ve ë 
5018 oe 7 0.16 
E: = /:0.0537x ! 8 
z 0.9481x + 0.0094 z € 3:90.0837x 1 0,0083 
A 0.16 R?=0.9212 = R° =0.9237 
0.14 0.14 
0.14 0.16 0.18 0.2 0.22 0.24 0.26 0.14 0.16 0.18 0.2 
Image transect Reh) Image transect Rib) 
(a) (b) 
Figure 4. Comparison of band 1 (IKONOS band 1) reflectance 
from in-situ transects and image transects values along (a) 
Shiraho Reef: 343 points and (b) Fukido River mouth area: 435 
points.