was used to help locate it in the bathymetric data and the 
probable location of the transect line was determined by 
measuring a 50+5-meter distance from the reference coral. An 
XYZ format of the distribution points was exported and 
converted to elevation by inverting the depth values simply by 
multiplying each value with -1. A polygon was created to clip 
the bathymetric data points, which were then interpolated to 
raster using Inverse Distance Weighted (IDW). Neighborhood 
statistics (mean) using a circular neighborhood with different 
radiuses was applied to determine the mean value of pixels at a 
certain radius (1m, 7m, etc.). The neighborhood statistic results 
were subtracted from the original interpolation to determine the 
relative elevation, which were most probably corals if these 
areas are higher than its surroundings 
Pre-processing of underwater videos. The underwater video 
used had good resolution and water clarity to sufficiently 
recognize features of organisms. This video was cut into 
snapshots, which had to have overlapping areas in order to 
reference one image to another. Four approaches for 
registration and georeferencing were applied to the photos and 
were the following: (1) sequential image registration, (2) 
control images to bathymetric data, (3) control images to tape 
distance and (4) uncontrolled mosaic to bathymetric data. The 
first three approaches resulted to extremely warped images and 
mosaics, since the video did not have corrections for pitch, roll 
and yaw, thus the fourth approach (uncontrolled mosaic to 
bathymetric data) was used for this research. Due to this 
difficulty with image registration and georeferencing, the 
snapshots were manually ‘stitched’ together in Adobe 
Photoshop instead, to create a photo-mosaic, which was an 
alternative in order to obtain a dataset with better resolution and 
water clarity (Hill & Wilkinson, 2004). The snapshots used 
were those that were most parallel to the seafloor in order to get 
the most upright views of the features. Each photo was 
subjectively scaled and adjusted to fit with the first image used. 
Since no georeferencing was applied, mosaicking in ENVI 4.3 
and ArcGIS 9.3 cannot be accomplished, thus, it was done 
instead in Adobe Photoshop. Auto blend was applied to the 
uncontrolled photo-mosaic to eliminate the lines made by the 
boundaries of each snapshot. Finally, linear stretching was 
applied using ENVI 4.3. The processes that were applied and 
used on the snapshots were done in order to end up with a 
photo mosaic that may not be accurately made but was useful 
for classification purposes. 
Georeferencing. The locations of the corals in the relative 
bathymetry were determined based on its similarity with 
features present in the mosaicked photos of the transected area. 
Higher values represented the elevated features, which 
suggested that these were the corals, while lower values 
represented non-elevated features (most probably sand or 
rubble). Since the bathymetric data from the MBES had correct 
geographic location, the uncontrolled mosaic was then 
georeferenced based on it, to give the mosaic scale and rotation 
corrections. 33 control points, used to georeference the mosaic 
to the relative bathymetry, were evenly spread out through the 
entire photo gaining a root-mean-square error (RMSE) of 
0.64185. 
Masking: Before the georeferenced mosaic could finally be 
used for classification, a mask for the area outside the said 
image was built and applied on the mosaic such that areas that 
were not useful had a value of zero in order to remove the 
unnecessary parts of the photo during classification. 
Pixel-based image classification: Four classifications were 
used on the unstretched and linear stretched mosaic, namely: 
supervised classification, unsupervised classification, 
unsupervised with supervised classification (also known as 
hybrid classification in remote sensing (Richards & Jia, 2006)) 
and OBIA. The supervised classification requires training 
pixels that have known spectral values and class. There are 
multiple algorithms under supervised classification (Richards & 
Jia, 2006) but the algorithm used in this research was the 
Maximum Likelihood where regions of interest (ROIs) were 
manually chosen and were used as training pixels to guide the 
algorithm in its classification. Five ROI classes were used and 
were the following: dark sand, light sand, dark coral, light coral 
and rubble. Two training classes were assigned for sand due to 
the lighting condition and shadows though linear stretching was 
somehow able to compensate for the attenuation by the water 
column. The spectral separability between chosen pairs of ROI 
for a given file was calculated using the Compute ROI 
Separability option of ENVI 4.3. The values ranged from 0 to 
2.0, which indicate how statistically separate an ROI pair was. 
ROI pairs with pair separation values greater than 1.9 had good 
separability while those having values below 1 had very poor 
separability (ENVI User’s Guide). Due to the very poor 
separability between rubble and dark sand (~0.1823), maximum 
likelihood was ran again using only 4 ROIs, excluding rubble to 
maintain a high pair separation. 
Unsupervised classification determines classes purely on the 
spectral differences of features. The iterative self-organizing 
data analysis (ISODATA) computes for the class means, which 
are evenly distributed in the data space. It then clusters 
iteratively the rest of the pixels using minimum distance 
techniques (Richards & Jia, 2006). ISODATA was used on 
both mosaics, where it was programmed to produce 4 to 5 
classes with five iterations. Combining of classes was applied 
in order to get 3 classes only. 
Hybrid classification is the combination of unsupervised and 
supervised classifications (Richards & Jia, 2006) and was done 
to further explore which method would yield the highest 
accuracy. This was accomplished by creating a 5-iteration 
ISODATA classification set to produce 20 to 40 classes. The 
resulting 40 classes were then manually combined to form a 
classification map with three classes. 
Classification using object-based image analysis: The last 
classification, object-based image analysis, was done using the 
software eCognition 8.7 by Definiens &. The developed rule set 
was done through a Process Tree and had 4 general processes: 
image segmentation, classification, refining of classification 
and exportation. “Multi-resolution segmentation” was chosen 
as the method for image segmentation because the objects of 
interest in an image appear on different scales at the same time 
(Baatz &  Schápe, 2000). Segmentation settings and 
composition of homogeneity are all defined by the user and 
were replicated from eCognition Webinar, except for the scale 
parameter, which is an adjustable quantity that may be modified 
to meet the criteria for adaptability. In eCognition, the scale 
parameter is a measure of homogeneity and as it increases, the 
algorithm allows more merges and in turn, also lets the region 
become larger (Zhang & Maxwell, 2006). The image objects, 
produced during the image segmentation process were initially 
unclassified and went through the rule set for classification and 
refining of classification. The classification used spectral 
conditions (mean of blue and green bands) to remove the black 
boundary surrounding the image. It also used multiple spectral 
conditions (mean and standard deviation of different bands) to 
   
initial 
separa 
to obj 
deviat 
used. 
refine 
neighl 
in ord 
fish. / 
merge 
was 
geore! 
Manu 
refere 
digitiz 
the m 
assign 
expert 
them 
Accui 
43 g 
omiss 
and tl 
dividi 
numb 
does 1 
This i 
errors 
probl 
comm 
pixels 
a clas 
other 
ofac 
pixels 
a cert 
divide 
numb 
produ 
corre 
were 
relate 
produ 
exclu: 
repre: 
mode 
Com] 
the p« 
finall 
benth 
Geor 
regist 
togetl 
uncor 
were 
and a 
Perfo 
on th 
objec 
an a