Anand Mehta
, Eraldo Ribeiro
, Jessica Gilner
and Robert van Woesik
Department of Computer Sciences
Department of Biological Sciences
Florida Institute of Technology
Melbourne, FL 32901
Coral reef characterization, machine vision applications, texture classification, texture segmentation.
The development of tools to examine the ecological parameters of coral reefs is seriously lagging behind
available computer-based technology. Until recently the use of images in environmental and ecological data
gathering has been limited to terrestrial analysis because of difficulties in underwater image capture and data
analysis. In this paper, we propose the application of computer vision to address the problem of monitoring
and classifying coral reef colonies. More specifically, we present a method to classify coral reef images based
on their textural appearance using support vector machines (SVM). Our algorithm uses raw pixel color values
directly as sample vectors. We show promising results on region classification of three coral types for low
quality underwater images. This will allow for more timely analysis of coral reef images and broaden the
capabilities of underwater data interpretation.
The analysis of underwater images is a challenging
computer vision problem with several important ap-
plications in the fields of archeology, military, nav-
igation and biology. In recent years, stimulated by
the current availability of digital optics technology,
the demand for solutions for underwater image analy-
sis has increased significantly. In this paper, we de-
scribe a supervised texture classification method for
automatic classification of coral reef images. The
work reported here is part of a ongoing research ef-
fort aiming to develop an image-analysis system to
monitor temporal changes of coral colonies on coral
reefs. Our classification method uses a support vector
machine(SVM) classifier (Vapnik, 1995). Depend-
ing on the application, support vector machines have
several advantages when compared to classical super-
vised classification methods such as maximum likeli-
hood. In the particular case of texture classification,
support vector machines allow for very good class
separation even when the size of the feature vectors
is large and the number of training samples is limited.
Another advantage is that it allows for the use of raw
image data as feature vectors which is specially inter-
esting for natural random textures where geometrical
features and descriptors are difficult to obtain (Kim
et al., 2002; Li et al., 2003). This is particularly rel-
evant to the classification of coral reef images which
present rich patterns of colors, shapes and textures.
Coral reefs are among the most complex natural
systems on earth, both structurally and biologically.
Indeed, high structural topographic complexity allows
coral reefs to support more species than any other ma-
rine system. Coral reefs have recently risen to global
prominence in terms of their capacity to act as early
warning indicators of global climate change. They are
accurate proxies of thermal stress events, where corals
visibly pale (bleach), which often leads to mortal-
ity and relative shifts in species composition (Hoegh-
Guldberg, 1999). Yet, quantifying diversity and rela-
tive abundance of species on coral reefs is time con-
suming. Most serious are the delays in data compil-
ing and synthesis, where twelve months is not uncom-
In this paper, we apply computer vision to the
problem of monitoring and classifying coral reef
colonies. Our texture classification approach does not
require any explicit feature extraction. The feature ex-
traction is implicitly performed by the support vector
classifier. As a result, no pre-processing is required,
and the pixel raw data is directly used for both training
Mehtaa A., Ribeiro E., Gilner J. and van Woesikb R. (2007).
In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IU/MTSV, pages 302-305
and classification. This is possible since support vec-
tor machines implicitly perform feature extraction by
means of a kernel which is defined by a dot product
of two non-linear mapped patterns. Support vector
machines are binary classifiers in essence but multi-
class separation can be achieved by means of the one-
against-all decomposition procedure. This is equiva-
lent to decomposing the multiclass problem into mul-
tiple independent binary classification tasks. Alterna-
tively, better results can be obtained when multiclass
predictors are trained directly as described in (Cram-
mer and Singer, 2001). The classification presented
in this paper is a binary one.
The first method used to quantify benthic coral reef
organisms was borrowed from plant ecologists. It in-
volved the use of line intercept transect method (Loya,
1972), where fiberglass lines are laid over a given
reef and meticulous measurements are made, to the
nearest centimeter, of each benthic (bottom dwelling)
coral reef organism while the researcher is underwa-
ter. Only small spatial tracks of reefs can be ex-
amined using this technique. Other methods involve
similar small scale data gathering exercises, involving
quadrats and belt transects. Capturing subsections of
reef on still and video digital images has increased the
area of observation, but there are still delays in data
gathering and compilation due to manual processing.
There has been a considerable increase in the ef-
fort to provide remote sensing solutions to the prob-
lem of coral reef monitoring for both aerial and un-
derwater imagery. A review of the application of
both acoustic and optical imaging devices in analyz-
ing the physiology, behaviour, and interactions be-
tween benthic species can be found in (Solan et al.,
2003). While progress has been made in the use of
remote sensors, namely airborne scanners and satel-
lites (Mumby et al., 2004), ground truth comparisons
have revealed high error estimates rarely surpassing
60% accuracy (Hedley et al., 2004).
Solutions based on satellite imagery have lim-
ited resolution and are not suitable for monitoring
small variations in both shape and color of coral
colonies. Imaging coral colonies directly using under-
water video and high-resolution images represent a
contemporary consideration of coral reef assessment.
Our analysis is focused on underwater imagery. Our
classification method does not require the estimation
of any intermediary feature vector or histogram. The
general visual appearance of the corals is implicitly
extracted in the support vector machine classification
process. Our work follows the line of texture clas-
sification as described by (Kim et al., 2002) and (Li
et al., 2003). Li et al (Li et al., 2003) use translation-
invariant features generated from the discrete wavelet
frame transform together with a fusion scheme based
on multiple support vector classifiers, each with a dif-
ferent setting of the kernel parameter. Kim et al (Kim
et al., 2002) use raw pixel data as input for the support
vector classifier of textures. Their results are demon-
strated for sets and mosaics of Brodatz textures.
Support vector machines (SVM) are binary classifiers
that estimate the optimum separating hyperplane that
maximizes the margin between two classes. The mar-
gin can be defined as the distance of the closest point,
in each class, to the separating hyperplane. In sta-
tistical learning theory, this is equivalent to perform-
ing structural risk minimization on a nested set struc-
ture of separating hyperplanes (Vapnik, 1995; Burges,
Given a set of training examples (x
, y
) R
1}, i = 1, . . . , l, the objective is to determine the
function f : R
1}, from a class of functions,
such that f will correctly classify new examples
(x, y), i.e., f (x) = y, which were generated under the
same underlying probability distribution p(x, y) as the
training data. Support vector machines use hyper-
planes for class separation:
(w · x) + b = 0 w R
, b R, (1)
and the corresponding decision function is given by:
f (x) = sign[(w · x) + b] (2)
Solving for the optimal separating hyperplane
w consists of finding the solution of a constrained
optimization problem using quadratic programming,
where the optimization criterion is the width of the
margin between the classes (Vapnik, 1995; Burges,
1998). The optimal separating hyperplane can be rep-
resented as a linear combination w =
of a sub-
set of the training examples that lie on the margin.
These training examples or patterns carry all relevant
information about the classification problem and they
are called support vectors. Once the support vectors
are estimated, classifying a new test pattern x is
done using the following expression:
f (x) = sign[
(x · x
) + b] (3)
where the sign of f (x) determines the class member-
ship of x.
Usually, general pattern classification problems
cannot be solved using a linear classifier such as a sep-
arating hyperplane, and the classification procedure
should allow for non-linear separating surfaces. A
crucial property of the SVM is that both the quadratic
programming problem and the final decision function
depend only on dot products between the patterns.
This property allows the classification procedure to be
generalized to cases when the patterns are not linearly
separable. In this case, the decision function is a non-
linear function of the data. The method consists of
mapping the data into a higher dimensional dot prod-
uct space where they are considered linearly separa-
ble. Once the mapping is performed, SVM can be ap-
plied to the new space to calculate the decision bound-
ary. This linear boundary leads to a non-linear sepa-
ration surface in the original space. The mapping pro-
cedure is based on the Cover’s theorem (Cover, 1965)
that shows that a multidimensional space consisting
of nonlinearly separable patterns can be transformed
into a new feature space where the patterns are lin-
early separable with a high probability. The theorem
is valid under the conditions that the transformation
is nonlinear and the dimension of the feature space is
high. The kernel function k(x, y) = Φ(x) · Φ(y) is in-
troduced and does not require explicit knowledge of
Φ(.) The solution has the form:
f (x) = sgn[
k(x, x
) + b]
= sgn[
(Φ(x) · Φ(x
)) + b]. (4)
Table 1 lists the three most commonly used kernel
functions in SVM classification.
Table 1: Kernel functions used for SVM classification.
kernel name function
Polynomial k(x
, x
) = (x
· x
Radial basis function k(x
, x
) = e
Sigmoid kernel k(x
, x
) = tanh(x
· x
) Θ)
In this section, we describe the details of our clas-
sification algorithm. The algorithm consists of a di-
rect application of the support vector machine classi-
fier using raw pixel data as sample vectors. We chose
the radial basis function kernel to perform the experi-
ments where σ is manually selected and kept fixed for
all experiments. In our trials, the radial basis func-
tion kernel performed better than the polynomial (Ta-
ble 1). We plan to investigate approaches for the se-
lection of σ as suggested by (Muller et al., 1997).
We commenced by extracting a number of N × N
subregions of previously labeled coral images. Since
our current classification approach is a supervised
one, we selected the images to match a particular class
of coral reef that we wanted to classify. The pixel val-
ues inside these regions became the sample vectors to
be fed into the classifier for the training stage. During
the training phase, the support vectors were identified.
Once the support vectors were at hand, we classified
each pixel in an image according to the class predicted
by the support vector machine classifier. This is a sim-
ple algorithm that allows for very good texture classi-
fication results of coral reef images.
In this section, we present results for classification
of coral reef images. We applied the support vector
machine method for the classification of three types
of coral: corymbose Acropora, branching Acropora,
and tabulate Acropora. The classifier was trained for
about 100 subregions of each coral type. Each sub-
region had a size of 25 × 25 square pixels. The clas-
sification results were obtained for two classes only,
but multiclass classification can be obtained using a
one-against-all approach, as mentioned earlier. We
make use of the support vector machine implemen-
tation provided by the libsvm library by Chang and
Lin (Chang and Lin, 2001). Figures 1, 2, and 3 show
both the original test images(a) and the results of the
texture classification(b). Our algorithm achieved a
95% correct classification for all classified images.
However, several small regions were still misclassi-
fied. The main reason seems to be the lack of a proper
model for the underwater illumination. We plan to ap-
proach this problem in future implementations. The
(a) (b)
Figure 1: Coral class ification results. (a) corymbose Acrop-
ora, and (b) resulting labeled regions.
VISAPP 2007 - International Conference on Computer Vision Theory and Applications