Image-Based Coral Reef Classification and Thematic Mapping

Abstract

This paper presents a novel image classification scheme for benthic coral reef images that can be applied to both single image and composite mosaic datasets. The proposed method can be configured to the characteristics (e.g., the size of the dataset, number of classes, resolution of the samples, color information availability, class types, etc.) of individual datasets. The proposed method uses completed local binary pattern (CLBP), grey level co-occurrence matrix (GLCM), Gabor filter response, and opponent angle and hue channel color histograms as feature descriptors. For classification, either k-nearest neighbor (KNN), neural network (NN), support vector machine (SVM) or probability density weighted mean distance (PDWMD) is used. The combination of features and classifiers that attains the best results is presented together with the guidelines for selection. The accuracy and efficiency of our proposed method are compared with other state-of-the-art techniques using three benthic and three texture datasets. The proposed method achieves the highest overall classification accuracy of any of the tested methods and has moderate execution time. Finally, the proposed classification scheme is applied to a large-scale image mosaic of the Red Sea to create a completely classified thematic map of the reef benthos.

1. Introduction

For many, the term “remote sensing” is synonymous with satellite imagery or perhaps satellite and aerial imagery. Remote sensing, thus defined, implies a platform (overhead), a technology (usually passive optics) and a spatial scale (m to km pixels, regional to global extent). More broadly, however, remote sensing includes any method of detection or measurement made without directly contacting the object of study. Several technologies that are relevant for coral reef remote sensing, such as single- or multi-beam sonars, for example, fall under a more inclusive definition of remote sensing.

Underwater imagery is another remote sensing technology that has been used in virtually every aspect of marine science. In particular, images of the seabed have become vital tools for exploration, archaeology, marine geology and marine biology, among other fields. Underwater images, like terrestrial, hand-held images, tend to be overlooked even with a broad definition of “remote sensing”, yet they complement satellite, airborne and ship-based sensors by providing information at finer spatial resolution (< mm to cm scale pixels), albeit over a smaller area per image or per unit time of data collection.

The high spatial resolution afforded by underwater imagery of the seabed is fundamentally different from satellite, aerial and ship-borne sensors in the sense that images taken close to the seabed may be categorized as methods of direct remote sensing for the study of coral reef biodiversity (Turner et al. [1]). Direct remote sensing means that the actual individuals being studied (benthic organisms in this case) can be resolved. Overhead imagery and acoustic mapping, in contrast, are indirect methods for study in coral reef biology and ecology, because, rather than resolving individual organisms, they quantify proxies of biodiversity or they correlate with in situ measurements (Turner et al. [1]).

Direct remote sensing of the seabed complements indirect methods of coral reef remote sensing, because it allows different types of measurements to be made. For example, species identification for many benthic organisms is possible from underwater images, so species diversity indices can be calculated. Sizes of individual coral colonies can be measured and tracked through time. The fate of individuals impacted by bleaching or disease can be tracked through time. Many other uses for the imagery can be imagined, but in general, such measurements are infrequently used, due to two obstacles. First, until recently, it was difficult to obtain large numbers of seabed images across large enough areas to create statistically viable sample sizes. Second, extraction of quantitative information from seabed images is a laborious task that requires a significant time commitment of expert analysts.

Two technological advancements in the early 2000s qualitatively changed the acquisition of underwater imagery. One development was the widespread availability of digital, consumer-level still and video cameras. Digital cameras, along with associated advancements in batteries and storage, have largely removed constraints on the quantity of images that can be acquired. In 2000, a SCUBA diver with a Nikonos underwater camera was limited to 36 photographs per dive, but in 2013, the same diver with a digital camera can capture literally thousands of images in a single dive. The second development was the increased sophistication and availability of platforms, such as remotely operated vehicles (ROVs), autonomous underwater vehicles (AUVs) and towed bodies. These new platforms allow image capture at depths and across spatial scales that are not possible with divers or practical (for expense reasons) with manned submersibles.

Advancements in underwater image acquisition have largely removed the first obstacle to fully exploiting seabed images for coral reef remote sensing, but, ironically, they have exacerbated the second. It is now easy and common to collect more imagery during reef surveys that can be practically analyzed by hand. Nevertheless, images of reef surveys are useless to understanding reef processes unless examined by an expert analyst. Software packages, such as CPCe [2], have facilitated the transfer of information from imagery to analysis; however, humans are still “in the loop.”

An automated process for classifying underwater optical imagery would enhance the study of coral reef dynamics and processes by reducing the analysis bottleneck and allowing researchers to take full advantage of large underwater image datasets. Efforts to automate classification have been made for well over a decade [3], but no single algorithm is yet widely accepted as robust. A few of the most challenging obstacles to classification accuracy in coral reef environments include: significant intra-class and inter-site variability in the morphology of benthic organisms [4], complex spatial borders between classes on the seabed, subjective annotation of training data by different analysts, variation in viewpoints, distances and image quality, limits to spatial and spectral resolution when trying to classify to a free taxonomic scale, partial occlusion of objects due to the three-dimensional structure of the seabed, lighting artifacts due to wave focusing [5–7] and variable optical properties of the water column.

The existing methods for coral reef image classification have been tested on unique image datasets within the papers in which they were published; their utility has not been demonstrated on common datasets. Without comparison on standard datasets, it is impossible to assess the relative effectiveness and efficiency of these methods: Pizarro [8], Beijbom [4], Marcos [9] and Stokes and Deane [10]. All the previous algorithms used fixed classification schemes for all types of underwater optical image datasets. In general, however, characteristics, such as the size, number, resolution and color information of the images, vary from dataset to dataset. Therefore, a fixed recipe for automated classification for all types of optical image datasets might not always provide the best solution.

The goals of this work are to develop an improved method for classifying underwater optical images of coral reefs and to compare the proposed method with existing alternative approaches. The main difference of this new method, relative to previous algorithms, is that classification schemes should be malleable depending on the characteristics of each dataset (e.g., the size of the dataset, number of classes, resolution of the samples, color information availability, class types, etc.). In the proposed method, different combinations of features and classifiers are tuned to each dataset prior to classification to increase classification accuracy. Our new method is compared with the existing algorithms for accuracy and time requirement on six image datasets (Section 4). Finally, image classification methods were applied to a composite mosaic image to demonstrate the applicability of this approach for images that cover areas larger than a single image.

2. Review of Existing Algorithms for Automated Benthic Classification Using Optical Imagery

A short summary of existing algorithms for automated benthic classification using optical imagery is given in

Table 1. A brief summary of methods classifying benthic images. All the methods mentioned here are specific to coral reef habitat classification. The methods in bold are used in Sections 5 for performance comparison and are referred to by the underlined author’s names. The last column, N, contains the number of classes used for testing in each method, as reported in their respective papers.

Authors	Features	Classifiers	N
Beijbom [4]	Maximum response (MR) filter bank	Library for support vector machines (LIBSVM)	9
Padmavathi [11]	Bag of words using scale-invariant feature transform (SIFT) and kernel principal component analysis (KPCA)	Probabilistic neural network (PNN)	3
Stokes & Deane [10]	Color: (RGB histogram) Texture: discrete cosine transform (DCT)	Probability density weighted mean distance (PDWMD)	18
Marcos [12]	Texture: local binary pattern (LBP) Color: normalized chromaticity coordinate (NCC) histogram	Linear discriminant analysis (LDA)	2
Pizarro [8]	Color: normalized chromaticity coordinate (NCC) histogram Texture: bag of words using scale-invariant feature transform (SIFT), Saliency: Gabor filter response	Voting of the best matches	8
Mehta [13]	Pixel intensity	Support vector machines (SVM)	2
Gleason [14]	Multi-spectral data Texture: grey level co-occurrence matrix (GLCM)	Distance measurement	3
Johnson-Roberson [15,16]	Texture: Gabor filter response Acoustic	Support vector machines (SVM)	4
Marcos [9]	Color: normalized chromaticity coordinate (NCC) histogram Texture: local binary pattern (LBP)	3-layer feed-forward back projection neural network	3
Clement [17]	Texture: local binary pattern (LBP)	Log-likelihood measure	2
Soriano [18]	Color: normalized chromaticity coordinate (NCC) histogram Texture: local binary pattern (LBP)	Log-likelihood measure	5
Pican [3]	Texture: grey level co-occurrence matrix (GLCM) self-organizing map	Kohonen-Map	30

, with entries in bold corresponding to the published methods tested in Sections 4 and 5 of this paper.

Only published methods that exclusively used optical imagery and produced accurate multi-class classifications were chosen for comparison against the proposed classification scheme. Among these works, Marcos [9] used mean hue-saturation value (HSV) or normalized chromaticity coordinate (NCC) color features, which in many cases, are not discriminative enough for classification among classes. In the work of Pizarro [8], an entire image is classified as one class. Therefore, within-image heterogeneity cannot be classified or quantified. The method by Stokes and Deane [10] is time-efficient; however, the weighting of the color and texture features during feature vector creation were done manually. Beijbom [4] used multiple patch sizes; therefore, features in the patch are calculated several times, which is redundant and time-consuming.

Much additional work has been done on texture classification for other applications besides benthic imagery. Two papers in particular are worth mentioning here, as they reported the highest classification accuracy on standard texture datasets. Caputo [19] used filter bank outputs with a support vector machine (SVM) classifier that had a kernel based on the chi-square distance. This method worked very effectively on low-resolution texture datasets. Zhang [20] used an approach that represents images as distributions (signatures or histograms) of features extracted from a sparse set of key points locations and trained a SVM classifier with kernels based on the earth mover’s distance (EMD) and the chi-square distance. In Section 5, we compare the performance of our method on texture datasets with these two methods as well.

3. Proposed Algorithm

Our proposed method consists of seven steps to perform training and classification of underwater images. This method is a supervised classification scheme, which assumes that enough (at least 15 examples for training of each class) labeled training data is available. A unique feature of our approach is that it is based on a variable implementation scheme. Under this scheme, different options are available for each of the main processing steps. The configurable scheme allows the tuning of the processing pipeline to the characteristics of different datasets and leads to performance gain. The steps of the method are given in Figure 1 and discussed in brief below.

3.1. Step One: Image Enhancements

The image enhancement step aims to improve image comparability and, consequently, classification accuracy. Image enhancement improves the visibility of image features by counteracting some of the effects of the medium, such as light scattering, color shifts and blurring. The image enhancement step contains one sub-step that is always used (CLAHS) and three optional sub-steps to be used if color images were acquired in the dataset.

The first image enhancement sub-step (optional) is color correction using color difference optimization. This sub-step is optional, because it requires the presence of known color markers in the dataset [4] (Figure 2). Images from most cameras have a non-linear response to radiance. However, modifying color is more suited to the linear domain, as this operation is linear. Therefore, the color difference optimization involves the conversion of the camera response to a linear domain, followed by a color correction and a conversion back to the original non-linear domain. The nonlinear to linear red-green-blue (RGB) transformation and its inverse u e the following equations (for the green and blue channels, the same equation applies with different notations). Here, R represents the nonlinear red, and r represents the linear red. The gamma value (γ), in general, varies from 2.2 to 2.4, depending on the compression scheme of the camera. For JPEG, the gamma (γ) is normally 2.4. This paper used a value of 2.4 for gamma for all experiments, where applicable. s is the inverse of γ:

$$\begin{array}{l}r={\left(\left(\left(R/R_{\text{max}}\right)+0.055\right)/1.055\right)}^{\gamma }\\ \hspace{0.17em}\hspace{0.17em}R=0.5+255\times (1.055\times r^s)-0.055\end{array}$$

The coefficient for the red color channel modification is computed with the following equation:

$$C_R=\frac{{\alpha }_{R1}{\tilde{R}}_1{R}_1+{\alpha }_{R2}{\tilde{R}}_2{R}_2+{\alpha }_{R3}{\tilde{R}}_3{R}_3}{{\alpha }_{R1}{R}_1^2+{\alpha }_{R2}{R}_2^2+{\alpha }_{R3}{R}_3^2}$$

Here R̃₁ is the ideal red component value of the first marker, R₁ is the red component of the first marker of the working image, α_R1 is the corresponding weight of individual color channels and C_R is the red correction factor. In our case, we used the same weights for all three channels. However, since the red color attenuates very fast in the underwater environment, α_R1 could be given less weight than the other channels when applied to different data. Once computed, the red correction factor C_R is multiplied with the linear version of the red color channel to obtain an approximately corrected version.

The second image enhancement sub-step, (mandatory) is the Contrast Limited Adaptive Histogram Specification (CLAHS) [21]. CLAHS locally enhances the contrast of images by dividing the image into several subregions and by transforming the intensity values of each subregion independently to comply with a specified target histogram. This method works very effectively for underwater images of any size or resolution. For our implementation, we divided the entire image into image patches, which ranged in size from 64 × 64 to 256 × 256 pixels, depending on the types of classes present in the dataset. Then, each image patch was further divided into 4 × 4 subregions. CLAHS was applied to the subregions within each patch.

The third and fourth sub-steps are optional, as they require color images. The third image enhancement sub-step (optional) attempts to reduce the color difference caused by different lighting conditions in different images in the dataset using the comprehensive image normalization method [22]. With this normalization, the color features present in the images become more consistent, making color a better feature for classification throughout the data.

Finally, the fourth image enhancement sub-step (optional) is color channel stretching. In this sub-step, for each individual color channel, we determine the 1.5% and 98.5% intensity percentile, subtract the lower from all intensities in that channel and divide the result by the upper intensity percentile.

3.2. Step Two: Feature Extraction

Gehler et al. [23] showed that a combination of several texture and color features leads to better image classification results than any single type of feature. Our proposed method uses the Gabor filter response [24], grey level co-occurrence matrix (GLCM) [25–27] and completed local binary pattern (CLBP) [28] as texture descriptors and the opponent angle and hue color histograms (optional) as color descriptors [29] to define the features of each image patch to be used in the classification. This combination was chosen to represent core texture and color characteristics of given images. In the case of the Gabor filter, by defining four sizes and six orientations, we obtain 24 images of filter response. Using mean and standard deviation for each of these 24 images, a feature vector of 48 values is created. For GLCM features, we use the 22 features [25–27], as listed in Appendix 1. For CLBP features, we use the rotation invariant format, resulting in a histogram of 108 bins when concatenated for three window sizes of eight, 16 and 24 pixels. The opponent angle and hue color histogram features each use 36 bins, for a total of 72 features.

3.3. Step Three: Kernel Mapping

Kernel mapping is used to project the feature vectors to linearly separable feature space. We use chi-square and Hellinger kernels in this step. In the histogram features (in our case, the CLBP, opponent angle and hue histograms), the rarest bins contain more discriminative information, because most of the high frequency bins correspond to the background pixels. The chi-square and Hellinger kernels (also known as the Bhattacharyya coefficient) emphasis the low frequency bins (

Table 2. Chi-square and Hellinger kernel functions. Here, h and h′ are normalized histograms, where h′ is derived from h with first order differentiation. k is the kernel function, γ is the regularization coefficient, and i and j corresponds to histogram bin index.

Kernel Name	Formulation
Chi-square	$k(h,h^{\prime })=\text{exp}\left(-\frac{1}{\gamma }\sum _j\frac{{\left(h_j-{h\prime }_j\right)}^2}{h_j+{h\prime }_j}\right)$
Hellinger	$k(h,h^{\prime })={\sum }_i\sqrt{h(i)\times h^{\prime }(i)}$

). Also, L1 normalization is used to rescale all the features, so that they are comparable.

3.4. Step Four: Dimension Reduction (Optional)

This step performs a mapping of the data to a lower dimensional space in such a way that the variance of the data in the low-dimensional representation is maximized. With lower dimension (depending on the dataset; in our case, it was almost two-thirds of the original dimension), the complexity of the method reduces, as well as the hardware requirements. For highly correlated datasets of moderate size (datasets with training samples more than 5,000 and less than 12,000 are considered as moderate ones), it was found that reducing the dimension of the feature space using principal component analysis (PCA) or PCA with Fisher kernel mapping [30] increases the accuracy when classifying with k-nearest neighbor (KNN). Datasets where classes have high inter-class variability and low intra-class variability are referred to as highly correlated. In such cases, we transform the feature vector with PCA using exhaustive search over all possible dimensions to find the optimum dimension where the classification accuracy is maximum. This step is optional, as it is mainly used for reducing hardware requirements.

3.5. Step Five: Prior Settings

All of the classification methods used in the above scheme requires an estimate of the probability that an image patch will fall into any one of the defined classes. Two choices are possible for estimating this prior probability: (1) an existing estimate of the actual class frequency, such as the frequency within the training data, or (2) by assuming that an image patch has equal probability of falling into any of the defined classes. In natural underwater images, the estimated class frequency tends to produce better results than equal probability, because classes are seldom evenly populated [31]. For the data in this paper, class frequency is calculated from the training set.

3.6. Step Six: Classification

Our algorithm uses one of four types of classifiers depending on the characteristics of the data: support vector machine (SVM), k-nearest neighbor (KNN), multilayer perception back propagation neural network (NN) and probability density weighted mean distance (PDWMD). These methods are found through experimentation to be performing better than other available methods for classifying underwater optical imagery. Directions on how to select the appropriate classifier for a given datasets are presented in Section 5.

3.7. Step Seven: Map Classification

The last step of our approach, referred to as thematic mapping, applies the image classification to large area optical maps of the benthos. These large-area representations (commonly known as image mosaics) are obtained by registering and blending large sets of individual images acquired at close range [32–34].

The first sub-step to create a classified mosaic image consists of using a sliding window (of the same size as the training patches) to segment the image mosaic. The shifting of the window depends on the segmentation resolution desired. Large shifts provide coarser segmentation, lower resolution and lower computation expenses and vice versa. We used 16 pixels shifts for the particular example presented in this work.

Each result has a confidence level, which depends on the distance from the class boundary created by chosen classifier. Class labels are assigned if the results exceed a confidence threshold (the confidence threshold is arbitrarily chosen); otherwise, the patch is assigned a background label. After creating the initial segmentation using classification with the sliding window, the second sub-step involves the use of a morphological filter to check for consistency with the neighborhood surrounding each patch. In general, one object should be positively classified several times, as long as the object occupies a significant portion in the window and the shift size is much less than the patch size. The underlying assumption of morphological filtering is that each classified patch should have at least two patches in the eight-patch-neighborhood classified as the same class. Otherwise, the classified patch is re-assigned to the most prominent class label in the eight-patch-neighborhood. A potential drawback of this approach is that the final classification results may tend towards dominant, contiguous-cover classes reducing the representation of rare classes.

3.8. Parameter Tuning

There are several parameters in each option of the proposed classification framework that need to be tuned beforehand. To select the parameters of these options, exhaustive search was done on underwater optical image datasets (we used the EILAT (Eilat dataset), Rosenstiel School of Marine and Atmospheric Sciences (RSMAS), Moorea labeled corals (MLC) 2008 and EILAT 2 datasets). Once the parameters of all the sub-steps in the proposed method are tuned using the mentioned datasets, those specific parameters can be used for other underwater optical image datasets.

4. Methodology for Evaluating the Classifiers

4.1. Choosing Different Configurations

A unique feature of our classification framework is the ability to use different configurations, i.e., different combinations of sub-steps with different parameter values to suit the needs of a particular image dataset. Therefore, determining which of the various options to use in each step was critical. A series of experiments were conducted varying the choice of options within each step, while keeping the options for the other steps constant. In all the experiments, for each possible configuration, we used 10 crossfold random validation, meaning that 90% of the image patches were used for training and 10% were used for testing, over 10 iterations. With crossfold random validation, we reduced the possibility of having results biased by over-fitting [35].

To determine the best combination of image enhancement options, i.e., step 1, we used color correction for datasets containing color reference markers in the image. To select among the CLAHS, normalization and color channel stretching options, we ran the same classification with those different options, keeping the other steps fixed as follows: opponent angle and hue channel histograms, CLBP, Gabor filter response and GLCM in the feature extraction step, L1 normalization in the kernel mapping step, PCA in the dimension reduction step, class frequency prior settings and the KNN classifier.

The best combination of feature extraction options, i.e., step 2, was determined by varying the use of hue and opponent angle color histograms, CLBP, Gabor filter response and GLCM, while keeping the options for the other steps set as follows: CLAHS in the image enhancement step, L1 normalization in the kernel mapping step, PCA in the dimension reduction step, class frequency in the prior probability settings step and the KNN classifier.

To determine the best combination of kernel mapping options, i.e., step 3, we varied the use of the chi-square kernel, Hellinger kernel and L1 normalization, while keeping the options for the other steps set as follows: CLAHS in the image enhancement step, opponent angle and hue channel histograms, CLBP, Gabor filter response and GLCM in the feature extraction step, PCA in the dimension reduction step, class frequency in the prior probability settings step and the KNN classifier.

Finally, to determine the best combination of dimension reduction and classifier options, i.e., steps 4 and 6, we varied the use of PCA and the Fisher kernel with each of the four classifiers, while keeping the options for the other steps set as follows: CLAHS in the image enhancement step, opponent angle and hue channel histograms, CLBP, Gabor filter response and GLCM in the feature extraction step and L1 normalization in the kernel mapping step.

In the Section 5.5, our insights on how to decide on what configuration to use based on the properties of any new dataset are discussed. The insights are backed up by the results presented in Section 5.1. For our method, we are selecting among the options proposed in the framework (Figure 1) according to the characteristics of the dataset (e.g., size, number of classes and resolution of the samples, color information availability, class types and so forth). Since there is no single configuration that works best in all datasets, it is proposed to select appropriate options from the framework and use the tuned parameters of the methodologies to get the best performance. We tuned the parameters for all the methodologies using the underwater optical image datasets described in Appendix 2: EILAT, RSMAS, MLC 2008 and EILAT 2. As we are using these previously tuned parameters, no additional time would be required in tuning parameters for a new benthic habitat dataset. The datasets are summarized first in Section 4.2, then the methods for performance evaluation and comparison and thematic mapping in Sections 4.3–4.4, respectively.

4.2. Datasets

The relative performance of published classification algorithms has not been tested on common datasets. To address this issue, our proposed classification scheme and four other existing algorithms were tested against six image datasets that represent a variety of challenging environments. A seventh dataset is used along with datasets 1–3 to guide selection of the various options available in our proposed framework. Finally, an eighth dataset is used to illustrate the potential for generating thematic maps of large contiguous areas of the seabed using this automated classification framework. A brief summary of the image datasets used in this work is presented in

Table 3. A brief summary of the image datasets used in this work. N represents number of patches in each datasets. Detailed descriptions with sample patches are given in Appendix 2.

Name	Classes	N	Resolution	Color
EILAT	Sand, urchin, branches type I, brain coral, favid coral, branches type II, dead coral and branches type III	1,123	64 × 64	Yes
RSMAS	Acropora cervicornis, Acropora palmata, Colpophyllia natans, Diadema antillarum, Diploria strigosa, Gorgonians, Millepora alcicornis, Montastraea cavernosa, Meandrina meandrites, Montipora spp., Palythoas palythoa, Sponge fungus, Siderastrea siderea and tunicates	766	256 × 256	Yes
MLC 2008	Crustose coralline algae, turf algae, macroalgae, sand, Acropora, Pavona, Montipora, Pocillopora, Porites	18,872	312 × 312	Yes
UIUCtex	bark I, bark II, bark III, wood I, wood II, wood III, water, granite, marble, stone I, stone II, gravel, wall, brick I, brick II, glass I, glass II, carpet I, carpet II, fabric I, paper, fur, fabric II, fabric III and fabric IV	1,000	640 × 480	No
CURET	61 texture materials imaged over varying pose and illumination, but at constant viewing distance.	5,612	200 × 200	No
KTH-TIPS	sandpaper, crumpled aluminums foil, Styrofoam, sponge, corduroy, linen, cotton, brown bread orange peel, cracker B	810	200 × 200	No
EILAT 2	Sand, urchin, branching coral, brain coral and favid coral	303	128 × 128	Yes
Red Sea mosaic	Sand, urchin, branching coral, brain coral, favid coral, background objects	73,600	3,257 × 5,937	Yes

. Appendix 2 contains a more detailed description of the various datasets. Figure 3 shows an example of images patches from the EILAT dataset.

4.3. Criteria for Evaluation and Comparison with Other Methods

An error matrix and precision recall curve are used to evaluate the classification quality of each classification method. The error matrix quantifies the accuracy of each classified category as a percentage based on the total number of validation points in each category [36]. The overall accuracy (OA) of the classified dataset is derived from the error matrix as the sum of the number of correctly classified tested examples divided by the total number of tested examples. The precision-recall curve is computed by varying the threshold on the classifier (from high to low) and plotting the values of precision against recall for each threshold value. The average precision (AP) summarizes the precision-recall curve by measuring the area under the curve.

To compare the proposed framework, four underwater image algorithms (Pizarro [8], Beijbom [4], Marcos [9] and Stokes and Deane [10]) and two state-of-the-art texture classification methods (Caputo [19] and Zhang [20]) are used. Each algorithm is implemented as closely as possible to the original papers (the parameters are tuned with exhaustive search). The first four algorithms (initially developed for underwater imagery) are used to classify all six of the test datasets. The two texture-classification algorithms are only used to classify the three texture datasets. The results for each algorithm are evaluated in terms of the overall accuracy and average precision of the classification. The time required to execute each method is reported in Section 5.3. The effect of changing the number of training samples in the datasets is also evaluated in Section 5.3.

4.4. Mosaic Image Classification

Finally, our proposed algorithm and first four algorithms are implemented on a mosaic image to create a thematic map of a contiguous section of the seabed covering an area larger than a single image. The mosaic is created using the methods by [32–34] from survey images (283 high-resolution digital still images and rendered at 1 mm per pixel) of shallow water coral reefs in the Red Sea, near Eilat [37]. The mosaic covers an area of 19.2 square meters.

To train the classifier for the mosaic, we use entire 1,123 image patches in the EILAT dataset. The sliding window used to classify the mosaic is of size 64 × 64 pixels (same as the size of training image patches). The sliding window is moved with 16 pixels shift (either horizontally or vertically) per iteration. In total 73,600 image patches are classified within the mosaic image.

To examine the effects of morphological filtering, we created a ground truth by manually classifying the Red Sea mosaic with help of experts. We use this ground truth for a qualitative assessment of the proposed method and the other existing algorithms. The classification results (color coded) on the mosaic are presented and discussed in the Section 5.4.

5. Results

Having presented, in the previous section, the criteria and methods for evaluating the classifiers, this section now presents experimental results. These results allow for assessing the impact of different configurations, for evaluating and validating the proposed method, for comparing against the other selected approaches and for making recommendations on the best configuration for future datasets.

5.1. Evaluation of Different Configurations

Among the image enhancement options, (step 1), CLAHS alone produced the best results for the EILAT, RSMAS and EILAT 2 datasets. CLAHS with color channel stretching produced the best results for the MLC 2008 dataset (

Table 4. Overall accuracy (%) with different image enhancement options as evaluated with the EILAT, RSMAS, EILAT 2 and MLC 2008 datasets. The configurations for the other steps are fixed as follows. Feature extraction: completed local binary pattern (CLBP), grey level co-occurrence matrix (GLCM), Gabor filter response, opponent angle and hue channel histogram; kernel mapping: L1 normalization; dimension reduction: principal component analysis (PCA); prior: class frequency; classifier: k-nearest neighbor (KNN). CLAHS and NA stand for ‘contrast-limited adaptive histogram specification’ and ‘not applicable’, respectively.

	EILAT	RSMAS	EILAT 2	MLC 2008
No pre-processing	90.7	70.1	80.1	64.0
Color correction	NA	NA	NA	63.8
CLAHS	92.9	85.8	87.4	69.3
Normalization	70.7	64.5	58.8	58.2
Color channel stretching	67.1	58.8	62.9	70.5
CLAHS + Color correction	NA	NA	NA	70.9
CLAHS + Normalization	91.0	85.2	87.3	68.1
CLAHS + Color channel stretching	91.4	82.4	81.7	72.7
CLAHS + Color correction + Color channel stretching	NA	NA	NA	73.2

). CLAHS, color correction and color channel stretching produced the best result for the MLC 2008 dataset.

Regarding the feature extraction options, (step 2), the combined feature vector of color histogram (hue and opponent angle), CLBP, Gabor filter response and GLCM worked most effectively on EILAT, RSMAS and MLC 2008 datasets (

Table 5. Overall accuracy (%) with different feature extraction methods as evaluated with the EILAT, RSMAS, EILAT 2 and MLC 2008 datasets. In this experiment, fixed configurations on the rest of the steps are as follows. Image enhancement: contrast limited adaptive histogram specification (CLAHS); kernel mapping: L1 normalization; dimension reduction: principal component analysis (PCA); prior: Class frequency; classifier: k-nearest neighbor (KNN). Different combinations of completed local binary pattern (CLBP), grey level co-occurrence matrix (GLCM), Gabor filter response and color histogram (hue + opponent angle) are evaluated in this experiments.

	EILAT	RSMAS	EILAT 2	MLC 2008
CLBP	91.3	74.7	84.8	55.1
Gabor	85.7	61.2	65.3	39.4
GLCM	70.9	62.9	58.2	46.8
Color histogram (hue + opponent angle)	64.2	81.7	53.0	41.3
CLBP + Gabor	90.5	75.0	87.7	54.9
CLBP + GLCM	92.2	75.8	86.4	57.1
Gabor + GLCM	87.6	72.1	77.5	46.3
CLBP + GLCM + Gabor filter response	93.4	83.5	91.2	62.4
CLBP + GLCM + Gabor filter response + color histogram (hue + opponent angle)	94.7	89.6	87.3	65.7

). For the EILAT 2 dataset, the color features are not discriminative enough to aid the classification performance in the combined feature vector. In general, if color images are available, color histograms can be used together with texture descriptors. The combination of CLBP, Gabor filter response and GLCM (texture descriptors) works better than any single feature in all four datasets.

With respect to the available kernel mapping options (step 3), the best results are obtained by using both the chi-square and Hellinger kernels along with L1 normalization, as illustrated in

Table 6. Overall accuracy (%) with different normalization and kernel mapping methods as evaluated with the EILAT, RSMAS, EILAT 2 and MLC 2008 datasets. In this experiment, fixed configurations on the rest of the steps are as follows. Image enhancement: contrast limited adaptive histogram specification (CLAHS); feature extraction: completed local binary pattern (CLBP), GLCM, Gabor filter response, opponent angle, hue channel histogram; dimension reduction: principal component analysis (PCA); prior: class frequency; classifier: k-nearest neighbor (KNN).

	EILAT	RSMAS	EILAT 2	MLC 2008
Nothing	87.5	84.8	87.3	61.9
L1 normalization	91.9	87.6	87.4	64.0
Chi-square kernel mapping	85.7	87.6	88.1.	63.3
Hellinger kernel mapping	85.8	87.9	88.5	62.2
Kernel mapping (chi-square, Hellinger)	89.2	88.1	89.3	65.7
L1 normalization + kernel mapping (chi-square, Hellinger)	93.4	89.7	91.1	66.5

.

Finally, for the EILAT and RSMAS datasets the best classifier (step 4) and dimension reduction (step 6) options are the ones using the KNN classifier with dimension reduction by PCA and Fisher kernel mapping (

Table 7. Overall accuracy (%) with different dimension reduction and classification methods (support vector machine with radial basis kernel having one to all scheme (SVM), k-nearest neighbor (KNN), multi-layer back projection neural network (NN) or probability density weighted mean distance (PDWMD)) as evaluated with the EILAT, RSMAS, EILAT 2 and MLC 2008 datasets. In the table, principal component analysis (PCA), Fisher kernel (F), combination of principal component analysis and Fisher kernel (P+F) and no dimension reduction (ND) are applied. Moreover, SVM, KNN, NN and PDWMD are represented as S, K, N and P, respectively. In this experiment, fixed configurations on the rest of the steps are as follows. Image enhancement: contrast limited adaptive histogram specific (CLAHS); feature extraction: completed local binary pattern (CLBP), grey level co-occurrence matrix (GLCM), Gabor filter response, opponent angle and hue channel histogram; kernel mapping: L1 normalization; prior: class frequency.

	EILAT				RSMAS				EILAT 2				MLC 2008
	ND	PCA	F	P+F	ND	PCA	F	P+F	ND	PCA	F	P+F	ND	PCA	F	P+F
S	94.3	91.9	86.7	90.2	92.1	89.0	81.8	88.2	93.1	91,5	88.9	90.7	75.2	71.7	61.2	70.1
K	91.7	93.4	85.4	94.9	91.4	92.8	87.4	93.5	88.0	92.2	85.1	92.7	69.5	73.9	64.5	74.9
N	89.9	88.1	83.2	79.7	88.4	86.4	86.5	80.2	92.5	91.1	86.3	89.2	77.4	75.5	67.9	73.0
P	91.2	90.5	84.9	87.1	87.9	86.5	85.5	84.3	89.2	89.4	80.7	85.1	79.8	73.3	66.6	71.3

). For the EILAT 2 dataset, the SVM classifier without dimension reduction produces the highest overall accuracy. For MLC 2008, dataset, the maximum overall accuracy is achieved with probability density weighted mean distance classifier (PDWMD) and no dimension reduction.

Using the experimental results above (Tables 4, 5, 6 and 7), we select the configurations shown in

Table 8. Selected configuration for our method. Three different configurations are used for (1) EILAT and RSMAS datasets, (2) MLC 2008 dataset and (3) Columbia-Utrecht Reflectance and Texture (CURET), Kungliga Tekniska Högskolan (KTH), University of Illinois at Urbana-Champaign (UIUC) datasets. Contrast limited adaptive histogram specific (CLAHS), completed local binary pattern (CLBP), grey level co-occurrence matrix (GLCM), principal component analysis (PCA) and k-nearest neighbor (KNN) are used in the table as acronyms.

Steps	EILAT/RSMAS	MLC 2008	CURET/KTH/UIUC
Image enhancement	CLAHS	CLAHS, color correction, color channel stretching	CLAHS
Feature extraction	Opponent angle histogram, hue channel histogram, CLBP, GLCM, Gabor filter response	CLBP, GLCM, Gabor filter response, Opponent angle histogram, hue channel histogram	CLBP, GLCM, Gabor filter response
Kernel mapping	L1 normalization, chi-square kernel, Hellinger kernel (for CLBP, color histogram)	L1 normalization, chi-square kernel, Hellinger kernel (for CLBP, color histogram)	L1 normalization, chi-square kernel, Hellinger kernel (for CLBP, color histogram)
Dimension reduction Prior	PCA, Fisher kernel Class frequency	None Class frequency	PCA, Fisher kernel Class frequency
Classifier	KNN	Probability density weighted mean distance	KNN

for the respective datasets to be compared with other published methods. For texture datasets (Columbia-Utrecht Reflectance and Texture (CURET), Kungliga Tekniska Högskolan (KTH), University of Illinois at Urbana-Champaign (UIUC)), we use the same configuration as the EILAT dataset without color histograms as feature descriptors. The texture types, distribution and size of the grey texture datasets are similar to those of the EILAT dataset.

5.2. Evaluation of Proposed Classification Method

The proposed image classification method is evaluated using the MLC 2008 dataset by constructing a precision-recall curve and error matrix. Precision is defined as the fraction of retrieved instances that are relevant, while recall corresponds to the fraction of relevant instances that are retrieved. High recall indicates that an algorithm is capable of returning most of the relevant results. High precision indicates that an algorithm returns more relevant results than irrelevant. The precision-recall curves (Figure 4) and error matrix (

Table 9. The error matrix of our proposed method tested on the MLC 2008 dataset. The classes in both row and columns corresponds to A = Acropora, C = CCA, MA = Macroalgae, MO = Montipora, PA = Pavona, PP = Pocillopora, P = Porites, S = Sand and T = Turf. Within the main 9 × 9 cell portion of the table, the given number corresponds to the raw count of the number of validation image patches that fell into a particular target/output combination.

		Target Class
		A	C	MA	MO	PA	PP	P	S	T
Output Class	A	146	1	1	0	0	2	2	0	1
	C	9	2408	139	51	29	17	38	116	55
	MA	4	59	372	0	2	10	5	11	4
	MO	0	19	3	336	2	6	2	9	5
	PA	0	31	13	4	202	0	0	4	2
	PP	14	0	6	3	0	691	11	2	8
	P	9	32	7	5	3	11	727	11	40
	S	0	67	3	6	4	1	10	650	60
	T	6	18	16	11	6	6	24	45	817

) for our proposed method applied to the MLC 2008 dataset result in 85.3% overall accuracy and 75.3% average precision. The highest precision value is observed for Pocillopora, and the lowest for the macroalgae class, as shown in Figure 3.

Table 9 shows the error matrix of our method when applied on the MLC 2008 dataset. The highest number of examples in this dataset is from the crustose coralline algae (CCA) class, which also has frequent overlaps with other classes. The error matrix reflects this fact, as other classes are highly confused with CCA (as shown in the 2nd row of Table 9).

5.3. Comparison with Other Methods

Table 10. The error matrices of the tested other methods as applied to the MLC 2008 dataset. The classes in the rows and columns of each error matrix correspond to A = Acropora, C = CCA, MA = Macroalgae, MO = Montipora, PA = Pavona, PP = Pocillopora, P = Porites, S = Sand and T = Turf. Within the main 9 × 9 cell portion of each error matrix, the given number corresponds to the raw count of the number of validation image patches that fell into a particular target/output combination.

		Ground Truth											Ground truth
		A	C	MA	MO	PA	PP	P	S	T			A	C	MA	MO	PA	PP	P	S	T
Output Class	A	68	28	23	12	0	32	17	1	7	Output Class	A	112	17	12	1	0	21	12	0	13
	C	1	2226	94	30	24	30	35	84	111		C	3	2366	62	22	25	19	28	67	43
	MA	4	251	228	2	8	29	16	0	22		MA	2	166	318	2	3	28	5	3	33
	MO	0	100	7	169	0	21	21	14	84		MO	4	115	7	247	2	1	10	10	20
	PA	0	33	0	0	197	0	8	4	6		PA	0	38	8	3	177	0	7	6	9
	PP	9	69	36	17	0	554	51	0	8		PP	16	62	21	0	1	625	8	1	10
	P	0	112	28	12	0	60	531	29	67		P	9	100	8	8	2	4	629	8	71
	S	0	198	3	10	2	5	25	600	105		S	0	178	4	5	2	3	18	686	52
	T	2	156	32	34	0	8	38	94	628		T	2	94	12	9	1	4	53	46	771
Marcos											Stokes & Deane
Accuracy: 68.7%											Accuracy: 78.3%
		Ground Truth											Ground Truth
		A	C	MA	MO	PA	PP	P	S	T			A	C	MA	MO	PA	PP	P	S	T
Output Class	A	74	37	9	5	6	22	17	2	16	Output Class	A	148	24	9	2	2	27	16	0	8
	C	38	1888	106	55	31	94	84	178	161		C	6	2790	68	35	25	28	66	129	148
	MA	26	140	301	3	10	26	13	16	25		MA	2	194	421	15	10	21	10	1	28
	MO	4	81	6	175	13	22	43	35	37		MO	1	128	15	309	4	4	20	15	25
	PA	7	37	6	20	147	3	17	4	7		PA	1	50	4	1	216	9	21	2	7
	PP	102	136	34	14	10	363	29	13	43		PP	5	52	17	5	6	815	16	1	13
	P	13	119	22	35	27	27	457	41	98		P	8	150	17	13	10	20	740	31	69
	S	6	248	17	28	9	23	52	479	86		S	0	255	1	4	2	4	36	823	61
	T	10	185	28	21	7	38	87	92	524		T	3	320	25	13	1	24	69	61	724
Pizarro											Beijbom
Accuracy: 58.2%											Accuracy: 73.7%

shows error matrices for all four of the compared classification methods when tested on the MLC 2008 dataset. The proposed framework achieves the highest overall accuracy (OA) and highest average precision for five datasets among six (

Table 11. Overall accuracy (OA) (%) for each method/dataset. The highest overall accuracy obtained for each dataset is shown in bold.

	Marcos	Stokes & Deane	Pizarro	Beijbom	Caputo	Zhang	Our
EILAT	87.9	75.2	67.3	69.1	NA	NA	96.9
RSMAS	69.3	82.5	73.9	85.4	NA	NA	96.5
MLC 2008	68.7	78.3	58.2	73.7	NA	NA	85.5
CURET	20.8	49.7	38.1	86.5	98.6	98.5	99.2
KTH	25.5	88.9	48.3	36.3	95.8	96.7	98.3
UIUC	14.6	56.9	19.9	32.2	98.2	99.0	97.3

and

Table 12. Average precision (%) for each method/dataset. The highest average precision for each dataset is shown in bold. NA represents not applied.

	Marcos	Stokes & Deane	Pizarro	Beijbom	Caputo	Zhang	Our
EILAT	85.1	73.5	58.1	64.2	NA	NA	97.2
RSMAS	59.2	81.2	67.1	79.9	NA	NA	96.2
MLC 2008	49.5	61.3	46.4	64.9	NA	NA	74.8
CURET	14.7	30.1	29.1	81.5	98.2	98.1	98.4
KTH	23.5	84.8	39.7	34.6	95.7	96.4	97.7
UIUC	14.7	43.5	18.2	30.4	97.2	98.5	96.9

). For the EILAT dataset, all methods work reasonably well, though our method achieves the highest overall accuracy (96.9%). The second highest OA is achieved with the Marcos classification (87.9%). The results of the RSMAS dataset are similar to those of the EILAT dataset, except the second most accurate method is the Beijbom classification, with an overall accuracy of 85.4%. For the MLC 2008 dataset, our method results in 85.5% overall classification accuracy.

The proposed method achieves almost similar classification accuracy when tested against traditional and texture-only image datasets (Table 11). Most of the other methods, however, perform significantly worse against texture-only datasets. The failure of the methods by Marcos [9], Stokes & Deane [10], Pizarro [8] and Beijbom [4] on standard texture datasets indicates that these methods rely heavily on color information. Heavy reliance on color information may limit the robustness of classification algorithms, since color can be inconsistent or absent in underwater datasets.

The two state-of-the-art texture-classification algorithms (Caputo [19], Zhang [20]) perform well relative to the proposed method. For the UIUC dataset, our method attains 97.3% overall accuracy, but can not beat Zhang [20], which achieves 99% OA. The UIUC dataset has higher resolution than the other datasets; therefore, dense descriptors, such as our method, might be influenced by the background information resulting in inaccurate classification. Our method is mainly suitable for small patch sized image datasets, which is useful for classified mapping on mosaics.

For CURET and KTH datasets, which have a smaller patch size, our proposed method has higher overall accuracy than the texture-only method (Zhang [20]). We compare our proposed method with their method in the same experimental conditions and acquire 99.2% accuracy for the CURET dataset and 98.3% for KTH datasets, where the reported as state-of-the-art Caputo [19] obtains 98.6% accuracy for CURET dataset and Zhang [20] obtains 96.7% for the KTH dataset.

The classification methods by Marcos and Stokes & Deane are computationally efficient (Figure 5). Our method requires less time than the methods by Pizarro [4] and Beijbom [8]. For the method of Pizarro, the time for the generation of 500 visual word vocabulary is considered, which normally can be computed offline. If that time is taken out, this method can be considered as fast as our method. In this context, the method by Stokes and Deane can be considered very efficient. All these methods are implemented in Matlab 2010a and tested on Intel core i5-2430M CPU with 2.4 GHz speed and 6 GB RAM.

The performance consistency is tested as a function of the percentage of training samples used. As the percentage of the training samples for the MLC 2008 datasets changes from 10% to 60%, accuracy increases gradually in all the methods. This result reflects the fact that more training samples can produce stronger classifiers (Figure 6).

5.4. Classifying Image Mosaics

Applying the classification method to segment the full mosaic of Red Sea survey images takes approximately seven hours on an Intel core i5-2430M CPU with 2.4 GHz speed and 6 GB RAM. The image patches are of size 64 × 64 pixels with a sliding window of 16-pixels shift per iteration, result in 73,600 image patches to classify in the full mosaic.

The classification results of each method on the Red Sea mosaic image are shown in Figure 7. Misclassifications tend to be on the borders of objects (Figure 8). Morphological filtering alters some results by maintaining neighborhood consistency (Figure 8, right). Morphological filtering increases the classification accuracy from 82.8% to 83.7% (Figure 8). In Figure 9, the left is the input mosaic image and on the right is the segmented and color coded thematic map.

5.5. Recommendations for Classifying Future Datasets

The application of multiple classification methods to six image datasets results in the following recommendations concerning future image classifications.

• If the dataset contains low contrast or blurred images, CLAHS works very well as an image enhancement step. If color markers are available in raw images, color correction can be performed to enhance the color constancy.
• For texture description, the concatenation of GLCM, Gabor filter response and CLBP works consistently well. If images have good reliable color information, then opponent angle and hue color channel histograms can be added, with the texture descriptor assigning equal weights to both color and texture descriptors.
• In all the cases of image patch classification, sparsely populated bins within histograms possess higher distinctive information than the high frequency bins. This statement is based on the assumption that high frequency bins often represent the background of the object and contain less distinctive information. The chi-square and Hellinger kernels can be used to modify bin counts and boost the population of low frequent bins. L1 normalization of the feature vector is necessary in all cases before applying the classifier for training and testing.
• If the dataset is small (datasets with training samples less than 5,000 are considered as small ones), then PCA and Fisher kernel mapping works very effectively to reduce the feature dimension. However, for large (Datasets with training samples more than 12,000 are considered as large ones) datasets, almost all the features become useful, and the dimension reduction is much less effective. In large datasets, almost all the dimensions become discriminative, and only a few dimensions are reduced with PCA. Therefore, for large datasets, this dimension reduction step can be avoided.
• Class frequency works well as prior in all the cases.
• For smaller datasets, the KNN classifier has the best performance. However, as the datasets get larger, the effectiveness of this method reduces, owing to the higher storage requirements, lower efficiency in classification response and lower noise tolerance. Some recent works [38] address this problem of KNN. However, SVM (linear SVM with one against the rest scheme) and neural networks (multilayer perception with back projection configuration) can be appropriate classifiers for bigger datasets. For underwater images, the classification based on probability density weighted mean distance (PDWMD) from the tail of the distribution by Stokes and Deane [10] works efficiently, both in terms of time and efficiency.
• Morphological filtering can increase the accuracy of the classification results.
• For smaller datasets, the KNN classifier has the best performance. However, as the datasets get larger, the effectiveness of this method reduces because of high storage requirements, low efficiency in classification response and low noise tolerance. Some recent works [38] address this problem of KNN. However, SVM (linear SVM with one against the rest scheme) and neural networks (multilayer perception with back projection configuration) can be appropriate classifiers for bigger datasets. For underwater images, the classification based on probability density weighted mean distance (PDWMD) from the tail of the distribution [10] works efficiently both in terms of time and efficiency.
• In our method, all the features used are partially or completely scale and rotation invariant. These features are therefore able to mitigate the effects of limited scale variation of individual classes. For larger scale variation, it is important to have enough training examples of individual classes at different scales. In the future, multi-resolution mapping might be useful for benthic habitat covering large scale variations of individual classes.

6. Conclusion

Large datasets of benthic images, which have been enabled by digital acquisition and innovative platforms, such as remotely operated vehicles (ROVs) or autonomous underwater vehicles (AUVs), provide a new opportunity for remote sensing of coral reefs, as well as a challenge. The opportunity lies in the types of measurements that can be made from direct remote sensing of benthic organisms. The challenge lies in efficiently extracting biological or ecological information from the raw images. Some form of automated analysis will be required to make full use of this rich data source.

Our proposed method presents a novel image classification scheme for benthic coral reef images that achieved the highest overall classification accuracy of any of the tested methods and had moderate execution times. This paper used six standard datasets to compare the set of methods that are representative of the state-of-the-art in automated classification of seabed images. On state of the art Moorea labeled corals (MLC) dataset [4], our method achieved 85.5% overall accuracy, whereas all the other compared methods attained less than 80%, including the method by Beijbom [4].

The proposed method can be configured to the characteristics (e.g., size, number of classes and resolution of the samples, color information availability, class types and so forth) of individual datasets. We provided guidelines for choosing the appropriate configuration for future classification of reef images. The results suggest that using a selective combination of various preprocessing steps, feature extraction methods, kernel mappings, priori and classifiers for various datasets can give better results than using a single method for all datasets. The results can be extended over large continuous areas by using mosaics of underwater images. On the Red Sea mosaic image, our proposed method resulted in 83.7% overall accuracy, which is at least 8% higher than the other methods tested.

The experimental part of our work allowed us to identify classification problems that are specific to underwater images. On one hand, there are many classes in underwater imagery that have samples with very clear differences in shape, color, texture, size, rotation, illumination, view angle, camera distance and light conditions. On the other hand, there are overlapping classes that look almost the same from specific angles and distances. Finding optimal patch size and patch shift are still open questions. Moreover, additional challenges, such as motion blurring, color attenuation, refracted sunlight patterns, water temperature variation, sky color variation and scattering effects, have to be reduced to maintain the image quality and the reliable information content. These issues highlight areas where future research may continue to improve the accuracy and efficiency of automated classification methods.

Applying automated classification techniques to mosaic composites produces a rapid (in terms of expert annotation time) technique of characterizing reef communities that can be used to track changes over time. Quantifying benthic community composition over the scale of hundreds of square meters by automated analysis of underwater image mosaics is a novel capability in coral reef science and provides a new spatial scale from which reef dynamics can be observed and studied. In addition to the time saved by applying a single training dataset to a large-area reef mosaic, applying automated classification schemes to mosaics can also reduce computational time, since it bypasses the redundancy of classifying highly overlapping images. Furthermore, the techniques presented here are not uniquely limited to coral reef classification and may prove useful in other ecosystems.