A comprehensive evaluation of classification algorithms for coral reef habitat mapping: challenges related to quantity,quality, and impurity of training samples

The preparation of control data is a primary concern in many supervised classification schemes. In coral reef mapping, this issue becomes more severe for three reasons: (1) control samples, located beneath the water, are quite difficult and costly to access; (2) because of the high spatial variability of coral reef habitats, it is very difficult to obtain high-quality samples; and (3) pure training samples are also hardly achievable. These issues, namely quantity, quality, and impurity challenges, are the main focus of this study. Three classification algorithms, including Maximum Likelihood Classifier (MLC), Artificial Neural Networks (ANNs), and Support Vector Machines (SVMs), are comprehensively evaluated, and their requirements for control data are determined. To accomplish this, rich field data, collected from diving off of Lizard Island in eastern Australia, and Landsat-8 images are used as the input data. With respect to accuracy, ANN is best, as it can deal with the complexity of coral reef environments; however, it requires a higher number of training samples (i.e. ANN cannot manage the quantity challenge). On the other hand, SVM shows the best resistance against the quantity and impurity challenges. Being aware of these points, a coral reef map is produced, for the first time, of the northern Persian Gulf, a coral habitat with very special environmental conditions. In this region, SVM achieved 68.42% overall accuracy, even though a very limited field work campaign was conducted to provide the control data.