Classifier selection and cover area estimation of hard bottom in underwater belt transect thumbnails

Abstract

Classifier implementations of k-nearest neighbor (KNN), support vector machines (SVM), logistic regression (LR), and a feedforward neural network (NN) are assessed in terms of their performance in the classification between hard and soft bottom cover from thumbnails of underwater image mosaics. Input features were chosen from a set of 7 texture descriptors and appropriate parameters were tuned to maximize performance. The maximum accuracy values are 89% for both SVM and LR classifiers in Calatagan AM, and 88.4 ± 4.5% for NN in Calatagan PM. The performance of the KNN classifier is low for both datasets at 88.3 ± 2.5% and 86.3 ± 1.7% for Calatagan AM and PM respectively, and with low recall values. Errors across all classifiers are within range of each other. An application to the estimation of hard bottom cover and mapping is discussed.