A comparison of the deep learning methods for solving seafloor image classification task

Abstract

An accurate interpretation of seabed images is relevant for evaluating and monitoring ecosystem states, assessing environmental impact, mapping seabed sediment, tracking life forms, and carrying out many other tasks. The main scope of this study is to establish an automated, accurate and efficient classification system of seabed images using state-of-the-art techniques. A convolutional neural network and other deep learning algorithms have been used to solve the seafloor images classification problem. To test the proposed techniques, images of five benthic classes were used, and the classes include “Red algae”, “Sponge”, “Sand”, “Lithothamnium” and “Kelp”. The task has been solved with the overall accuracy of 92.78% using a dataset consisting of 18356 image regions and the 10-fold cross-validation to assess the performance. The advantages of the convolutional neural network are presented and compared with the results obtained using two other deep learning techniques following different dataset formation approaches. The comparison of the classification results has shown that deep learning methods are suitable for seabed images classification, where the best precision has been shown by the convolutional neural network.