Creation and evaluation of deep learning models for the classification of coral reef structure images

Abstract

This thesis introduces a convolutional neural network (CNN) model to classify structured corals in a marine ecosystem monitoring system. The proposed model achieves continuous monitoring with a low computational cost and an absolute precision of 91.41%, surpassing the reference project of Gomez-Rios Gómez-Ríos et al., 2019a. It was found that using transfer learning with pre-trained neural networks and implementing l2 regularization improved the accuracy and avoided overfitting problems. However, limitations were faced due to the limited data availability in the StructureRSMAS database. For future development, it is planned to expand the database and explore new neural network architectures further to improve the generalizability and performance of the model. The proposed model offers an accurate and efficient classification of structured corals, which can be helpful to scientists, conservationists, and managers of marine ecosystems in their conservation efforts and informed decision-making