<![CDATA[Although benthic habitats represent some of the largest, most diverse, and productive ecosystems on Earth with great environmental, and economical value, they are increasingly threatened and declining in many locations worldwide. Every year, numerous underwater images are collected for monitoring these habitats. Still, the manual labelling process remains tedious and time-consuming, creating a huge gap between data collection and extraction of meaningful information. In this study, an automated framework is proposed for single-label classification and semantic segmentation of benthic habitats using convolutional neural networks (CNNs). The framework integrates and evaluates various pre-trained CNNs, bagging of features (BOF), color spaces, and texture descriptors for benthic habitat classification. Furthermore, the classified images served as training and validation samples to assess the semantic segmentation performance of pre-trained CNNs with different architectures (e.g., ResNet-50, AlexNet, Xception, etc.). Both high- and low-quality underwater images of benthic habitats collected from six diverse study areas located off Australia and Japan were used to evaluate the proposed framework. The analysis revealed that the ResNet-50 FC1000 combined with BOF, color space, and texture attributes yielded the highest automatic classification accuracy. Moreover, the ResNet-50 network outperformed all the tested networks for automatic semantic segmentation of benthic habitats. Overall, the presented framework enhanced the automation of benthic habitat classification and semantic segmentation processes.]]>
