| dc.description.abstract | This thesis work explores multimodal learning techniques for habitat classification using remotely sensed and visual data. Autonomous Underwater Vehicles (AUVs) play a vital role in marine scientific surveys, providing efficient data collection and observations over marine ecosystems. Benthic habitat mapping, which involves classifying seabed sites into different habitat categories, is a key objective in marine ecology. AUVs capture visual imagery of the seabed, while multibeam sonars collect bathymetry data. By correlating visual imagery with features from the bathymetry data, reliable habitat classification models can be developed. This study investigates self-supervised learning approaches, particularly contrastive learning, to enable robust classification and image-content prediction. Results show that contrastive learning on bathymetry data achieves test accuracy rates of approximately 59% and 63% for patch sizes 16x16 and 32x32, respectively. In contrast, visual imagery achieves over 86% accuracy. Multimodal learning, combining visual images with bathymetry patches, yields accuracies of about 71% and 72% for different patch sizes. Separate networks with shared loss achieve accuracies of over 71%. This work demonstrates the feasibility and effectiveness of multimodal learning techniques in habitat classification, leveraging the strengths of both visual and bathymetry data. Future work involves exploring additional self-supervised multimodal learning approaches to improve underwater data analysis. | |
