QUANTIFYING AND COMPARING UNCERTAINTIES WITH BAYESIAN METHODS FOR CLASSIFICATION OF SEAFLOOR IMAGES

Abstract

Our oceans are rapidly changing, and mapping the ocean floor is an important task for managing our resources. While vast amounts of seafloor (benthic) image data are available, annotating this data manually remains a significant challenge, even for expert oceanographers. The time and cost required to annotate millions of images are considerable, and additional complications arise from issues such as visibility, orientation, and image quality. To address these challenges and assist oceanographers in constructing more accurate ocean floor maps, Machine Learning (ML) techniques are employed to automate the classification of benthic images. This study utilizes the BenthicNet dataset, which comprises over 11 million seafloor images. Various ML techniques are applied, including point estimate methods for class memberships, Bayesian Neural Networks (BNNs), and Monte Carlo dropout inference. The primary focus of this thesis is the evaluation of model uncertainty in the classification of ocean data images using these techniques. Experiments were conducted to assess the reliability and robustness of the models.