Deep Learning Methods for Breaking Ocean Waves

Breaking ocean waves evolve through a complex and turbulent process which dissipates large amounts of energy in short periods of time. This thesis examines this process through the use of images and computer vision models. Training strategies for models of classification and segmentation tasks are examined and compared. Segmentation models of the breaking region are developed and a self-supervised pre-training task is assessed in improving the generalisation of models when training with limited labelled data. A stereo vision experiment is deployed in challenging conditions in a remote location on Inishmaan, one of the Aran Islands in Galway bay off the West coast of Ireland. The difficulties encountered are documented and recommendations for improvements on the stereo setup are presented. An alternate dataset of stereo images are reconstructed to a three dimensional representation of the sea surface and the directional spectra is estimated using a wavelet based method.