Synthetic Aperture Radar (SAR) satellites produce vast archives of high-dimensional ocean imagery, capturing complex multi-scale surface patterns induced by sea-air interaction processes. To estimate geophysical parameters such as turbulence fluxes, scientists traditionally apply domain knowledge to manually select physically meaningful features, an implicit form of dimensionality reduction. This thesis explores whether Variational Autoencoders (VAEs) can automate this process of dimensionality reduction by learning compressed latent representations directly from raw SAR ocean imagery.
Using 220,000 Sentinel-1 Wave Mode images co-located with ERA5 reanalysis data, VAE architectures were trained across four latent dimensionalities (32, 64, 128 and 256). The multi-scale complexity of SAR scenes introduced a strong frequency bias: standard pixel-wise losses such as Mean Squared Error failed to capture fine-scale detail, and conventional metrics such as PSNR and SSIM proved insufficient to measure this. Frequency Focal Loss (FFL) was incorporated to address reconstruction in the spectral domain, alongside a dynamic weight matrix that refocuses the loss on difficult-to-learn features. Dynamically annealing loss term weights had a striking effect on reconstruction quality, and subsequent hyperparameter optimisation using Optuna further confirmed that loss function tuning dominates over architectural choices. This sensitivity to loss function design is a central finding of this work.
VAEs successfully reconstructed large- and intermediate-scale ocean patterns at latent dimensions of 128 and 256. For air-sea flux estimation, three configurations were compared: a direct CNN regressor, a frozen VAE encoder with regression head, and a jointly trained VAE. All three underperform the physics-informed approach of O'Driscoll et al.\cite{o2023obukhov}, suggesting unsupervised deep learning alone cannot extract flux-relevant information, though task objectives incorporate