Machine Learning Applications in Physical Oceanography

The ocean is a complex, dynamic system shaped by nonlinear interactions across space and time. Traditional methods like numerical simulations and Eulerian measurements often fall short in resolution, interpretability, or real-time applicability. This dissertation introduces a machine learning framework combining Gaussian Process Regression (GPR) and Vision Transformers (ViT) to enhance the analysis of both Lagrangian and remotely sensed ocean data. First, GPR is evaluated as a probabilistic tool for reconstructing velocity fields from sparse drifter data. Using two model-based datasets—a double-gyre simulation and a Navy Coastal Ocean Model convergence region—GPR yields accurate velocity fields with quantifiable uncertainty, particularly where sampling is dense or flow is structured. Next, ViT is applied to SAR imagery for sea surface classification. Trained on TenGeoP-SARwv and evaluated on AI4Arctic datasets, the ViT outperforms CNNs, especially in structure-based categories, and generalizes well across varying polarizations and resolutions. Attention maps offer interpretability by visualizing key decision regions. Finally, the framework integrates both GPR and ViT to detect and analyze Regions of Interest (ROIs) in real-world drifter datasets. Trained on synthetic dynamical regimes, the ViT model successfully identifies patterns like saddle points and spiral sinks in the SPLASH and LASER experiments. GPR then reconstructs velocity fields within these ROIs, revealing fine-scale dynamics and transitions. This dissertation contributes: (1) an interpretable ML pipeline for ocean velocity reconstruction, (2) a ViT-based system for dynamical regime identification, and (3) a combined ROI detection and interpretation framework that advances the application of ML in oceanography.