Deep learning methods for sea surface height forecastingRus, Marko (Avtor) Kristan, Matej (Mentor) Ličer, Matjaž (Komentor) sea-level forecastingstorm surgedeep learningspatio-temporal modelingoperational oceanographyGlobal mean sea-level rise driven by anthropogenic climate change has critically increased the frequency and severity of coastal flooding, particularly in semi-enclosed regional basins like the Adriatic Sea, where complex dynamics involving storm surges, astronomical tides, and basin seiches threaten densely populated communities. Traditional numerical ocean models, such as NEMO and SCHISM, have historically been the standard for operational forecasting. However, their predictive accuracy is inherently limited by the constraints of explicit physical modeling, such as inevitable process simplifications and uncertainties in the initial ocean state. Furthermore, these models are computationally expensive, struggling to efficiently provide the probabilistic ensemble forecasts necessary for estimating uncertainty. By learning these complex, non-linear sea-level dynamics directly from observational data, deep learning offers a powerful way to bypass both physical and computational bottlenecks. To address these limitations, this dissertation explores and validates the hypothesis that deep-learning architectures can provide a fast and highly accurate alternative for operational sea-level forecasting by systematically evolving the HIDRA (HIgh-performance Deep tidal Residual estimation method using Atmospheric data) model family. The deep convolutional neural network HIDRA2 directly predicts full single-point sea-surface height (SSH) over a 72-hour horizon, establishing itself as the first data-driven model to significantly outperform the operational numerical ocean model (NEMO) in the northern Adriatic during extreme storm surges. To better capture long-range spatio-temporal atmospheric dependencies, this architecture is subsequently enhanced in HIDRA-T, which integrates transformer-based encoders to achieve superior precision and recall in storm surge forecasting. HIDRA3 utilizes a joint spatial latent state to process data from multiple stations simultaneously, allowing it not only to implicitly reconstruct missing observations and maintain reliability during multi-sensor blackouts, but also to improve overall forecasting accuracy by incorporating a broader spatial context. HIDRA-D, by combining tide gauge records with sparse, unevenly distributed satellite absolute dynamic topography (ADT) and learning low-frequency spatial components in the Fourier domain, successfully generates dense 2D basin-scale forecasts. Because tide gauges and satellite altimetry utilize unaligned vertical datums, HIDRA-D inherently learns to automatically align these sources, effectively aligning local coastal stations against the global geoid, a highly valuable capability for broader oceanographic applications. Models of the HIDRA family have transcended academic validation to achieve operational success. The architectures have been successfully verified not only in the Adriatic but also in the Danish and Estonian waters, and were successfully tested at several tide gauges by the Spanish public port authority agency Puertos del Estado, achieving better accuracy than relevant operational numerical ocean models. Furthermore, the models are actively used in daily operational forecasting by the Slovenian Environment Agency (ARSO), and HIDRA2 is implemented as part of the European Commission's Destination Earth On Demand Extremes digital twin as an early detection system, triggering computationally expensive numerical models only when extreme hazards are detected.20262026-09-28 17:25:02Doktorsko delo/naloga188810VisID: 37215sl