Sea surface temperature (SST) plays a pivotal role in air–sea interactions, with implications for climate, weather, and marine ecosystems, particularly in regions like the Caribbean Sea, where upwelling and dynamic oceanographic processes significantly influence biodiversity and fisheries. This study evaluates the performance of foundational models, Chronos and Lag-Llama, in forecasting SST using 22 years (2002–2023) of high-resolution satellite-derived and in situ data. The Chronos model, leveraging zero-shot learning and tokenization methods, consistently outperformed Lag-Llama across all forecast horizons, demonstrating lower errors and greater stability, especially in regions of moderate SST variability. The Chronos model’s ability to forecast extreme upwelling events is assessed, and a description of such events is presented for two regions in the southern Caribbean upwelling system. The Chronos forecast resembles SST variability in upwelling regions for forecast horizons of up to 7 days, providing reliable short-term predictions. Beyond this, the model exhibits increased bias and error, particularly in regions with strong SST gradients and high variability associated with coastal upwelling processes. The findings highlight the advantages of foundational models, including reduced computational demands and adaptability across diverse tasks, while also underscoring their limitations in regions with complex physical oceanographic phenomena. This study establishes a benchmark for SST forecasting using foundational models and emphasizes the need for hybrid approaches integrating physical principles to improve accuracy in dynamic and ecologically critical regions.
Read full abstract