SEA-ViT: Forecasting Sea Surface Currents Using a Vision Transformer and GRU-Based Spatio-Temporal Covariance Model

Teerapong Panboonyuen

Abstract: Forecasting Sea surface currents is critical for maritime navigation, environmental monitoring, and climate analysis, especially in regions like the Gulf of Thailand and the Andaman Sea. In this work, we present SEA-ViT, a novel deep learning framework that combines the Vision Transformer (ViT) with bidirectional Gated Recurrent Units (GRUs) to model spatio-temporal covariance for accurate prediction of sea surface currents (U, V) using high-frequency radar (HF) data. The name SEA-ViT, short for Sea Surface Currents Forecasting using Vision Transformer, underscores the model's focus on ocean dynamics and its utilization of ViT to enhance predictive performance. SEA-ViT leverages over 30 years of historical data, incorporating ENSO indices (El Niño, La Niña, and neutral phases) to capture the intricate interplay between geographic coordinates and climatic variability. This approach significantly advances forecasting capabilities, aligning with the objectives of Thailand's Geo-Informatics and Space Technology Development Agency (GISTDA) to improve maritime region analytics. Our model outperforms state-of-the-art baselines, including CNNGRU and other leading time-series architectures. Code is available at https://github.com/kaopanboonyuen/gistda-ai-sea-surface-currents.