A deep learning approach for near-coastal sea surface temperature prediction

Accurate near-coastal sea surface temperature (SST) prediction remains challenging due to the limitations of numerical ocean models in resolving fine-scale coastal dynamics. This study proposes a novel deep learning framework specifically designed for station-level SST forecasting in nearshore regions. The framework employs a seasonal stratified sampling strategy to capture thermodynamic patterns across the annual cycle while preventing temporal distribution shift. Building upon the Segment Recurrent Neural Network (SegRNN) architecture, we identify a fundamental information compression bottleneck that causes forecast smoothing. To address this limitation, an Attention-Enhanced Parallel Multi-step Forecast (Attn-PMF) strategy is developed, enabling the model to directly retrieve high-variance features from historical sequences through global attention mechanisms. Validated using four years (2021–2024) of hourly observations from 31 coastal stations in the East China Sea, the proposed framework demonstrates superior performance compared to the operational FIO-COM numerical model, particularly for lead times beyond 48 hours. Results show that the Attn-PMF strategy effectively preserves high-frequency variability and mitigates forecast degradation, providing reliable predictions for coastal management and marine safety applications.