CSTFnet: Cross-Scale Spatiotemporal Predictive Framework for Seasonal Prediction of Sea Surface Temperature

Sea surface temperature (SST) is critical for climate prediction, environmental monitoring, disaster warning, and various other oceanographic applications. SST prediction traditionally relies on numerical simulations, which effectively consider dynamical evolutions with explicit physical interpretations, but suffer from numerous challenges, such as limited timeliness and intensive computational demands. Recent advancements in deep learning have demonstrated its capability to capture spatiotemporal information through automatic feature representation learning, however the high-dimensional spatiotemporal correlation in ocean data, particularly the dependencies across different scales, remains inadequately explored. To fill this gap, this study proposes a cross-scale spatiotemporal fusion prediction network for seasonal prediction. Specifically, we explore the latent relationships within ocean spatiotemporal processes by tensor decomposition to derive mapping results across different scales. The fusion learning of the mapping results is achieved through dilated convolutional long short-term memory networks to achieve seasonal prediction of spatiotemporal process information. Extensive experiments demonstrate the effectiveness and superiority of the proposed model over existing state-of-the-art methods in medium- to long-term predictions.