Multiscale Spatio-Temporal Attention Network for Sea Surface Temperature Prediction

Abstract

Accurate prediction of sea surface temperature (SST), a crucial indicator of global climate and ecosystem changes, holds significant economic and social benefits. Deep learning has shown preliminary success in modeling the dynamic spatial-temporal dependencies within SST signals, yet it remains challenging to obtain precise SSTs due to the inherent variabilities across multiple temporal and spatial scales, driven by distinct physical processes. In this paper, we propose a novel multi-scale spatio-temporal attention network, named MUSTAN, tailored for the SST prediction problem. MUSTAN achieves multi-scale fusion through a progressive scale expansion paradigm, where sub-scale representations are iteratively merged with its counterpart scale units, enabling the propagation of fine-scale SST changes across broader scales. For each scale, MUSTAN introduces temporal attention to characterize dynamic SST patterns in different ocean regions, and spatial attention to capture intricate SST evolution interplay among these regions. Extensive experiments conducted on datasets from the Bohai Sea, Yellow Sea, and South China Sea consistently validate the effectiveness and superiority of our design, outperforming the state-of-the-art methods on SST prediction tasks.