STVformer: A spatial-temporal-variable transformer with auxiliary knowledge for sea surface temperature prediction

Abstract

Sea surface temperature (SST) is a crucial indicator among the various factors influencing ocean dynamics. It significantly impacts weather patterns, ocean circulation, and marine biodiversity. SST variation is affected by multiple factors such as solar radiation and air-sea heat exchange, which contribute to the complexity of accurately predicting sea surface temperatures. The challenges of SST prediction tasks stem from the difficulty in modeling the coupling relationships between dynamic ocean variables and capturing long-term spatio-temporal dependencies. Existing data-driven methods for SST prediction overlook the physical relationships between ocean variables, and struggle to effectively capture long-term features. In this work, we propose a spatio-temporal-variable transformer model (STVformer) consisting of multi-variable feature representation module and spatio-temporal-variable saliency modeling module for SST prediction. STVformer first models the physical relationship among auxiliary variables including short-wave radiation (SWR), long-wave radiation (LWR), latent heat flux (LHF) and sensible heat flux (SHF) based on the heat budget equation. Then, it leverages the saliency self-attention mechanism and the spatio-temporal attention mechanism to effectively learn the spatio-temporal-variable correlations and long-term dependencies. Extensive experiments are carried out on two datasets to validate the effectiveness of STVformer. The experimental results demonstrate that STVformer surpasses existing methods in SST prediction.