STCA-DeFormer: a spatio-temporal Transformer with seasonal decomposition and cross-modal attention for sea surface temperature prediction

Abstract Sea surface temperature (SST) is a key physical variable for characterizing marine environmental conditions. It holds critical scientific significance and application value in marine disaster early warning, global climate change modeling, and fishery resource management. Aiming at the challenges of complex spatiotemporal feature coupling and difficult capture of multiscale variation patterns in SST prediction, this paper proposes a predictive model called STCA-DeFormer (Spatio-Temporal Cross-Attention Decomposition Transformer). The model integrates Seasonal and Trend decomposition using Locally estimated scatterplot smoothing (STL) and a spatio-temporal Transformer with cross-modal cross-attention mechanisms. First, the model decouples the original time series into trend, seasonal, and residual components via STL decomposition. Then, it constructs an enhanced geographic-time encoding module. This module incorporates a 2D convolutional attention mechanism to capture spatial heterogeneity features. Meanwhile, it employs cross-modal cross-attention to model the physical field coupling between SST and auxiliary variables (e.g., sea level pressure, 2-meter air temperature). Finally, a multi-component parallel Transformer architecture is adopted to achieve long-range temporal dependency modeling. Experiments using satellite-derived data over a study area in the South China Sea (SCS) show that the proposed method outperforms the convolutional long short-term memory (ConvLSTM) baseline model and other comparative models. The study demonstrates that the spatiotemporal Transformer architecture effectively improves SST prediction accuracy in complex environments. It provides a new technical pathway for marine disaster early warning and climate modeling.