Deep learning for global sea surface temperature anomaly forecasting with balanced spatiotemporal features

With climate change intensifying, sea surface temperature anomalies (SSTAs) have become a key climate indicator. However, their nonlinear, non‑stationary, and periodic characteristics pose challenges to traditional forecasting methods. This paper proposes a hybrid forecasting framework that integrates three-level temporal decomposition with deep learning, extending single-point prediction to multi-region spatiotemporal forecasting. First, a hybrid SCV module integrating singular spectrum analysis (SSA), complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and variational mode decomposition (VMD) is established to disentangle the original signal into trend, periodic, and stochastic components, which are subsequently reconstructed via fuzzy entropy.These components are then fed into deep learning models (e.g., bidirectional gated recurrent unit (BiGRU)) for prediction. Second, we combine temporal features with spatial features to propose the SCV-spatiotemporal feature coupling block (STFCB) model. This model uses a feature pyramid network (FPN) and an intrinsic mode function (IMF)-guided convolutional block attention module (IMF-guided CBAM) to dynamically fuse spatial information. The results demonstrate that SCV-STFCB significantly outperforms the temporal-only SCV with feature attention (SCV-FA) egret swarm optimization algorithm (ESOA) model, with BiGRU achieving a mean square errora mean square error of 0.000073 and $R^2$ of 0.995903, indicating that the proposed spatiotemporal enhancement method improves the test-period predictive performance.