A Hybrid Mamba–ConvLSTM Framework for Multi-Day Sea Surface Temperature Forecasting at 0.05^∘ Resolution

Accurate multi-day sea surface temperature (SST) prediction at sub-mesoscale resolution is challenging due to nonlinear ocean dynamics, heterogeneous multi-source observations, and error accumulation during autoregressive rollout. This paper proposes a hybrid Mamba–ConvLSTM framework that combines recurrent local spatiotemporal encoding with selective state-space long-range spatial modeling. The ConvLSTM branch captures local spatial patterns and short-range temporal dependencies through convolutional gating, while the Mamba branch captures long-range spatial dependencies across each frame through cross-direction window scanning and maintains temporal coherence via persistent hidden states across successive time steps. A physically informed preprocessing stage aligns 0.083${}^{\circ }$ reanalysis variables to the 0.05${}^{\circ }$ OSTIA target grid via a Grow-and-Cut strategy and extracts gradient-based advection and diffusion proxy features under boundary-aware finite differencing. During autoregressive rollout, auxiliary variables are held at their last observed values and physical proxies are recomputed from the predicted SST, following a clearly specified protocol. Experiments on a South China Sea benchmark compare the proposed model against nine baselines—including persistence, daily climatology, ConvLSTM, PredRNN, ConvGRU, TCTN, PANN, Swin-UNet, and ViT-ST—under an identical data-split, normalization, and rollout protocol. Evaluation with RMSE, MAE, SSIM, $R^2$, and anomaly correlation coefficient (ACC) shows that the proposed model achieves a 10-day average RMSE of 0.512 ${}^{\circ }$C, outperforming the strongest learning-based baseline ViT-ST by 5.0% and the persistence forecast by 21.0%. Ablation studies, sensitivity analyses, seasonal evaluation, and statistical significance testing verify the contribution of each component and the robustness of the results.