Accurate reconstruction of sea surface temperature (SST) from satellite observations remains challenging due to persistent data gaps caused by cloud cover and atmospheric contamination, especially in monsoon-influenced tropical basins such as the South China Sea (SCS). Here, we introduce the Peripheral Feature–Assisted Fourier-Transform Convolutional Bidirectional LSTM (PFTCBiLSTM), a deep learning framework that integrates Fourier-based convolutional LSTM units with a peripheral feature encoder–decoder incorporating shortwave radiation, sea surface wind speed, and spatiotemporal coordinates. This design enables the model to capture multi-scale SST variability while embedding relevant physical drivers, thereby supporting robust reconstructions under severe data degradation. Using daily Himawari-8 SST from 2015–2025, the PFTCBiLSTM consistently outperformed the advanced FTC-LSTM baseline, achieving R² = 0.98, SSIM = 0.98, MAE = 0.21 °C, and RMSE = 0.35 °C. A multi-stage training strategy, tailored to seasonal regimes and coastal complexities, reduced both bias and error variance across open-ocean and shallow-water environments. Independent validation with in-situ NOAA iQuam data confirmed strong agreement, particularly in cloud-prone and dynamically active regions. Robustness testing with adversarial perturbations demonstrated stability against realistic observational noise, while Shapley Additive Explanations (SHAP) indicated that SST at the prediction time step, seasonal timing, and recent SST history dominate the reconstruction process, with auxiliary variables gaining importance in coastal zones. By coupling physical interpretability with high statistical and structural fidelity, the PFTCBiLSTM advances operational SST monitoring and offers a transferable framework for reconstructing other cloud-affected satellite variables. Its application can enhance climate monitoring, ecosystem modeling, and marine hazard prediction in some of the world’s most observationally challenging ocean regions.