BF-CNet: A Bayesian-informed probabilistic fusion network to address cloud-induced data gaps in ocean remote sensing

• We introduce BF-CNet, a Bayesian fusion network integrating uncertainty features and spatiotemporal inference to address cloud-induced gaps. • Using the SSIM-optimal quantile, BF-CNet delivers more accurate, structurally faithful pixel reconstructions, surpassing existing methods. • BF-CNet produces 99-quantile predictive distributions, with PICP and ECE confirming reliable, well-calibrated uncertainty for ocean monitoring. Ocean remote sensing is essential for a wide range of marine applications, yet persistent cloud cover often introduces large spatiotemporal gaps that undermine data continuity and reliability. To address this challenge, we propose the Bayesian Field Convolutional Neural Network (BF-CNet), a probabilistic fusion network designed for cloud-affected reconstruction and forecasting. BF-CNet consists of two cooperating modules: the Uncertainty Feature Extraction Module (UFEM), built upon Bayesian Neural Fields, which performs pixel-wise posterior inference from cloud-free observations to reconstruct missing regions with quantified uncertainty; and the Spatiotemporal Inference Module (SIM), a ConvLSTM-based network that propagates these probabilistic representations through time for short-term forecasting. The two modules are trained sequentially within a cascaded framework. A hybrid loss function that combines per-quantile mean absolute error (MAE) and the continuous ranked probability score (CRPS) is adopted to balance deterministic accuracy and probabilistic calibration. Furthermore, a comprehensive evaluation framework integrating deterministic (MAE, RMSE, PSNR, SSIM) and probabilistic (PICP, ECE) metrics is employed to jointly assess reconstruction fidelity and uncertainty reliability. Trained on datasets with physically realistic cloud masks, BF-CNet achieves substantial improvements over advanced baselines (e.g., cloud-region reconstruction MAE = 0.367, RMSE = 0.643, SSIM = 0.943, PSNR = 34.32 dB; ECE = 0.011), while maintaining computational efficiency and robust generalization across different oceanic variables. These results demonstrate the effectiveness of probabilistic deep learning models in addressing missing-data challenges in ocean remote sensing, providing a reliable tool for spatiotemporal ocean monitoring and operational decision-making under cloudy conditions.