SAR and optical imagery for dynamic global surface water monitoring: addressing sensor-specific uncertainty for data fusion

Satellite-based surface water monitoring is essential for traking the spatiotemporal dynamics of global water bodies. However, most existing systems rely on a single mission or sensor modality, constraining both accuracy and temporal coverage. To overcome these limitations, we propose a multi-mission data fusion framework that integrates SAR Sentinel-1 and optical Sentinel-2 observations. Two U-Net convolutional neural networks were trained independently on the S1S2-Water dataset: one using Sentinel-1 sigma-nought backscatter (VV/VH) and the other using Sentinel-2 RGB and NIR bands, with terrain slope incorporated as ancillary input in both models. Predictive uncertainty is quantified via Monte Carlo dropout embedded within the networks, modeling pixel-wise predictions as Gaussian distributions. These probabilistic outputs are subsequently fused using a Bayesian framework and refined through sensor-specific exclusion masks. Evaluation across 16 geographically diverse test sites demonstrates that the fused probabilistic predictions achieve an overall IoU of 89%, highlighting the synergistic benefits of uncertainty-aware, multi-sensor integration. Furthermore, we show that model evaluation restricted to cloud-free optical imagery introduces substantial bias, limiting applicability for near-real-time monitoring. The proposed framework improves temporal availability, robustness, and reliability, advancing multi-satellite approaches for global surface water monitoring.