Monte Carlo–Based Uncertainty Propagation for Probabilistic Water Masking from Satellite Remote Sensing Reflectance Product

Accurate and reliable delineation of surface water from optical satellite imagery is a pre-requisite for many hydrological applications. In inland and riverine environments, water masking is a major source of uncertainty due to optically complex waters, mixed land–water pixels and strong adjacency effects. Conventional water masks provide no explicit measure of classification confidence and do not account for uncertainty in pixel classification, particularly along riverbanks, under bridges, in building-shadows, and in highly dynamic systems, where small changes in reflectance or threshold parameters can lead to unstable water boundaries.

In this study, we present a generic Monte Carlo–based water detection framework that explicitly propagates Sentinel-2 ACOLITE remote sensing reflectance uncertainty through multiple spectral threshold-based water indices (NDWI, MNDWI, AWEI, and MBWI), resulting in per-pixel water occurrence probabilities. These indices are evaluated independently and combined using a deterministic voting-based fusion scheme. This decision logic is further constrained by physically motivated reflectance thresholds in the near-infrared and shortwave infrared bands, together with a low-signal filter, to suppress shadows and dark non-water surfaces that commonly generate false positives in index-based approaches.

The method is demonstrated as a proof of concept using a Sentinel-2 acquisition over the Seine River in Paris characterized by complex optical conditions. High-confidence water pixels dominate the main river channel, while intermediate probabilities are concentrated along riverbanks, bridges, and narrow tributaries. Within the final detected water mask, the mean water probability reaches 0.98, with more than 97% of water pixels classified with high confidence (P ≥ 0.9). Classification uncertainty is very low overall, indicating strong consistency across Monte Carlo realizations. Intermediate probabilities (0.3 < P < 0.7) represent less than 1% of detected water pixels and are spatially confined to water–land transition zones. Sensitivity experiments indicate that total water extent is weakly affected by increasing reflectance perturbation, whereas uncertainty increases systematically at water–land boundaries. By explicitly quantifying water-detection uncertainty, this Monte Carlo framework provides a statistically robust foundation for subsequent water-quality retrieval and uncertainty propagation.