Integrating AI-derived SAR Flood Extent Map Uncertainties in Remote Sensing Data Assimilation for Flood Forecasting

Numerical hydrodynamic models are widely used to simulate and forecast river water surface elevation (WSE) and flow velocity, over lead times ranging from hours to several days. Their predictive skill, however, is limited by multiple sources of uncertainty related to simplified governing equations, numerical solvers, forcing and boundary conditions, and model parameters, e.g. friction coefficients, obtained through calibration. These uncertainties propagate to model outputs and can significantly affect flood forecasts. Data Assimilation (DA) provides a robust framework to reduce such uncertainties by sequentially combining numerical model predictions with observations as they become available, while explicitly accounting for their respective error statistics. 

In this work, a joint state-parameter EnKF is implemented to reduce uncertainties in upstream time-varying inflow discharges and spatially distributed friction coefficients through the assimilation of in-situ WSE observations. The performance of the EnKF strongly depends on ensemble size and on the spatial and temporal density of the observing network. However, the limited availability and continued decline of in-situ river gauge stations, particularly in floodplains, motivate the integration of remote-sensing (RS) observations into the DA framework, and with that the uncertainties associated with the flood extent maps.

Recent advances in deep learning (DL) have significantly improved automatic SAR-based flood extent mapping. Nevertheless, most existing approaches provide deterministic flood extent maps without associated uncertainty estimates, which are essential for stochastic DA methods. To address this, we here rely on a unified DL framework, called  Density-Aware Conformal Flood Mapping (DACFM), that explicitly quantifies two complementary sources of uncertainty in SAR-derived flood maps: (i) DL model’s knowledge-related uncertainty, caused by finite training data or model misspecification, and (ii) SAR data-related uncertainty arising from image noise and flood/non-flood class ambiguity. DL model’s knowledge uncertainty is characterized using feature density analysis in the latent space of a density-aware neural network, while data-related uncertainty is quantified via softmax entropy. These uncertainty estimates are operationalized through conformal risk control at a user-defined risk level (α, δ), enabling the rejection of out-of-distribution samples and the generation of set-valued predictions for in-distribution inputs. Such a method of uncertainty estimation was evaluated across diverse real-world flooding contexts, including built-up areas, vegetated regions, and bare soil, demonstrating improved uncertainty quantification.

The proposed approach is demonstrated using a high-fidelity TELEMAC-2D hydrodynamic model of the Ohio River reach between the Cannelton and Newburgh locks and dams. RS-derived flood extent products from Sentinel-1 SAR are assimilated in the form of wet surface ratios (WSR) over selected floodplain subdomains, each accompanied by uncertainty estimates derived from the DL-based flood mapping framework. Flood reanalyses for the major flood events of February and April 2025 yield significant WSE error reduction. Independent flood extent maps derived from Sentinel-2, and Landsat-8 optical images were also used to validate the experiments.