Li Y., Matgen P., Chini M.
IEEE Transactions on Geoscience and Remote Sensing, vol. 64, art. no. 4202520, 2026
Accurate flood mapping from synthetic aperture radar (SAR) imagery is essential for disaster response and risk management. Although deep learning (DL) models have advanced SAR-based flood mapping, they typically lack mechanisms to quantify and communicate predictive uncertainty, limiting their reliability in high-stakes applications. This study proposes a unified framework that captures and conveys two key sources of uncertainty in flood mapping: uncertainty arising from the model's limited knowledge, and uncertainty caused by noise and ambiguity inherent in the data. The former is estimated via feature density in the latent space of a density-Aware neural network, while the latter is quantified using softmax entropy. These uncertainties are then communicated through conformal risk control under a user-defined risk level (α δ ): inputs associated with high model-knowledge uncertainty are flagged as out-of-distribution (OOD) and abstained from prediction, while data-related uncertainty guides the construction of set-valued predictions for in-distribution (ID) samples. Experiments on diverse real-world flood scenarios-including flooded built-up (FB) areas, flooded vegetation (FV), and flooded bare soil-demonstrate substantial gains in uncertainty quantification and predictive reliability. The proposed density-Aware neural network achieves an average area under the receiver operating characteristic (AUROC) of 0.921 in OOD detection, outperforming Bayesian neural networks (BNNs) (0.684) and deep ensembles (DEs) (0.603). OOD abstention improves predictive safety, reducing the average water false negative rate from 36.9% to 28.5% under a risk level of (α =0.05$ and δ =0.1$ ), while prediction sets reduce average miscoverage from 4.4% to 1.6% compared to standard singleton predictions. By explicitly quantifying and communicating the uncertainty in DL predictions, this work strengthens trustworthy, risk-informed decision support for SAR-based flood mapping.
