Evidential Deep Learning for Uncertainty-Aware Global Flood Extent Segmentation

Flood extent mapping from satellite imagery plays a critical role in disaster response and flood risk management, particularly as flood events become more frequent and severe under a changing climate. At its core, the task involves classifying each pixel in an optical satellite image as flooded or non-flooded. Recent deep learning-based segmentation models have demonstrated strong performance at the global scale. However, despite their accuracy, most existing approaches provide deterministic predictions and offer limited information on the reliability of individual pixel-level outputs. This lack of uncertainty information constrains their operational applicability, especially in high-risk scenarios where models may exhibit overconfident but incorrect predictions.

To address this limitation, we extend a global flood extent segmentation framework by explicitly incorporating uncertainty quantification. Specifically, an Evidential Deep Learning (EDL) approach is integrated into a UNet++ architecture within the ml4floods framework, enabling simultaneous prediction of flood extent and associated pixel-wise uncertainty. Within the EDL formulation, network outputs are interpreted as evidence and parameterised using a Beta distribution, providing a principled estimate of predictive uncertainty. Furthermore, total uncertainty is decomposed into aleatoric and epistemic components, allowing clearer interpretation of whether uncertainty arises from data ambiguity or from limited model knowledge.

The proposed approach is evaluated using the extended WorldFloods global flood dataset. Preliminary results indicate that the EDL-enhanced model maintains promising segmentation performance while producing informative uncertainty maps. Elevated uncertainty is consistently observed in misclassified regions and along land-water boundaries, where optical signals are inherently ambiguous. These results demonstrate that uncertainty estimates offer valuable insight into model reliability and support operational decision-making by highlighting areas that require closer inspection. In practice, uncertainty-guided triage can help prioritise expert review and resource allocation, focusing attention on regions where decision risk is highest.