Flooded Road Detection using Deep Learning and Street Map Semantics for Humanitarian Aid Support

Flood disasters disrupt critical infrastructure such as roads, posing significant challenges for humanitarian aid organizations. In this work, we develop an AI method to detect flooded roads employing Deep Learning models trained for flood water detection in combination with semantic information of roads derived from street maps. One potential application is to inform disaster response teams about road accessibility using data collected from sources such as drones. We employ DeepLabV3+, UNet++, and SegFormer-b5 models previously trained on the BlessemFlood21 dataset to predict flooded areas. As street map input we employ road data extracted from OpenStreetMap and align it with the BlessemFlood21 image data. By combining the predicted flood areas with the street map information, we provide spatially resolved information on flooded road segments visualized in an orthomosaic highlighting the affected road parts. We assess the effectiveness of our methodology in identifying impacted road segments qualitatively and quantitatively. The proposed semi-automatic method for detecting flooded roads could serve as a foundation for assessment and route planning in the delivery of humanitarian aid.