Satellite Image-Based Flooded Road Detection Using Semantic Segmentation

Abstract

During floods, submerged roads cut off rescue teams from affected people, yet most satellite flood products map general inundation and do not say which roads can still be used. This work presents a deep learning pipeline that identifies accessible routes during floods from satellite imagery. A U-Net with an ImageNet-pretrained ResNet34 encoder takes co-registered pre-event and post-event four-band tiles from the SpaceNet 8 Germany subset and predicts a road map and a flooded-road map directly. Training combines weighted binary cross-entropy with Dice loss to counter severe class imbalance, and tiles are split into training, validation and test sets by geographic block to avoid spatial leakage. The predicted road mask is skeletonised into a road graph, edges with submerged stretches are removed, and a shortest-path search returns the fastest remaining route between two locations, or reports that none exists. On the held-out test set the model reaches a flooded-road F1 score of 0.74 and a road F1 score of 0.79. In 1,000 sampled origin–destination queries, 93% of the returned routes avoid every truly flooded segment, a route is found for 95% of reachable destinations, and the mean detour is 9% longer than the ideal flood-aware route. Using pre-event imagery and a Dice term raises the flooded-road F1 score from 0.42 to 0.74 over a post-event, cross-entropy baseline. The results show that flood-aware routing from satellite imagery alone is practical for emergency planning.