Search

Flooding is one of the most common natural disasters, causing emergencies and significant damage to infrastructure. Amid crises, entities such as the World Food Program employ remotely sensed images, usually via drones, to analyze the situation quickly and plan life-saving interventions. Suitable Computer Vision (CV) tools are necessary to assist experts of first responder teams with the analysis of the image data, increasing their productivity and enabling strategic resource allocation. Many state-of-the-art CV-tools use supervised Deep Learning (DL) techniques, for which labeled training data is needed. In this paper, we present the BlessemFlood21 dataset to support the development of DL-based CV-tools for the task of advanced flood analysis in non-coastal river flood scenarios. In particular, we address the problem of segmenting water from RGB images. The presented resulting dataset consists of high-definition, georeferenced RGB images labeled with precise water masks. The water masks are derived using a proposed semi-automatic human-in-the-loop strategy based on the use of additionally imaged NIR data together with a classical Random Forest approach. We assess the resulting dataset and provide a baseline for future advancements in flood mapping by training and evaluating three different state-of-the-art segmentation models (UNet++, DeepLabV3+, SegFormer-B5). Further, we showcase the potential use of the dataset with several experiments. In particular, we compare with the Floodnet dataset and, we employ the segmentation models trained on our dataset together with a Digital Elevation Model to showcase the potential to estimate flood water levels.