NeuenahrFlood dataset and an improved human-in-the-loop strategy for efficient flood water segmentation

Effective disaster response during floods requires quickly identifying flooded areas using aerial RGB drone images. However, detecting floodwater is challenging because water, mud, and soil appear similar, and manual analysis is impractical with large datasets. We introduce an improved human-in-the-loop labeling strategy using the Segment Anything Model 2 (SAM 2) together with imagery from the 2021 Bad Neuenahr flood event. We incorporate Near-Infrared (NIR) data into a false-color representation, enhancing water visibility to generate labeled data (RGB image, water mask). While manual refinement is needed due to complex flood shapes, using sparse prompts reduces adjustments compared to traditional methods. Our labeled RGB dataset enables training Deep Learning models to detect floodwater in RGB images without NIR data. We present NeuenahrFlood21, an RGB labeled dataset for water segmentation during river floods, and training state-of-the-art models on it confirms improved automated flood detection capabilities.