Abstract
Effective disaster response during floods requires quick identification of flooded areas. The typical data source for idenntification is aerial Remote Sensing (RS) imagery, often inexpensive RGB drone images.
Reliable flood water detection in RGB images is thus essential. However, detecting water in RGB flood imagery remains challenging, because water, mud, and soil appear similar. Moreover, the sheer volume of data makes manual analysis impractical.
To automate water detection, Computer Vision (CV) and Deep Learning (DL) techniques are employed, but state-of-the-art DL methods require large amounts of labeled data; a resource that is time-consuming to produce.
We introduce an improved human-in-the-loop strategy which creates labeled data consisting of pairs (RGB image, water mask) from aerial RGB and Near-Infrared (NIR) imagery of the 2021 Bad Neuenahr flood.
For our labelling strategy, we integrate the NIR data via a false-color representation. We then apply the Segment Anything Model 2 (SAM 2) on these false-color NIR representations.
Because flooded regions have complex shapes, the initial results require manual refinement. However, by leveraging sparse prompts to identify water, these adjustments are less time-consuming compared to traditional methods.
The final labeled RGB dataset serves to train DL models to detect flood water regions in RGB images without additional NIR information. As a result of the proposed labeling strategy, to foster further flood detection research, we provide NeuenahrFlood21, an RGB labeled dataset for the task of water segmentation in RGB images. NeuenahrFlood21 matches typical acquisition parameters during a river flooding event and adds to existing data resources with varied flood and vegetation patterns. We evaluate the benefit of the proposed labeling strategy by training state-of-the-art models on NeuenahrFlood21, confirming the enhanced automated flood detection capabilities.
