U-Net-based waterline segmentation for flood disaster response

Myzina A., Safin R., Tsoy T., Martinez-Garcia E., Magid E.

Floods are among the most destructive natural disasters, causing human casualties and severe economic damage. Effective monitoring of flooded areas requires automated systems capable of real-time perception and decision-making. This paper proposes a water surface segmentation model based on the U-Net architecture, trained entirely on synthetic data generated in the Gazebo simulator with variations in shoreline shape, water level, and surface color. The model demonstrated high segmentation accuracy on a test set (Dice coefficient = 0.9663, IoU = 0.9530) and maintained robustness across diverse scenarios, ranging from natural lakes to complex urban environments. When integrated into a UAV navigation algorithm, the system enabled real-time flight along detected flood boundaries. These results confirm feasibility of applying U-Net-based segmentation for UAV-assisted flood monitoring and search-and-rescue operations. Future work focuses on validating the approach with real-world data and adapting the network to resource-constrained onboard platforms.

Keywords: robotics, UAV, Semantic segmentation, U-Net, Gazebo, Search And Rescue, flood monitoring

In book

Proceedings Volume 14073, Fifth International Conference on Artificial Intelligence, Automation, and Algorithms (AI2A 2025)

[б.и.], 2026.