Improving flood detection in arid regions using Sentinel-1 interferometric coherence and machine learning

Kurzfassung

Floods are among the most devastating natural disasters, affecting about 1-in-4 people globally. The increasing frequency and intensity of extreme weather events have led to unprecedented flooding impacts, particularly in arid regions. The low soil permeability in arid regions means that short periods of heavy rain can cause rapid surface runoff, erosion, and infrastructure damage. Moreover, arid regions often lack the infrastructure and resources to cope with such disasters. Satellite-based remote sensing has become crucial for near real-time flood mapping and monitoring and rapid response and rescue operations. While optical satellites are limited by cloud cover, Synthetic Aperture Radar (SAR) satellites are increasingly utilized due to their relatively longer wavelengths which penetrate clouds, and their ability to collect information in different modes of polarization. However, current SAR-based flood detection methods struggle to differentiate between water and dry sandy surfaces, as both exhibit similar low-amplitude backscatter characteristics. This creates a critical gap in our ability to monitor and respond to floods in arid regions. We present a methodology that combines SAR amplitude and interferometric coherence data for flood detection in arid regions. We leverage ESA Copernicus Sentinel-1 data and employ a Random Forest classifier to integrate multiple SAR features, including temporal coherence and backscatter information. The predicted flood map is validated against reference flood maps derived using cloud-free Sentinel-2 optical imagery. The methodology is tested through three real-world flood events in Iran, Turkmenistan, and Pakistan. Our analysis reveals that combining coherence information with amplitude-based methods improves flood detection accuracy from 12% to 25% across the three test cases, with strong performance in areas where traditional methods typically fail. Using permutation feature importance analysis, we identified three key parameters: coherence and pre/post-flood amplitude changes all in vertically transmit and vertically receive. By focusing on these features, our model maintains the same accuracy while reducing processing time by 33%, making it more suitable for emergency response. The model also demonstrates robust performance across different geographical regions: successfully detecting floods in previously unseen locations without retraining the model. This geographical transferability of the model suggests the potential for a standardized flood detection system including arid regions. The increasing availability of open-access SAR data and advances in cloud computing have made handling and computing calculations of SAR data more feasible. With multiple space agencies launching new SAR missions, there are opportunities to test and adapt this methodology across different sensors and integrate it into operational flood mapping systems.