Towards Accurate Flood Mapping in Arid Regions: Sentinel-1 SAR-based insights and explainable machine learning. 

Abstract

Floods are not only frequent but also one of the costliest natural disasters. The use of satellite remote sensing is a cost-effective and widely adopted method for near real-time flood monitoring. Optical satellite imagery excels at distinguishing water from other land cover types by leveraging the spectral behavior in visible and infrared ranges. However, a major limitation of optical sensors is their inability to penetrate through clouds. This results in images with missing information, impeding their use for flood monitoring. In the past decade, Sentinel-1 Synthetic Aperture Radar (SAR) imagery has emerged as a valuable tool in operational flood management, overcoming the challenges posed by optical sensors. SAR is an active imaging technique that provides cloud-free images day and night by utilizing specular reflection from smooth water surfaces. In SAR imagery, water appears dark due to its unique backscatter characteristics. While SAR amplitude has been widely used for flood detection and monitoring, it tends to overestimate flooded areas, especially in arid and semi-arid regions, because the radar backscatter over sand and open water surfaces is similar.  In our study, we explore the potential of Sentinel-1 amplitude and interferometric coherence in arid-flood mapping. We conduct multiple case studies and employ the random forest method to train, test, and validate our model predictions against flood masks derived from cloud-free optical imagery. We design several scenarios to investigate the contribution of different layers of information in improving flood mapping accuracy in arid regions along with feature importance analysis to understand the role of each feature to reduce model complexity. Our results demonstrate the effectiveness of fusing amplitude and coherence information in flood mapping,  as compared to coherence or amplitude alone. By utilizing the key features derived using permutation feature importance, flood mapping accuracy was significantly improved by approximately 50%, while also reducing response time, which is crucial for effective emergency management. The findings hold promise and emphasize the versatility of the proposed approach across different sensors and scenes. This offers significant potential for global flood mapping in arid regions, particularly in countries with limited resources. As future missions and advancements in SAR systems continue to evolve, the detection capabilities for floods will further improve, leading to enhanced flood management in arid areas.