Basin-wide flood depth and exposure mapping from SAR images and machine learning models

Abstract

Recent years recorded an increasing number of short duration - high-intensity rainfall events in the Indian subcontinent consequent with urban and riverine flash floods. Rapid assessments of flooded areas are key for effective mitigation strategies and disaster risk plans, as well as to prepare operative policies for future events. Herein, we present an integrated methodology for rapidly mapping the flood extent, and depths based on Synthetic Aperture Radar (SAR) images and a digital elevation model (DEM). Incessant rain during August 2019 brought heavy riverine flooding in southern India, killed at least 280 people, and displaced about one million inhabitants from low-lying areas. We used SAR images by Sentinel-1 before, and during the flooding, and the MERIT DEM which enabled us to map the flood extent and flood depth of the inundation zones. Because the coverage of Sentinel-1 scene was limited to the Kabini river section during the flood period, flood extent and depth maps for the adjacent basin was generated by mapping the susceptibility for flooding using the training set obtained from the flood time Sentinel-1 images, and a set of predictive variables derived from DEM using random forest model. Qualitative analysis and cross-comparison with a numerical flood model proved the proposed approach is highly reliable with an accuracy value of 90% and 86% respectively for training and validation data, thus allowing a precise, simple, and fast flood mapping. The methodology presented here could be applied to other flooded areas having incomplete inventory in the context of flood risk assessment.