Optimizing the accuracy of flood extent mapping using multitemporal stack of Sentinel-1 SAR data with machine learning approach for Wardha River, Chandrapur District (India)

Abstract

Floods are the most common, destructive and frequently occurring natural disasters on the earth in terms of economic damages and affected lives. A flood can be an inconvenience or a catastrophic event, resulting in long-term economic and environmental consequences. Flood extent mapping identifies and delineates the areas that are inundated. The study focuses on the flood event of Wardha river near Chandrapur on 12th August, 2022. In this study, we proposed an ensemble averaging model (EAM) for optimizing the accuracy of flood inundation mapping that discriminates flood waters from the non-flood waters using stack of multitemporal Sentinel-1 satellite imagery. Sentinel-1 uses C-band microwave signals to measure backscatter from the Earth's surface with its synthetic aperture radar (SAR) sensor that can penetrate clouds and collects data regardless of weather conditions. The results of the proposed model were compared with other machine learning models such as SVM, RF and MLC. The result analysis reveals that the overall accuracy, Kappa coefficient (KC) and area under curve (AUC) values for the proposed model (OA = 98%, KC = 0.97, AUC = 0.986 for training and OA = 97%, KC = 0.96, AUC = 0.957 for testing dataset) outperformed the other models. The result may help people and town planners in identifying safe and risky areas in the study area.