Automatic flood detection by leveraging deep convolutional neural networks

Abstract

Early and precise flood detection from satellite imagery can help with rescue preparation and damage assessment for geophysical systems. Historically, hand-crafted methods have been used to automatically identify water in satellite pictures, but they frequently lack the precision and robustness required for accurate and early flood detection. This work dives into the 2018 floods in Kerala and some areas of Maharashtra using Synthetic Aperture Radar (SAR) images to provide a successful flood mapping framework using multi class U-net segmentation on RBG images produced after the preprocessing with the masks obtained using Random Forest Classification with an accuracy 97.98%.