SAR-FloodNet: A Patch-based Convolutional Neural Network for Flood Detection on SAR Images

Abstract

Flooding and other natural calamities may wreak havoc on people's lives and property. Following the incident, a detailed estimate of the affected area is necessary to undertake rescue operations by deploying an emergency response team. Obtaining an accurate estimate of a flooded region has traditionally required many human resources, and it is time-consuming. In this work, a patch-based Convolutional Neural Network (CNN) model is proposed for quickly detecting floods on remote sensing data. Since SAR - synthetic aperture radar images are acquired with active sensors in any weather condition, these images are considered for flood mapping during both day and night. The SEN 12-FLOOD dataset consisting of SAR images covering the flood events in Western Africa, Iran, and Australia is used to assess the model's performance. The SAR images are divided into small patches, and these patches are fed to the network for training and prediction. The maximum vote is considered for deciding whether the region covered by an image is flooded or not. The proposed SAR-FloodN et outperforms the existing pre-trained models in flood detection, giving an accuracy of 95%.