Flood Detection Using Multispectral Images and SAR Data

Abstract

Remote sensing imagery analysis is a very crucial task in regard to climate or disaster monitoring. Satellite images can capture the ground surface conditions and give a huge amount of information in a single image. In recent days, with the availability of multi-temporal satellite data, monitoring of flood events have become pretty easy. It gives accurate and real time flood information. Flood is one of the most disastrous natural disasters in Assam, India. It is necessary to predict or monitor flood events to minimise the overall damage caused due to floods. There are many scientific approaches which have been made operational in flood monitoring related activities. However Deep Learning based approaches are not yet fully exploited so far to monitor and predict flood events. We propose flood detection in real-time with the help of multispectral images and SAR data using Deep Learning technique Convolutional Neural Network (CNN). The satellite images are from Sentinel-2 and the SAR data are from Sentinel-1. The CNN was trained with 100 images for 100 iterations. CNN has shown excellent performance in image-oriented tasks like classification, segmentation and feature extraction. Recently Deep learning techniques are used extensively on remote sensing data due to their high resolution and the former’s extensive computing capability. The study area comprises of 2 districts namely Barpeta and Kamrup of Assam, India. We have obtained an accuracy of 80% in detecting flood. Based on our result, deep learning may be vigorously explored in various other disaster detection or monitoring activities.