Water Stream Segregation Using U-Net Architecture On Sentinel 2 A Dataset

Abstract

 Climate is a prolonged environmental style and its change is a natural process. As the climate is constantly changing, monitoring the surroundings globally is a need of the hour that requires attention. There’s is need to monitor places that have suffered due to diverse climate changes like heavy rainfall, drought etc. at the time of harsh climate condition. First objective of this study deals with observing areas remotely that have experienced floods or water retention and particularly demarcate the boundaries of water from remote Sentinel-2A images. This segregation of water is essential for handling diverse climate changes. Using satellite image for this task provide huge coverage and fast and quick upgradation. Deep learning algorithm, U-Net is used in study which is aiding in automating the surrounding monitoring process, benefiting from sentinel satellite images. Another objective of this study is to effectively observe and train the machine to anticipate the region bordered by water, which is enhancing deep learning-grounded methodological framework. In order to accomplish our objective a U-Net construction was implemented that allows learning of the key water classes label features from input Sentinel-2A. The result shows encouraging and promising accuracy value of 99.4 %. Performance measures used in our research are Mean intersection over union using Jaccard coefficient which results as 0.91 and total loss using focal loss and dice loss which results to 0.53