Flood Surveillance Using Deep Learning

Abstract

Climate change is one of the biggest problems facing mankind. Increased flooding is one of the effects of climate change. Because floods are so severe, they can cause additional problems that can take only 24 hours to be seen in the affected areas. The paper deals with the extensive use of Deep Learning to identify flooded areas. Instead of using machine learning algorithms such as Decision Tree and Random Forest, a U-net architecture is used that will be able to locate and demarcate by doing classification on every pixel. The dataset consisted of VV and VH synthetic aperture radar (SAR) images which were converted to single RGB images. The dataset was augmented and an UNet model was created using the PyTorch library. The dataset was passed through six models which differed in number of epochs, learning rate and optimizer. Finally, the models were analyzed using cross entropy loss and MIOU.