UAV Image Analysis of Flooded Area Using Convolutional Neural Networks

India has seen numerous flood events with severe infra structural damages and fatalities in recent years. UAV assisted technologies can contribute towards preparedness and response during these disasters. UAV images that capture a bird's eye view of the flooded area can be utilized for situation assessment and feedback. A major bottleneck identified here is the lack of a suitable data set. This work utilizes existing publicly available video data to create annotated data set of flooded areas in Kerala with 3 classes. This data set is then used to train YOLOv3 and YOLOv4 and the resulting models are analyzed. Within this framework we study the network behaviour by varying the loss function utilized and by feeding patches of images as input. It is seen that our method resulted in models that have high average precision values. This work provides a framework which can be utilized to generate focused data set to expand the number of classes involved and the situations analyzed.