Analyzing Drone Imagery of Flooded Regions with Deep Neural Networks

Abstract: Flooding is the world’s most prevalent natural disaster, causing a large amount of fatalities and severe economical consequences each year. In this thesis, drone imagery of flooded regions has been analyzed by deep neural networks in order to facilitate disaster prevention and response. The deep neural networks have been used to do image segmentation of buildings, roads and water. Two deep learning algorithms have been compared, the instance segmentation network Mask R-CNN and the semantic segmentation network Deeplabv3+, showing that Deeplabv3+ provides better segmentation masks for this type of imagery with a mIoU score close to 0.9 for buildings and water. Moreover, two post-processing methods have been implemented to investigate if they can improve the segmentation results. The implemented methods are morphological opening and closing operations as well as fully-connected conditional random fields. The experimental results show that these post-processing tools are able to slightly improve the results from the deep neural networks.