Flood Image Segmentation of UAV Aerial Images using Deep Learning

Abstract

An essential component of image analysis is flood segmentation, which makes it possible to identify flooded areas from aerial or satellite data. Unmanned aerial vehicles (UAVs) are acknowledged as useful instruments that offer extensive data for comprehending the extent of floods. The study uses the latest deep learning and image processing methods, concentrating on semantic image segmentation using the U-Net and DeepLabv3 architectures. A comparative analysis demonstrates the DeepLabv3 model's impressive performance of segmentation, with a precision of 0.7242, recall of 0.8163, IOU of 0.7414, dice loss of 0.1786, and overall accuracy of 0.9216. With competitive performance, the U-Net model achieves 0.8281 precision, 0.7937 recall, 0.5266 IOU, 0.1791 Dice loss, and 0.8974 accuracy. Performance of both models are well at segmentation flooding related feature, with U-Net exhibiting almost better precision.