Deep Learning for Arbitrary-Shaped Water Pooling Region Detection on Aerial Images

Abstract

Recent rapid development in Unmanned Aerial Vehicles (UAVs) have extensively promoted several types of civilian tasks. In this paper, we propose and compare two different deep learning and convolutional neural network methods to detect and extract the region of water pooling areas, such as gutters, abandoned ponds, tires, and other water retaining areas on rooftops, using UAVs based aerial images. The performance comparison between the YOLOv4 algorithm and the Mask-RCNN algorithm was explored in the case study to identify the best deep learning method for detecting these uneven regions of water pooling. Experimental results show that the Mask-RCNN approach efficiently detects these uneven areas in an aerial image while simultaneously generating a high-quality segmentation mask for each instance. On the other hand, YOLOv4 detects the best bounding box for the area of interest. The mean average precision (mAP) scores for Mask-RCNN and YOLOv4 are 71.67% and 57.9% respectively. The Mask-RCNN system has shown promising results on test images and video clips. Such real-time detection systems would eventually help to identify mosquito breeding sites to assist the dengue eradication as well as to identify suitable water resources for daily uses, thereby facilitating a better community health system.