Semantic Segmentation of Lotus Leaves in UAV Aerial Images via U-Net and DeepLab-based Networks

In the applications of remote sensing images, separating or segmenting the desired objects from the vast background is usually challenging. For example, calculating the number of single crops on agricultural land and the area of crops occupied per unit field, and quantizing the amount of new leaf buds in a forest area. Even though the targets of interest are visually recognizable, sometimes a large amount of them make an automatic segmentation tool preferable than the manual operation.In this work, the real-world problem we attempt to solve is to segment the lotus leaves from the background including water, field, and other plants. The segmentation result can be used to either monitor the growth rate of lotus or evaluate the severity of pests and diseases according to different color contrast. The look-down color image data were collected through a quadrotor UAV at different altitudes up to 20 meters. The image taken at the lowest altitude was chosen to delineate the groundtruth manually. We adopted two deep learning methods, U-net and DeepLab_v3+, to perform automatic leaves segmentation. The results show that both methods have similar performance but the latter one requires more computation. Then the conditional random fields (CRF) are applied as a post-processing to clean up the fragmented segmentation result.