Semantic segmentation and path planning for orchards based on UAV images

Abstract

A new semantic segmentation network, Space Atrous Depthwise Network (SADNet), was proposed to segment orchard unmanned aerial vehicle (UAV) images. In SADNet, a depthwise identity convolution block (DIC) was introduced to optimise the network feature extraction capabilities. An atrous spatial pyramid pooling module (ASPP) was introduced to fuse multi-scale semantic information. To improve the receptive field of the network, the concurrent spatial and channel squeeze and excitation module (scSE) were also introduced to focus the network on the pixels that need attention. The performance of SADNet on the orchard dataset was found to be better than that of DeepLabv3, PSPNet, and R2U-Net. The pixel accuracy and mean intersection over union reached 93.61 % and 88.28 %, respectively. MIoU of SADNet reached 86.2 % on the public dataset Pascal VOC 2012. Then, based on the results of orchard segmentation, we proposed a new method for constructing orchard raster maps. Based on this method, we planned the intra- and inter-row paths of the orchard. The planned paths have the smallest root mean square error of 3.19 m from the optimal path. The orchard segmentation and path planning method based on UAV images can plan the path of the whole orchard at the same time and can assist agricultural robots in pesticide spraying and fruit harvesting.