Semantic segmentation of sparse 3D point cloud based on geometrical features for trellis-structured apple orchard

Abstract

Orchard operations such as mechanical pruning and spraying are heavily affected by tree architectures. Quantified inputs (e.g., cutting locations for mechanical pruning, and canopy distribution and density for variable-rate precision spraying) are necessary information for achieving precise control of these orchard operations. Even in planar orchard systems, trees grow differently. Therefore, it is essential to measure the canopy at the individual tree level. A three-dimensional (3D) light detection and ranging (LiDAR) sensor imaging system was developed to estimate the main canopy specifications. The LiDAR sensor was installed on a utility vehicle and driven alongside tree rows in an apple orchard. A total of 1,138 frames of point cloud data were acquired from 69 apple trees in a tall spindle architecture. An algorithm was developed in the MATLAB environment to segment trellis wires, support poles, and tree trunks in these point cloud images. The results indicated that the proposed algorithm achieved overall accuracy values of 88.6%, 82.1%, and 94.7%, respectively, in identifying the corresponding three objects. Furthermore, canopy density and depth maps were created with the distribution of points in the point cloud images. The outcomes from this study provide baseline information for precision orchard operations such as mechanical pruning and precision spraying.