Individual nursery trees classification and segmentation using a point cloud-based neural network with dense connection pattern

Nurseries are used to cultivate a variety of tree species. Obtaining some specific information like the tree species, positions of crowns and trunks can enhance the efficacy of nursery management. Due to the robustness to illumination, the point cloud-based neural network models have become extensively employed in segmenting and classifying individual trees from large-scale data. However, few studies have focused on further processing the point clouds of individual trees. Therefore, D-PointNet++ (Dense PointNet++) is proposed in this paper for classifying tree species and segmenting different parts of trees (crowns, trunks, pots and supporting poles). D-PointNet++ utilizes a dense connection pattern in the feature extraction module, inspired by the architecture of DenseNet. Additionally, the proposed model uses a gating system and concatenation as fusion operations to combine point cloud features with different dimensions to improve accuracy. The point cloud data of seven different types of garden trees in the nursery was collected using a laser sensor. The experimental results demonstrate that D-PointNet++ surpasses two representative baseline methods, PointNet and PointNet++, in terms of both classification and segmentation accuracy. For the self-made nursery dataset, the classification overall accuracy (OA) and class accuracy (mAcc) of D-PointNet++ can reach 92.65% and 92.54%; the average Intersection over Union (mIoU) and mAcc can reach 89.90% and 92.18%, respectively. The proposed D-PointNet++ can provide more accurate information on each tree and is beneficial to the management of the nursery.