Unsupervised Semantic Segmenting TLS Data of Individual Tree Based on Smoothness Constraint Using Open-Source Datasets

Unsupervised segmentation of terrestrial laser scanning (TLS) data into wood and leaf is the key to studying forest carbon storage, photosynthesis, and canopy radiation. Further segmentation of wood data into the trunk and larger branch (TLB) and the remaining branch (RB) is of great significance and challenges for dust retention and soil heavy metal enrichment. We proposed an unsupervised, automatic semantic segmentation method based on TLS data of individual trees. The method first performs initial segmentation based on plane fitting residuals and neighborhood normal angle, which can extract smooth and connected regions in the point cloud. Then, the geometric features of segmented clusters are quantified to approximate RB or leaf features. Finally, the segmentation of TLB, RB, and leaf is realized by combining different clusters from bottom to top with geometric features and neighborhood relations. The segmentation performance of our method was evaluated with 104 tree samples from 23 tree species in two open-source datasets from Indonesia, Peru, and Guyana and from Canada and Finland. The microaverage precision of our method is 93.61%. The microaverage recalls of TLB, RB, and leaf are 97.08%, 86.44%, and 89.62%. Compared with the well-known method of separating wood and leaf, our method has 33.56% higher sensitivity, 1.82% higher specificity, 20.52% higher precision, and 0.217 higher F1-score. Besides, we estimated the surface area and volume of TLB, and the surface area and volume of RB based on the segmented data. The above parameters have good consistency compared to those calculated based on manually separated point clouds (Pearson correlation coefficient (PCC) of 0.55–0.93).