From Tree Segmentation to Stem Extraction: A Robust DBH Estimation Framework for Complex Forests using Handheld LiDAR and PointNet++

         Diameter at Breast Height (DBH) is a critical parameter for global carbon cycle modeling and forest biomass estimation. Conventionally, Individual Tree Segmentation (ITS) serves as the necessary prerequisite for DBH extraction. However, in high-density natural forests, this dependency becomes a bottleneck: severe canopy overlap and understory occlusion often cause traditional ITS algorithms to fail, severely limiting the accuracy of subsequent DBH retrieval.
          To address this challenge using handheld LiDAR, this study proposes an "Interference Rejection" strategy, shifting the focus from the challenging full-tree segmentation to targeted "stem semantic extraction." We argue that for DBH retrieval, separating the entire tree structure is unnecessary. Therefore, based on the PointNet++ framework, our method actively identifies and filters out non-essential interference (e.g., canopy foliage and shrub noise) from the LiDAR point clouds, isolating clean stem points directly.
             We validated this framework in the boreal forests of Mohe, China. Experimental results demonstrate the significant advantage of our approach. While recent mobile laser scanning studies typically report DBH estimation RMSEs ranging from 1.5 to 3.0 cm due to segmentation errors in complex environments, our "stem-focused" strategy achieved a superior RMSE of 1.26 cm. This workflow effectively bypasses the limitations of traditional segmentation, providing a highly automated and precise solution for forest inventory.