Individual tree segmentation from UAS Lidar data based on hierarchical filtering and clustering

Accurate Individual Tree Segmentation (ITS) is fundamental to fine-scale forest structure and management studies. Light detection and ranging (Lidar) from Unoccupied Aerial Systems (UAS) has shown strengths in ITS and tree parameter estimation at stand and landscape scales. However, dense woodlands with tightly interspersed canopies and highly diverse tree species challenge the performance of ITS, and current research has not delved into the impact of mixed tree species and different leaf conditions on algorithm accuracy. Therefore, this study firstly evaluates the performance of open-source ITS methods, including both deep learning and non-deep learning algorithms, on data with mixed tree species and different leaf conditions, then proposes a hierarchical filtering and clustering (HFC) algorithm to mitigate the influence and improve the robustness. Hierarchical filtering consists of intensity filtering, Singular Value Decomposition (SVD) filtering, and Statistical Outlier Removal (SOR). Hierarchical clustering involves point-based clustering, cluster merging, and filtered point assignment. Through experiments on three distinct UAS Lidar datasets of forests with mixed tree species and different leaf conditions, HFC achieved the optimal segmentation results while maintaining high robustness. The variations of F1-score are 1-3 percentage points for mixed tree species and 1-2 percentage points for different leaf conditions across different datasets.