Abstract: Objectives: The tree skeleton can reflect the branching structure of a tree, which is essential for analyzing its geometry and improving forest management. Unmanned Aerial Vehicle-Light Detection and Ranging (UAV-LiDAR) is an effective technique for extracting tree skeletons due to its flexibility and convenience in data acquisition. However, challenges such as incomplete branch point clouds and interference from leaf points in UAV-LiDAR data affect the accuracy of the results of skeleton extraction. Methods: An improved space colonization algorithm is proposed. The skeleton grows from a root point, determined by the lowest ten points in the UAV-LiDAR data. The growth influence space of skeleton points is defined using a dynamic radius adjustment mechanism based on search feedback. For the new skeleton points calculated from the influence space, a cylinder constraint is applied to limit their deletion range. As the skeleton points iteratively grow to branch tips, the growth is constrained by a layered concave hull algorithm. The complete skeleton points are then used to calculate branch levels and branching angles, which are compared with TLS data for accuracy assessment. Results: Three peach trees were processed, and the skeleton points were generally consistent with the original tree shapes. The main trunks and primary branches were correctly identified, with secondary branches extracted with an accuracy above 75% and an angular error of 5.125°. Compared with the original space colonization algorithm and other skeleton extraction methods (e.g., AdQSM, Laplacian-based), the proposed algorithm shows better overall and detailed performance with higher accuracy. Conclusions: By effectively addressing incomplete branch points and leaf point cloud interference in UAV-LiDAR data, the improved space colonization algorithm produces high-accuracy skeletons, offering robust data support for tree structure analysis and enabling applications in forest 3D modeling, resource assessment, and related fields.