Mean Shift-Based Clustering in Airborne Laser Scanning Data for 3D Single-Tree Detection in Forested Areas

As a non-parametric iterative mode-seeking algorithm, mean shift can be applied in digital forest resource monitoring using airborne laser scanning to analyze the forest point cloud and detect single trees. In this paper, we propose a mean shift-based computational scheme to segment and cluster the 3D point cloud data for the aim of detection and delineation of individual trees. The scheme employs mean shift iterative procedures on the data sets to cluster points with similar vegetation modes together. A defined complex multimodal feature space, containing information about the discrete target points’ geometrical shape and energy intensity, and a two-staged sequential segmentation procedure are adopted in this approach. Denoising and filtering operations between the coarse and fine segmentation stages are performed to improve the overall detection performance by clearing unwanted features, eliminating tiny clusters and pruning canopy structures. The proposed approach is validated on 20 test plots of a dense and multi-layered forest located in South China. Experimental results reveal that it can work effectively and when compared to the conventional raster-based methods, its accuracies are relatively high: it can detect 84 percent of the trees (“recall”) and 91 percent of the identified trees are correct (“precision”).