Individual tree detection and segmentation from unmanned aerial vehicle-LiDAR data based on a trunk point distribution indicator

Abstract

The light detection and ranging (LiDAR) systems mounted on unmanned aerial vehicle (UAV) platforms can provide high-density point cloud data for accurate individual tree detection and segmentation, which is needed for precision forestry. Individual trees can be detected and segmented based on tree trunk detection. It is a challenging task in forests characterized by high understory vegetation and varying point densities of trunks caused by obstruction from the upper canopy. We propose an approach to detect tree trunks and segment individual trees from UAV-LiDAR data. First, a trunk point distribution indicator (TPDI) was used to detect potential tree trunk positions (PTPs). Then random sample consistency (RANSAC)-based 3D line fitting was applied to each PTP to differentiate tree trunks from understory vegetation. Finally, a trunk-based region-growing segmentation method was applied to segment individual trees, and the result was refined through analysis of crown shape and vertical profiles. The approach was tested at three study sites in Eucalyptus plantations, which were characterized by overlapping crowns and relatively high understory vegetation. F-scores ranging from 0.920 to 1.000 were derived in 12 plots, and the accuracies increased with the tree heights. The comparative shortest-path algorithm for tree trunk detection and segmentation was applied for comparison and derived much lower F-scores (0.526–0.867). The proposed approach was also evaluated by replacing TPDI with a similar indicator. The comparison result indicated that the proposed approach was especially advantageous in forests characterized by relatively low tree heights and high understory vegetation.