A novel algorithm of individual tree crowns segmentation considering three-dimensional canopy attributes using UAV oblique photos
Lingting Lei, Tian Yin, Guoqi Chai, Yingbo Li, Yueting Wang, Xiang Jia, Xiaoli Zhang
Abstract
Extracting the individual tree crowns (ITCs) information is significant to fine forest resource investigation and carbon storage estimation. UAV oblique photogrammetry (UOP) can obtain point cloud data with high density due to higher presence of overlap, which has potential for tree crown information extraction. Many ITCs segmentation methods have been proposed to extract individual tree information. However, accurate ITCs segmentation using UOP data remains a challenge due to the uncertainties in environments with high heterogeneity of forest canopy vertical structure. Here, we proposed a novel approach, adaptive-kernel bandwidth mean-shift algorithm (AMS) considering three-dimensional canopy attributes, to segment ITCs using UOP data in complex forest environment. First, we developed a kernel bandwidth model with automatic adaptive parameter assignment using tree height derived from UOP data and applied it to the mean-shift algorithm. We demonstrated the generality of our algorithm in different tree species plots of a subtropical forest in China with overall precision f ≥ 0.72 and crown width rRMSE ≤ 0.13. Compared with the fixed-kernel bandwidth mean-shift algorithm (FMS) and seed region growth algorithm (SRG), the average f and rRMSE of the AMS algorithm were improved by 0.04 and 0.12, 0.16 and 0.11 respectively. Then, we evaluated the segmentation effect of AMS algorithm with point cloud densities of 25%, 50%, 75% and 100% respectively. We found that the segmentation accuracy decreases with decreasing point cloud density, but 75% of the point cloud density can satisfy most ITCs segmentation needs. In addition, we used LiDAR data (∼45 pts/m2) obtained by UAV to validate generalization ability of our approach, and the average r, p and f reached 0.97, 0.73 and 0.83. These results showed that the AMS algorithm can solve the ITCs segmentation problem in forests with complex structures using UOP and LiDAR point cloud data, which can support the accurate survey and scientific management of forest resources and provide basic data for the accurate estimation of forest carbon sinks.