A novel self-similarity cluster grouping approach for individual tree crown segmentation using multi-features from UAV-based LiDAR and multi-angle photogrammetry data

Abstract

Automatic collection of tree-level crown information is essential for sustainable forest management and fine carbon stock estimation. UAV-based light detection and ranging (LiDAR) and UAV-based multi-angle photogrammetry (UMP) data depict the 3D structure of forests at a fine-grained level by generating detailed point clouds, making them potential alternatives to labor-intensive forest inventories. However, the accuracy of the individual tree crown segmentation algorithms that have been developed is unstable in forest stands with high terrain undulation and high canopy density, mainly due to the various crown sizes and interlocking crowns resulting in varying degrees of over- or under-segmentation. Here, we propose self-similarity cluster grouping (SCG) algorithm for individual tree crown segmentation that integrates multivariable calculus of crown surfaces and spectral-texture-color spatial information of crown. Firstly, according to the property that DSM and its multi-order gradient information can characterize the crown surface variation and concavity-convexity features, first- and second-order edge detection operators were used to preliminarily determine the crown patch edges in order to reduce under-segmentation. Then, we developed a self-similarity weight function controlled by the spectral, texture and color spatial information of the tree crown patches to increase the similarity difference between adjacent crown patches of the same tree and those of neighboring trees, and designed the strategy for cluster grouping crown patches to complete individual tree crown segmentation. The performance of the proposed SCG algorithm was verified in Mytilaria, Red oatchestnu, Chinese fir and Eucalyptus plots in subtropical forests of China using LiDAR and UMP data. The overall accuracy of F-score (f) was above 0.85 for crown segmentation, and the rRMSE for crown width, crown area and crown circumference extractions reached 0.13, 0.22 and 0.14, respectively. On this basis, we evaluated the effect of spatial resolution of DSM on the segmentation accuracy of SCG algorithm, and found that the crown segmentation accuracy was proportional to the spatial resolution. Compared to the normalized cut algorithm, marker-controlled watershed algorithm and threshold-based cloud point segmentation algorithm, the SCG algorithm improved the overall accuracy f of individual tree crown segmentation by 0.06, 0.13 and 0.05 for LiDAR and 0.06, 0.21 and 0.10 for UMP, respectively. Furthermore, the effectiveness and generalizability of the SCG algorithm was verified in other Mytilaria, Red oatchestnut, Chinese fir and Eucalyptus plots in subtropical forests and Larch and Chinese pine plots in temperate forests using UMP data. The crown segmentation accuracy was better than 0.82, and the crown width extraction accuracy was up to 89 %. Overall, our proposed SCG algorithm reduces the over- and under-segmentation in complex forest structures and provides technical support for accurate crown information extraction at both plot and forest stand levels.