Boundary Feature-Enhanced and Sector-Constrained Deep Learning for Individual Tree Extraction From TLS Data

Individual tree extraction from terrestrial laser scanning data provides a foundation for forest investigation and monitoring. When current methods are applied to forests with complex spatial structure and a densely closed canopy, potential issues of under-segmentation or over-segmentation commonly exist. In this study, according to the characteristic of highly overlapping areas resulting from the competition among neighboring canopies, we propose a novel individual tree segmentation method based on boundary feature-enhanced deep learning with the constraint of sector-based region. First, we introduce a sector-based region delineation method to narrow the focus region when segmenting adjacent trees, as opposed to traditional cylindrical area construction. Then, in each sector, according to the symmetric and asymmetric competition theory, a series of boundary features are proposed to reflect the geometric characteristics of the overlapping area. Finally, we reformulate individual tree extraction as binary classification in each sector and propose a Boundary Feature-Enhanced Network (BFE-Net) for separation of central and neighboring trees. Verified on two datasets, consisting of 7 plots across temperate and boreal forests, the proposed method attains high single tree extraction accuracy, with mIoU and OA improved by more than 1.9% and 7.4% respectively, compared with state-of-the-art methods. Besides, our method shows a more stable performance when dealing with point density change, data occlusion, and canopy overlap.