TSS-Net: An efficient deep learning framework for individual tree segmentation in terrestrial laser scanning data

Accurate individual tree segmentation from terrestrial laser scanning (TLS) data is essential for forest inventory and ecosystem monitoring, yet remains challenging due to overlapping canopies, irregular tree structures, and varying point densities in complex forest scenes. This paper proposes TSS-Net (Tree Segmentation and Scoring Network), an end-to-end deep learning framework that integrates feature extraction, instance grouping, and quality assessment into a unified architecture for automated tree segmentation directly from raw point cloud data. Unlike previous methods requiring multi-stage training or extensive preprocessing, TSS-Net employs a sparse UNet backbone with dual prediction heads for semantic segmentation and offset prediction, combined with an offset-guided clustering algorithm that eliminates manual parameter tuning and separate ground filtering steps. The framework incorporates a quality-aware scoring mechanism that automatically filters low-confidence predictions, making it particularly suitable for operational forest inventory applications. Comprehensive validation on two diverse forest datasets—a plantation forest in Guangxi (1,362 trees, 4 species) and an urban landscape forest in Xiong’an (1,095 trees, 26 species)—demonstrates robust performance with an mIoU of 0.83 and F1-score of 0.8320, while achieving efficient training convergence in 3-5 hours and inference in approximately 5 seconds per scene. The proposed approach shows improvements over traditional methods and general-purpose deep learning techniques through forest-specific architectural adaptations, contributing to automated forest monitoring and tree-level parameter extraction.