Answers are grounded in this paper and cite their sources.
While terrestrial laser scanning (TLS) and personal laser scanning (PLS) provide high-quality forest data, accurate tree species identification remains challenged by the trade-off between geometric detail and computational scalability. This study presents TCMFE-Net, an object-level deep learning framework featuring a general multi-feature fusion and extraction unit (GMFE-Unit), which combines XYZ coordinates with raw reflectance intensity. Evaluated on the TocknerPLS and STPCTLS datasets, comprising major coniferous and deciduous species (including spruce, pine, beech and oak), TCMFE-Net achieves overall accuracies of 94.5% and 94.9%, respectively, yielding competitive advantages over conventional machine learning and state-of-the-art transformer-based architectures. A systematic analysis of sampling strategies reveals that while farthest point sampling (FPS) maintains skeletal quality, random sampling (RS) provides a three times increase in inference speed. The effectiveness of global feature aggregation depends on structural preservation; while global descriptors improve performance under FPS, they may introduce noise under RS. However, the fusion of raw intensity features offsets structural information loss in RS, achieving a better trade-off between precision and efficiency. These results demonstrate the potential of TCMFE-Net for automated single-tree species classification in close-range forest surveys. The implementation is available at: https://github.com/LiTao-0797/TCMFE-Net.
No takes yet. Share an insight, caveat, or question.
Penggen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: