TLNet: A deep learning framework for tree detection in forest point clouds using multi-layered forest structure

Individual-tree detection underpins forest inventory, growth monitoring, silvicultural planning, and disturbance assessment. Light Detection and Ranging (LiDAR) provides the 3D structure needed for this task; however, severe occlusion, irregular and range-dependent point density, and platform-dependent sampling geometry and visibility still hinder consistent performance across sites and sensors. We present Tree-Layer Network (TLNet), a deep learning framework that addresses these challenges by slicing forest point clouds into a stack of thin horizontal slabs that mirror the natural multi-layered organization of forest strata (understory, midstory, and overstory). Within each slab, 3D points are embedded via sparse-voxel convolutions and mapped onto a 2D grid of local features. To capture vertical context, TLNet fuses these grid feature maps in two directions: a bottom-up pass that propagates stable stem and lower-crown cues upward, and a top-down pass that injects overstory information downward. Learnable scalar gates adaptively re-weight each slab’s grid features to emphasize the most informative forest layers, and a lightweight decoder predicts a tree-location heatmap and coordinate offsets for precise localization. To mitigate limited real-world labels, we pre-train TLNet on fully labeled synthetic forests and fine-tune on datasets from three LiDAR platforms: terrestrial laser scanning, mobile laser scanning, and unmanned aerial vehicle-borne laser scanning. Across both coniferous- and deciduous-dominated sites, TLNet attains an average F1 score (harmonic mean of precision and recall) greater than 0.88, outperforming three point-based deep-learning backbones and three platform-specialized heuristic detectors. These results suggest improved cross-platform consistency and accurate stem localization for downstream inventory and silvicultural analyses.