Lightweight 3D LiDAR Rubber Tree Detection with Knowledge Distillation and Efficient Attention

Abstract This study investigates a compact LiDAR-based detector for rubber-tree trunk detection using a controlled simulated plantation benchmark together with two real-world forest-domain evaluations. Starting from MobilePIXOR, the backbone is modified with LiDAR-Adapted Omni-Scale Blocks (LAOSB) to strengthen multi-scale BEV feature extraction while reducing model size and computational cost. The simulation data are used as a controlled within-domain benchmark rather than as a scene-disjoint test of real-world generalization. Across three random seeds, LAOSB improves AP50:95 from 0.713 ± 0.001 to 0.732 ± 0.010 while reducing the parameter count from 0.596M to 0.404M and the computational cost from 12.91 to 11.23 GFLOPs. Further structured depth reduction produces the S3 Hard student, which achieves 0.735 ± 0.003 AP50:95 with 0.246M parameters and 10.94 GFLOPs, while LCB-BA Combined KD attains 0.738 ± 0.005 with 0.248M parameters and 11.05 GFLOPs. Matched hard-supervision and distillation controls show that the contribution of knowledge distillation is small and configuration-dependent. On the 354-frame acquisition-disjoint DigiForests evaluation, S3 Hard reaches 0.2874 ± 0.0172 AP50 and 0.0853 ± 0.0077 AP50:95 over 3,707 stems within 12 m. On the auxiliary manually annotated 50-frame FinnWoodlands pilot, S3 Hard reaches 0.2090 ± 0.0758 AP50 and 0.0720 ± 0.0265 AP50:95 over 903 trunks within 12 m, while visibility-stratified analysis shows a pronounced degradation under partial and heavy occlusion. Controlled point-dropout evaluation further indicates substantial operating-point sensitivity as LiDAR sparsity increases. DigiForests is treated as the primary external quantitative benchmark, while FinnWoodlands provides supporting manually annotated box-level and occlusion evidence; neither external evaluation is interpreted as direct scene-disjoint validation on natural rubber plantations.