TSD-LitePT: Transition-Sensitive Detail Enhancement for UAV LiDAR Transmission-Corridor Segmentation

UAV-borne LiDAR segmentation of transmission corridors is hindered by the anisotropic geometry of elongated conductors, semantic ambiguity near component junctions, and the loss of sparse insulator evidence during hierarchical sampling. To address these challenges, we propose TSD-LitePT, a three-stage extension of LitePT. First, Learnable Axial Rotary Position Encoding learns independent scales for the x-, y-, and z-axis rotary phases, adapting positional sensitivity to corridor geometry. Second, Transition-Modulated Semantic Refinement derives a segmentation-driven point-wise response from decoded features and uses it to gate semantic residuals, refining ambiguous representations without boundary annotations. Third, Sparse-Target Detail Calibration uses decoded semantics, normalized coordinates, and transition context to calibrate shallow high-resolution features, then aggregates serialized-neighborhood evidence into an insulator prior. A bounded, confidence-aware residual injects this prior into the insulator logit while suppressing changes at confident non-target points. Across three runs on the original five-class corridor dataset, TSD-LitePT achieved an mIoU of 0.9200±0.0003 and an insulator IoU of 0.7236±0.0016 (mean ± standard deviation), compared with 0.8818±0.0061 and 0.5548±0.0435, respectively, for LitePT. After separate retraining on an independent transmission corridor, TSD-LitePT achieved an mIoU of 0.8909, compared with 0.8653 for LitePT. The largest improvement occurred for insulators, while power-line and ground performance remained close to the baseline. These results show that coordinating positional adaptation, semantic refinement, and controlled detail recovery improves minority-component segmentation without disturbing dominant classes.