LFT-YOLO: Lightweight Feature Transfer and Adaptive Upsampling for Small-Object Pine Wilt Disease Detection in UAV Imagery

Abstract

Monitoring forest diseases is essential for sustainable forestry management and ecological security. However, unmanned aerial vehicle (UAV)-based forest inspection remains challenging because diseased pine canopies are often small, densely distributed, partially occluded, and easily confused with complex background patterns caused by illumination variations, shadows, and surrounding healthy vegetation. To address these challenges, this paper proposes a lightweight detector, termed Lightweight Feature Transfer YOLO (LFT-YOLO), based on the YOLO26n architecture. The central design of LFT-YOLO is a lightweight small-target information preservation–reconstruction pathway. First, a Local Feature Injection (LFI) module is introduced to establish an independent shallow-feature transfer path, enabling high-resolution spatial and texture information to be directly aligned with and injected into the P3 feature space, thereby alleviating the loss of fine-grained information for small targets during progressive downsampling and multi-scale fusion. Second, conventional fixed upsampling in the neck is replaced with DySample-based adaptive upsampling to improve spatial reconstruction and cross-scale feature alignment. Experiments on the public Pests and Diseases Tree-Low-Level (PDT-LL) dataset show that LFT-YOLO achieves a Precision of 87.75%, a Recall of 87.02%, and an mAP@0.5 of 93.31%, corresponding to a 0.51-percentage-point improvement over the YOLO26n baseline. More importantly, the size-specific evaluation demonstrates that the performance gains are concentrated on small targets: compared with YOLO26n, small-object AP@0.5, AP@0.75, AP@0.5:0.95, Precision, and Recall improve by 2.15, 2.55, 2.15, 2.10, and 2.30 percentage points, respectively. These results indicate that the proposed architecture provides more pronounced benefits for the dominant small-object detection regime while largely preserving the performance of medium- and large-scale targets. LFT-YOLO contains only 2.43 M parameters and requires 5.80 GFLOPs. On an NVIDIA Jetson Orin Nano Super 8 GB platform using TensorRT FP16 under a fixed 15 W power mode, the model achieves 90.39 ± 0.40 FPS, with a P95 inference latency of 11.37 ms and an average power consumption of approximately 8.53 W. These results demonstrate that LFT-YOLO provides a favorable accuracy–efficiency trade-off and is suitable for real-time UAV-based detection of small and densely distributed pine wilt disease targets under resource-constrained edge-computing conditions.