A lightweight non-intrusive framework for early-stage damage symptom detection of wood-boring pests in trees.

Early detection of wood-boring pest damage is crucial for preserving forest ecosystems and preventing economic losses in the timber industry. To address the challenges of extremely small early-stage damage symptoms, complex bark backgrounds, and deployment constraints on edge devices in forestry environments, this study proposes a light-weight non-intrusive detection framework, termed BH-YOLO, for early-stage damage symptom detection of wood-boring pests in trees. Here, "non-intrusive" refers to a fully contact-free and non-destructive detection paradigm that relies solely on RGB image acquisition and avoids invasive sensing methods such as drilling or sensor embedding, thereby balancing field diagnostic accuracy and edge-side deployment efficiency. Unlike conventional detectors that sacrifice accuracy for lightweight design, BH-YOLO achieves a superior accuracy-speed trade-off, making it particularly suitable for real-time deployment on resource-constrained edge devices. First, a Multi-Scale Ghost Convolution (MSGConv) module is integrated into the backbone to replace standard convolutions, leveraging parallel multi-scale kernel fusion to augment the model's perceptive capacity for irregular boreholes while effectively reducing topological complexity. Second, a Multi-Scale Attention Module with Channel Shuffle (MSAMCS) module is designed to supersede the conventional Bottleneck, significantly strengthening the extraction of core semantic features within complex textural environments. To further capture cryptic infestation symptoms, an Efficient Multi-scale Channel Attention (EMCA) module is embedded following the SPPF module, enhancing discrimination sensitivity toward minuscule targets through dynamic weight redistribution. Finally, Depthwise Separable Convolutions (DSConv) are employed to systematically replace the standard convolution layers and C2f modules in the neck network, achieving a drastic compression of parameter count and computational overhead. Experimental results demonstrate that BH-YOLO achieves an mAP@0.5 of 92.2% on the constructed real-world agricultural scene dataset, representing a 3.2% improvement over the baseline model, while the F1-score reaches 90.2%. In addition, the model's parameter count, floating-point operations, and model size are reduced by 66.7%, 64.6%, and 64.5%, respectively, while achieving an inference speed of 259.2 FPS for real-time mobile deployment-demonstrating its strong potential for edge-side deployment in real-world forestry monitoring scenarios. The proposed BH-YOLO model has been successfully deployed on Android mobile terminals, demonstrating robust performance across diverse illumination conditions and tree species, and specifically addresses the challenge of detecting extremely small, cryptic early-stage boreholes against complex bark textures, providing an efficient and automated solution for early monitoring and control of wood-boring pests while supporting forest ecosystem protection and reducing ecological and economic losses.