Abstract:

Objective: This study addresses the challenges of detecting forest pests, collected by pest monitoring lamps, that are characterized by high diversity, complex sizes, and dense occlusions. We propose a multi-scale pest detection algorithm (MPD-YOLO) aimed at improving the accuracy of forest pest detection, thereby providing a new technological path for monitoring and protection of pests in forest areas. Method: First, a dataset comprising images of forest pests from regions such as Beijing, Hebei, and Shanxi was constructed. This dataset includes 18 categories of forest pests from various scenarios, providing a robust foundation for algorithm training and evaluation. Second, to improve the detection of small pests, MPD-YOLO employs 3D convolutions to extract deeper scale-sequence features from small targets, and introduces a scale sequence feature fusion module to enhance the model’s sensitivity to small objects. Finally, to address the challenges posed by dense occlusions and the coexistence of multi-scale pests, the MPD-YOLO method incorporates a triple feature encoding effective long-range aggregation network. This module fuses features across different scales, enhancing the interaction between multi-layer feature maps and improving the model’s detection performance. Result: Extensive experiments were conducted on the constructed dataset under identical test conditions, and the MPD-YOLO method was compared with mainstream object detection algorithms (YOLO series, Faster R-CNN, etc.). MPD-YOLO demonstrated superior pest recognition accuracy, with F1-score and mAP of 88.43% and 91.92%, respectively, outperforming the second-best method, YOLOv8x, by 1.45% and 1.22%. Additionally, compared to the original YOLOv7 model, MPD-YOLO achieved improvements of 0.72% in F1-score and 2.8% in mAP, confirming its advantages in handling complex object detection tasks. Ablation studies further analyzed the contributions of the scale sequence feature fusion and triple feature encoding modules, validating the effectiveness of each design component. Conclusion: The proposed MPD-YOLO algorithm significantly improves the detection performance of multi-scale pests in complex environments, demonstrating high robustness in challenging detection tasks. This algorithm provides a novel technical pathway for forest pest monitoring and protection.

Key words: forest pest control, forest ecosystem, small target detection, YOLOv7, feature fusion