Real-time soybean pest detection system integrating UAV and Jetson based on improved YOLO.

Abstract

Insect pests significantly reduce soybean yield and quality. To overcome limitations of current detection models, such as high computational complexity, excessive parameters, and low accuracy, this paper proposes a lightweight, high-precision, and robust portable soybean pest recognition system. The raw close-range dataset is first preprocessed from an Unmanned Aerial Vehicle (UAV) perspective to enhance representativeness. The backbone network is optimized using three modules: Hierarchical Weight Decoupling (HWD) for lightweight subsampling, the C3 module with Kernel-2 decomposition and Multi-Channel Attention (C3K2-MCA) to enhance feature extraction, and Spatial Pyramid Pooling-Fast with Squeeze-and-Excitation Version 2 (SPPF-SEV2) to strengthen feature representation while reducing complexity. The Normalized Wasserstein Distance (NWD) loss function further improves target detection accuracy and robustness. The optimized model achieves a high precision rate of 96.2%. When deployed on NVIDIA Jetson devices, the UAV-Jetson-YOLO system enables rapid, cost-effective pest detection for smart agriculture.