A multi-scale small object detection framework for tomato leaf pest and disease detection in complex natural environments.

INTRODUCTION: This paper proposes an improved detection framework, PLD-YOLO, to address the challenges of small object scale, strong background interference, and loss of shallow feature information during feature extraction in pest and disease detection under complex natural environments for intelligent plant protection. METHODS: The proposed framework enhances small object representation by introducing a high-resolution feature branch. It further improves contextual modeling capability and suppresses complex background noise through multi-scale convolutions and an adaptive feature mechanism. In addition, a dynamic small object weighting strategy is introduced to improve small object learning under scale imbalance conditions. RESULTS: This study systematically evaluates the proposed method on the Tomato-Village dataset constructed under natural environments and compares it with representative two-stage object detection methods, YOLO series models, and transformer-based detection models. Ablation studies, cross-dataset validation, and robustness evaluations under different field conditions further demonstrate the effectiveness and robustness of the proposed approach. The experimental results show that PLD-YOLO achieves excellent performance in terms of precision, recall, mAP@50, and mAP@50-95 while maintaining high inference efficiency. DISCUSSION: The proposed method provides a potential technical solution for intelligent plant protection and contributes to the advancement of precision agriculture.