Abstract
INTRODUCTION: As a major global food crop, maize faces serious threats from pests that significantly impact crop yield and quality. Accurate and efficient pest detection is crucial for effective agricultural management. However, existing detection methods demonstrate inadequate performance when addressing challenges including diverse pest morphologies, inter-species similarities, and complex field environments. This study introduces AMS-YOLO, an enhanced detection model based on YOLOv8n, to address these critical challenges in maize pest identification. METHODS: To improve pest detection performance, we developed three synergistic modules specifically designed to address the identified challenges. First, the SMCA attention mechanism enhances target recognition within complex environmental settings. Second, an MSBlock multi-scale feature fusion module improves adaptability to pests across different growth stages. Third, an AMConv optimized downsampling strategy preserves subtle features necessary for distinguishing similar pest species. These architectural improvements were integrated into the YOLOv8n framework to create the AMS-YOLO model. RESULTS: Experimental evaluation on a dataset comprising 13 common maize pests covering comprehensive developmental stages demonstrates the effectiveness of AMS-YOLO. The model achieved 90.0% precision, 89.8% recall, 94.2% mAP50, and 73.7% mAP50:95, surpassing the original YOLOv8n by 3.1%, 3.7%, 3.2%, and 4.0%, respectively. Comprehensive comparative experiments showed superior performance over existing detection methods including SSD, RT-DETR, and various YOLO variants. Deployment tests on Jetson Nano revealed that the model size is only 5.3MB, representing a 15.9% reduction compared to the original YOLOv8n, with 19.6% fewer parameters and 16% reduced computational requirements while maintaining low resource utilization. DISCUSSION: The proposed AMS-YOLO model successfully addresses key challenges in maize pest detection through targeted architectural improvements. The lightweight design enables extended field deployment while maintaining high detection accuracy, making it highly suitable for resource-constrained agricultural environments. This advancement demonstrates significant potential for supporting more targeted pest management decisions, contributing to precision pesticide application and resource optimization in field conditions, thereby advancing intelligent and sustainable plant protection.