Real-time pest monitoring with RSCDet: Deploying a novel lightweight detection model on embedded systems

Timely and accurate pest monitoring is critical for protecting crop health and ensuring agricultural productivity. While deep learning models have shown great potential for pest detection, their deployment in real-world agricultural settings remains challenging due to factors such as variable lighting, cluttered backgrounds, and resource-constrained hardware platforms. To overcome this, we proposed RSCDet (RepGhostConv-Subpixel Fusion-Cascading Attention-Detection), a novel lightweight deep learning architecture designed for high performance and edge deployment. This study focuses on its system-level optimization, deployment, and field validation on embedded devices for in-situ applications. We developed a fully integrated pest monitoring platform that combines the RSCDet with NVIDIA Jetson TX2 NX hardware, TensorRT inference acceleration, robot operating system (ROS)-based asynchronous communication, and 4 G LTE connectivity. The system enables an end-to-end, edge-side pipeline covering image acquisition, real-time inference, and remote result visualization. Under resource-constrained conditions, RSCDet achieved over 30 frames per second (FPS), maintained detection accuracy above 90 %, and reduced mean absolute error by up to 56.6 % compared to the YOLO series. In real-world aphid monitoring tasks, the system demonstrated 84.5 % counting accuracy in high-density aphid infestations and operated with a low latency of only 0.25 s for remote result feedback. This study provides a validated design and deployment strategy for intelligent systems in challenging field environments. The platform offers a scalable and energy-efficient solution for real-time monitoring and early intervention, with potential deployment on ground and aerial autonomous control devices.