An integrated edge AI prototype for smart agriculture: real-time pest detection, physical trapping, and multi-node...

Timely pest management is crucial for minimizing crop losses and reducing chemical pesticide reliance in sustainable rice agriculture; however, continuous large-scale field monitoring remains labor-intensive and difficult to achieve. To advance integrated pest management, this study presents a solar-powered edge AI prototype for pest detection and trapping, evaluated through laboratory closed-loop testing and simulated multi-node deployment analysis, supplemented by preliminary field observation under outdoor conditions. A field-collected dataset comprising 1,500 images of 12 distinct rice pests from actual paddies in Fujian, China, was constructed for model training and performance assessment. In the perception module, a Mixed Aggregation backbone and a Ghost Convolution pre-head design are adopted to reduce computational cost while preserving accuracy. TensorRT-accelerated inference enables efficient on-device execution, and ESP32-based peripherals support real-time actuation and long-range communication in a closed-loop laboratory testbed under low power consumption. A cost-coverage deployment framework with a coverage-balanced greedy strategy is introduced for simulated placement analysis in fragmented paddy fields under varying resource budgets. Experimental results show that the proposed configuration achieves 96.1% mAP@0.5 and 59.4 FPS, corresponding to +0.7% accuracy and +19.6% speed gains over the YOLOv13 baseline. Under synthetic fog, rain, and low-light degradations, the selected model achieves the highest mAP in 6 of 9 scenarios, indicating favorable robustness trends under controlled perturbations. The embedded pipeline sustains 44.5 FPS on Jetson Xavier NX with 14.8 ms closed-loop control latency and 36.8 ms end-to-end response latency, and simulated multi-node deployment analysis indicates an average 1.6% coverage gain under identical device budgets while preserving full network connectivity. Preliminary outdoor trials only indicate that the system's closed-loop workflow can operate under natural conditions and that trapping events were observed, serving as evidence of engineering feasibility; agronomic benefits still require further long-term quantitative evaluation.