Enabling scalable and energy-efficient weed detection using data-driven edge AI for precision agriculture

Abstract

Real-time weed detection is a key enabling technology for precision agriculture; however, deploying deep learning models on low-cost embedded platforms remains constrained by computational latency and energy consumption. In this study, we present a deployment-oriented evaluation of YOLO-based object detection models for weed detection under realistic edge-AI conditions. Multiple YOLO architectures (YOLOv8, YOLOv10, and YOLOv11) were trained on real agricultural field imagery and evaluated on a held-out unseen test set. Trained models were exported to an intermediate representation and compiled using the Hailo Dataflow Compiler with calibration-based quantization to generate accelerator-ready executable files, and deployed on a Raspberry Pi 5 integrated with a Hailo-8L inference accelerator. Results show that hardware-accelerated inference achieves substantial latency reductions (sub-5 ms per image under batch size 1) compared to CPU-based execution, while achieving F1-scores around 0.6 on the unseen test set. Although quantization introduces a moderate reduction in accuracy, the relative ordering of models remains consistent across deployment configurations. Energy efficiency analysis further demonstrates high throughput per watt and suitability for near-real-time processing. Overall, the results highlight the trade-offs between detection accuracy, inference latency, and energy efficiency, and demonstrate the feasibility of deploying YOLO-based weed detection models on low-cost edge platforms. Additional validation on continuous video streams and more diverse datasets is needed to confirm full real-world readiness.