Multiclass weed and crop detection using optimized YOLO models on edge devices
Arjun Upadhyay, G C Sunil, Samriddha Das, Joseph Mettler, Kirk Howatt, Xin Sun
Abstract
Real-time accurate weed identification remains a considerable challenge due to dynamic environmental conditions and the high visual similarities between weeds and crops in terms of shape, color, and texture. The objective of this study is to utilize deep learning (DL) models for developing optimized YOLO-based DL models for multiclass weed and crop detection, designed specifically for deploying on edge devices that possess low computational power. Specifically, eight different versions of the You Only Look Once (YOLO) object detectors were used for weed and crop species identification under various environmental conditions in real fields. The YOLO dataset consisting of 4828 annotated images, collected from multiple locations across the state of North Dakota was used for training and evaluation. The models were evaluated on an NVIDIA Jetson AGX Orin edge device for both accuracy and inference speed. Among the DL models employed, YOLO11n and YOLO11-edge-base outperformed others in terms of mean average precision (mAP) and inference speed. YOLO11n achieved an [email protected] value of 0.83 with an inference speed of 10 ms, while custom YOLO11-edge-base attained an [email protected] value of 0.819 with a reduced inference time of 6.7 ms on the edge device. The lightweight optimized YOLO models demonstrated a balance between accuracy and efficiency while achieving good performance in object detection and maintaining fast inference speeds. These findings indicate that the lightweight models developed in this research hold promise for real-time applications on edge computing devices with limited computational resources, such as smart spraying system and intelligent mechanical weeding platforms.