Weeds detection in saffron fields using an improved YOLOv۵ model

The excessive use of agricultural pesticides and inputs has caused severe environmental damages to agricultural ecosystems. By applying digital agriculture and variable rate application systems, different sections of a farm can be managed with varying levels of pesticides and inputs, which is beneficial both in terms of production costs and environmental issues. In this study, a weed and saffron plant detection model was designed and evaluated to develop a selective weed control system in saffron fields. The proposed weed detection model is based on the YOLOv۵ object detection model. Specifically, several CBS and C۳ modules in the YOLOv۵s model were replaced with Ghost Bottleneck and C۳Ghost modules, respectively. This was done to reduce the number of model parameters and make the network lighter, which increases the speed of image processing during model training and inference. Furthermore, to improve the detection accuracy of the proposed model, a coordinate attention (CoordAtt) layer was used. The results showed that the number of parameters in the proposed model was reduced by ۴۷% compared to the corresponding model in terms of network width and depth coefficients in YOLOv۵ versions. Meanwhile, among the six trained models, the modified Yolov۵s model demonstrated the best performance, achieving ۸۱% accuracy and ۶۷% recall. The detection accuracy of the proposed model was ۳.۹۳% higher than that of the best-performing YOLOv۵ algorithm. Due to the lightweight nature of the proposed algorithm, it can be used for real-time weed detection in agricultural fields to develop selective control systems.

Keywords:

Saffron , Weeds control , Object detection , Deep Learning , YOLOv۵