Weed detection in soybean fields using improved YOLOv7 and evaluating herbicide reduction efficacy

Abstract

With the increasing environmental awareness and the demand for sustainable agriculture, herbicide reduction has become an important goal. Accurate and efficient weed detection in soybean fields is the key to test the effectiveness of herbicide application, but current technologies and methods still have some problems in terms of accuracy and efficiency, such as relying on manual detection and poor adaptability to some complex environments. Therefore, in this study, weeding experiments in soybean fields with reduced herbicide application, including four levels, were carried out, and an unmanned aerial vehicle (UAV) was utilized to obtain field images. We proposed a weed detection model—YOLOv7-FWeed—based on improved YOLOv7, adopted F-ReLU as the activation function of the convolution module, and added the MaxPool multihead self-attention (M-MHSA) module to enhance the recognition accuracy of weeds. We continuously monitored changes in soybean leaf area and dry matter weight after herbicide reduction as a reflection of soybean growth at optimal herbicide application levels. The results showed that the herbicide application level of electrostatic spraying + 10% reduction could be used for weeding in soybean fields, and YOLOv7-FWeed was higher than YOLOv7 and YOLOv7-enhanced in all the evaluation indexes. The precision of the model was 0.9496, the recall was 0.9125, the F1 was 0.9307, and the mAP was 0.9662. The results of continuous monitoring of soybean leaf area and dry matter weight showed that herbicide reduction could effectively control weed growth and would not hinder soybean growth. This study can provide a more accurate, efficient, and intelligent solution for weed detection in soybean fields, thus promoting herbicide reduction and providing guidance for exploring efficient herbicide application techniques.