Weed identification in soybean seedling stage based on UAV images and Faster R-CNN

The natural environment in which field soybeans are grown is complex in terms of weed species and distribution, and a wide range of weeds are mixed with soybeans, resulting in low weed recognition rates. Weeds compete with soybeans for sunlight, water and nutrients, and if not managed in a timely manner, weeds may impede soybean growth and reduce yield. The seedling stage is the early stage of soybean growth, and the growth status of soybeans and weeds varies greatly, making it easier to identify and manage weeds. In this paper, a field soybean weed recognition method based on low altitude UAV images and Faster R-CNN algorithm is proposed by utilizing soybean seedling stage weed data collected at low altitude by UAV equipment. A dataset containing 4000 images of soybeans, weeds and broadleaf weeds was constructed and generated in PASCAL VOC format. First, the classification effects of four backbone feature extraction networks, ResNet50, ResNet101, VGG16 and VGG19, were compared to determine the optimal structure; second, the aspect ratio distribution and area distribution of the targets in the dataset were analyzed, and a suitable anchoring framework was designed according to the characteristics of the dataset itself, and the target was trained to be able to recognize soybean seedling weeds with different weed densities. detection model; and two classical target detection algorithms SSD, YOlOv3, YOLOV4, and YOLOV7 were compared. This experiment shows that the Faster R-CNN model with VGG16 as the backbone feature extraction network has the optimal recognition accuracy. By analyzing the characteristics of the dataset itself and optimizing the anchor frame parameters, the optimized model has an average recognition accuracy of 88.69 % for a single data frame, and an average recognition time of 310 ms, which can accurately recognize soybean seedlings and weeds of different densities. Comparing the optimized Faster R-CNN with mainstream target detection models, the average accuracy is 6.31 % higher than the SSD model, 5.79 % higher than the YOlOv3 model, 6.8 % higher than YOLOV4, and 2.92 % higher than YOLOV7. The results show that the optimized target detection model in this paper is more advantageous and can provide scientific guarantee for grass damage monitoring and control in UAV scale.