Xiaojun Jin1, Yanxia Sun2, Jun Che1, Muthukumar Bagavathiannan3, Jialin Yu3,4, Yong Chen1. 1. College of Mechanical and Electronic Engineering, Nanjing Forestry University, Nanjing, China. 2. School of Rail Transportation, Nanjing Vocational Institute of Transport Technology, Nanjing, China. 3. Department of Soil and Crop Sciences, Texas A&M University, College Station, TX, USA. 4. College of Forestry, Nanjing Forestry University, Nanjing, China.
Abstract
BACKGROUND: Precision weed control in vegetable fields can substantially reduce the required weed control inputs. Rapid and accurate weed detection in vegetable fields is a challenging task due to the presence of a wide variety of weed species at various growth stages and densities. This paper presents a novel deep-learning-based method for weed detection that recognizes vegetable crops and classifies all other green objects as weeds. RESULTS: The optimal confidence threshold values for YOLO-v3, CenterNet, and Faster R-CNN were 0.4, 0.6, and 0.4/0.5, respectively. These deep-learning models had average precision (AP) above 97% in the testing dataset. YOLO-v3 was the most accurate model for detection of vegetables and yielded the highest F 1 score of 0.971, along with high precision and recall values of 0.971 and 0.970, respectively. The inference time of YOLO-v3 was similar to CenterNet, but significantly shorter than that of Faster R-CNN. Overall, YOLO-v3 showed the highest accuracy and computational efficiency among the deep-learning architectures evaluated in this study. CONCLUSION: These results demonstrate that deep-learning-based methods can reliably detect weeds in vegetable crops. The proposed method avoids dealing with various weed species, and thus greatly reduces the overall complexity of weed detection in vegetable fields. Findings have implications for advancing site-specific robotic weed control in vegetable crops.
BACKGROUND: Precision weed control in vegetable fields can substantially reduce the required weed control inputs. Rapid and accurate weed detection in vegetable fields is a challenging task due to the presence of a wide variety of weed species at various growth stages and densities. This paper presents a novel deep-learning-based method for weed detection that recognizes vegetable crops and classifies all other green objects as weeds. RESULTS: The optimal confidence threshold values for YOLO-v3, CenterNet, and Faster R-CNN were 0.4, 0.6, and 0.4/0.5, respectively. These deep-learning models had average precision (AP) above 97% in the testing dataset. YOLO-v3 was the most accurate model for detection of vegetables and yielded the highest F 1 score of 0.971, along with high precision and recall values of 0.971 and 0.970, respectively. The inference time of YOLO-v3 was similar to CenterNet, but significantly shorter than that of Faster R-CNN. Overall, YOLO-v3 showed the highest accuracy and computational efficiency among the deep-learning architectures evaluated in this study. CONCLUSION: These results demonstrate that deep-learning-based methods can reliably detect weeds in vegetable crops. The proposed method avoids dealing with various weed species, and thus greatly reduces the overall complexity of weed detection in vegetable fields. Findings have implications for advancing site-specific robotic weed control in vegetable crops.