Computer vision-based phenotyping and yield prediction in soybean

Soybean suffers from subjective inefficiency and error-prone traditional phenotypic measurements. In this study, we used the improved YOLOv8 algorithm to accurately extract phenotypes from single plant soybean images and predict yields with an average accuracy of 97.9% for pod detection. Additionally, the YOLOv8 model incorporating dynamic snake convolution identified main stems and branches with mAP50 of 96.3%. Combined with the midpoint coordinate algorithm (MCA) to measure stem and branch lengths and the dynamic grey encoding algorithm (DGEA) to retain the seed number information to achieve accurate extraction of key phenotypic features. Based on these features, the individual plant weight was predicted by the voting regressor ensemble model, and the R² reached 0.90. This method provides key data support for soybean breeding.