DSBEAN: An innovative framework for intelligent soybean breeding phenotype analysis based on various main stem structures and deep learning methods

Although computer vision technology has demonstrated significant efficiency in rapidly identifying soybean phenotypic traits, traditional methods still effective in accurately distinguishing certain complex phenotypes. To further advance the analysis of soybean phenotypic traits, this study proposed the DSBEAN framework that combined soybean breeding technology and deep learning algorithms from a new perspective to analyze soybean phenotypic traits. The key components of DSBEAN framework were two innovative evaluation indicators: the length ratio of the pod growth area to the main stem (MLR) and the pod density within the pod growth area (PD), which were essential for refining understanding and analysis of soybean phenotypic traits in computer vision perspective. The DSBEAN framework consisted of three sections: 1) Main stem Node Detection and PGA Identification. An improved YOLOv5s model was designed for soybean main stem node detection, pod coordinate extraction, and pod growth area (PGA) partition. 2) Main stem Segmentation. The U-Net model was employed for soybean main stem segmentation. 3) MLR and PD Extraction: The previously identified soybean phenotypes were used to calculate the MLR and PD. To validate the DSBEAN framework, a new soybean image and label dataset (SILD) was constructed, and diverse comparison experiments were performed. From the experimental results, the number of pods predicted based on MLR, PD, and the number of main stem nodes reached a correlation level of 0.93, highlighting the significant potential of the DSBEAN framework for soybean phenotype identification. In addition, the proposed framework had the potential to provide new directions for phenotype identification of other crops.