Deep learning has gained considerable interest in agricultural breeding research. While advances in sequencing technologies have made genotypic data collection easier in genomic breeding, phenotypic data collection remains labor intensive and time consuming. Furthermore, as traditional phenotypic data collection relies heavily on manual processes, the results may vary based on the researcher’s skill and criteria. Thus, automated phenotypic data collection is essential for addressing these challenges. In this study, we aimed to develop a deep learning model using the YOLOv8 framework to measure the lengths of hypocotyls and roots in sprout vegetables such as mung bean, cowpea, and soybean. Our model automates the measurement process, accurately identifies the hypocotyl and root using Roboflow, and subsequently measures their lengths with high precision in various legume species. This approach addresses the challenges of extensive phenotypic data collection, which is essential for genetic breeding and agricultural improvement. Our deep learning model facilitates consistent and accurate data collection in large-scale studies by controlling variables influenced by the researcher’s skills and criteria. This reduces errors and enhances data reliability and accuracy, which are crucial for successful breeding practices and agricultural research.

Image-based digitalization of germplasm stock holds significant promise for accelerating plant breeding and crop improvement. This technology facilitates efficient germplasm characterization, evaluation, and management through the capture and analysis of visual phenotypes. However, widespread adoption is hindered by challenges that include image quality control, data analysis complexity, and phenotypic representation limitations. This study investigated these constraints and proposed strategies to address them. By managing technical challenges, refining phenotypic data extraction, and developing robust data analysis pipelines, researchers can fully leverage image-based digitalization to enhance germplasm utilization and contribute to sustainable agriculture.

Fast and accurate selection is essential for breeding to cope with rapid climate changes and a steeply increasing population. Consequently, technologies for high-throughput phenotyping (HTP) are emerging. These technologies, unlike conventional phenotyping methods, enable us to evaluate agronomic traits in a fast and massive manner. Thus, the HTP facility was built to acquire and analyze crop images using RGB sensors at the National Institute of Agricultural Sciences, Republic of Korea. By testing various conditions to acquire images, we determined the conditions for phenotyping using the RGB sensor as follows: exposure 30,000 ms, gamma 75, and gain 100 using LED lights in a blue background. Based on this condition, images from 96 individual plants of rice Dongjin cultivar were obtained every week to measure plant height and shoot area, which are directly associated with yield. The results obtained from the image analysis were compared with the manually collected results. The r² value between the projected plant height obtained from image analysis and the plant height obtained from manual measurement was 0.989. Furthermore, the r² value between the projected shoot area obtained from image analysis and the shoot area obtained from manual measurement was 0.981. These results show that image analysis is highly reliable and can be used for crop phenotyping. Therefore, we expect that the new method we developed will be used for breeding in the near future.