Crop phenotyping of important agronomic traits in field conditions at single-plant resolution has long been a major bottleneck in both genetic analysis (e.g. large-scale association/linkage analysis) and breeding applications (e.g. genomic prediction/selection). Despite growing interest, ultra-affordable, high-throughput and accurate phenotyping tools for maize ears remain limited. Here, we developed OpenEar, an open source, low-cost phenotyping system that combines a DIY maize ear imaging platform with a deep learning-based end-to-end phenotypic data extraction pipeline. The imaging platform is composed of 3D-printed parts and electronics components easily available from local retailers to perform high-quality 360° surface scanning of maize ears. Our pipeline first employs CNN-based models to identify normally-developed ears suitable for phenotyping, followed by reliable segmentation of ears and ear surface projection images by YOLOv11-based models, from which ten key traits are subsequently extracted. OpenEar demonstrates reliable agreement with manual measurements across a diverse set of ear- and kernel-related traits, including ear length (R2 = 0.972), ear diameter (R2 = 0.905), ear volume (R2 = 0.976), ear weight (R2 = 0.878), kernel number (R2 = 0.98), kernel row number (R2 = 0.888), kernel number per row (R2 = 0.852), kernel thickness (R2 = 0.705), kernel width (R2 = 0.515), and thousand kernel weight (R2 = 0.605). A user-friendly graphical interface is developed for manual inspection of ears after computer annotation. Manually annotated ear videos and images are publicly released as a resource for the crop phenomics community. Our study highlights the potential of DIY-based low-cost solutions to make phenotyping more accessible in crop genetic analysis and breeding.
Fan et al. (Sun,) studied this question.