MaizeEar‐SAM: Zero‐shot maize ear phenotyping

Abstract

Abstract Quantifying the variation in yield component traits of maize ( Zea mays L.), which collectively determine the overall productivity of this globally significant crop, is critical to plant genetics research, plant breeding, and the development of improved agronomic practices. Grain yield per acre is a function of the number of plants per acre ears per plant number of kernels per ear average kernel weight. Number of kernels per ear is the product of the number of kernel rows per ear ( kernel‐row‐number ) and the number of kernels per row ( kernels‐per‐row ). Traditional manual approaches to quantifying kernels‐per‐row, which exhibits strong environmental sensitivity, are time‐consuming, limiting the scale of data collected for either trait. Recent automation efforts using image processing and deep learning face challenges of high annotation costs and limited generalizability across diverse germplasm. We address these limitations by exploring large vision models for zero‐shot, annotation‐free maize kernel segmentation. Utilizing an open source large vision model—Segment Anything Model—we segment individual kernels in images of maize ears and implement a graph‐based algorithm to calculate kernels‐per‐row. Our approach effectively identifies kernels‐per‐row across diverse maize ears, demonstrating the potential of zero‐shot foundation vision models coupled with image processing routines to enhance automation and reduce subjectivity in agronomic data collection. All our codes are open‐sourced to democratize these frugal phenotyping approaches.