Single-kernel Near-infrared Reflectance Spectroscopy for Prediction of Phytoglycogen Content and Other Sweet Corn Seed Composition Traits

Abstract

Breeding for improved seed composition traits such as starch and sugars is constrained by the destructive and time-intensive nature of conventional phenotyping methods. This study evaluated the use of single-kernel near-infrared reflectance (skNIR) spectroscopy as a high-throughput alternative for seed composition analysis in sweet corn. Spectral data from 130 diverse genotypes were collected with subsequent quantification of phytoglycogen (PG), sugars, and starch using wet chemistry. These data were used to develop a partial least squares regression model for predicting PG content and enabling both genotype and individual (at the kernel level) selection. In addition, predictive models were developed for six additional traits: kernel weight, starch, sucrose, glucose, total sugars, and total carbohydrates. Multiple spectral preprocessing approaches were tested, with optimal performance achieved by averaging spectra from six kernels per genotype and applying normalization. Prediction accuracies for most traits were high (0.70 to 0.97). For the PG trait, validation accuracy at the genotype level reached 0.77, which is considered high in the context of complex seed composition traits and limited sample size. These results demonstrate that skNIR provides a viable, nondestructive approach for estimating seed composition traits and has the potential to enhance selection efficiency in breeding programs.