A novel high-throughput hyperspectral scanner and analytical methods for predicting maize kernel composition and physical traits
Jose I. Varela, Nathan D. Miller, Valentina Infante, Shawn M. Kaeppler, Natalia de León, Edgar P. Spalding
Abstract
Large-scale investigations of maize kernel traits important to researchers, breeders, and processors require high throughput methods, which are presently lacking. To address this bottleneck, we developed a novel flatbed platform that automatically acquires and analyzes multiwavelength near-infrared (NIR hyperspectral) images of maize kernels precisely enough to support robust predictions of protein content, density, and endosperm vitreousness. The upward facing-camera design and the automated ability to analyze the embryo or abgerminal sides of each individual kernel in a sample with the appropriate side-specific model helped to produce a superior combination of throughput and prediction accuracy compared to other single-kernel platforms. Protein was predicted to within 0.85% (root mean square error of prediction), density to within 0.038 g/cm3, and endosperm vitreousness percentage to within 6.3%. Kernel length and width were also accurately measured so that each kernel in a rapidly scanned sample was comprehensively characterized.