Phenomic selection and prediction of maize grain yield from near-infrared reflectance spectroscopy of kernels
High-throughput phenotyping technologies, which can generate large volumes of data at low costs, may be used to indirectly predict yield. We explore this concept, using high-throughput phenotype information from Fourier transformed near-infrared reflectance spectroscopy (NIRS) of harvested kernels to predict parental grain yield in maize (Zea mays L.), and demonstrate a proof of concept for phenomic-based models in maize breeding. A dataset of 2,563 whole-kernel samples from a diversity panel of 346 hybrid testcrosses were scanned on a plot basis using NIRS. Scans consisted of 3,076 wavenumbers (bands) in the range of 4,000–10,000 cm−1. Corresponding grain yield for each sample was used to train phenomic prediction and selection models using three types of statistical learning: (a) partial least square regression (PLSR), (b) NIRS best linear unbiased predictor (NIRS BLUP), and (c) functional regression. Our results found that NIRS data were a useful tool to predict maize grain yield and showed promising results for evaluating genetically independent breeding populations. All model types were successful; functional regression followed by the PLSR model resulted in the best predictions. Pearson’s correlations between predicted and observed grain yields exceeded.7 in many cases within random cross validation. Partial least squares regression also showed promise on independent breeding trials. More research on predicting phenotypic traits from spectra will provide better understanding how NIRS and other phenomic technology can be used in predicting phenotypes of breeding programs.