Hybrid kernels integrating genomic and multispectral data improve wheat genomic prediction accuracy

Genomic selection (GS) is transforming plant breeding by enabling more accurate and efficient identification of superior genotypes. However, its practical implementation remains challenging, as achieving high prediction accuracy is critical for its success. Several factors-including sample size, the degree of relatedness among individuals, and the complexity of target traits-significantly affect the predictive performance of GS models. To address these limitations, recent studies have explored the integration of genomic and phenomic information to enhance prediction accuracy. This integrated approach has shown promising results and continues to gain empirical support. In this study, we propose an alternative strategy to improve the efficiency and accuracy of GS by constructing hybrid kernels that combine genomic and phenomic information. Specifically, we generate two new kernels by combining the original genomic and phenomic kernels, aiming to capture complementary and previously unexploited sources of variation. We applied this approach to multi-year data from the winter wheat (Triticum aestivum L.) breeding program at Washington State University, using phenomic data collected via unmanned aerial vehicles (UAVs). Our results provide empirical evidence that integrating genomic and UAV-derived phenomic data through hybrid kernel modeling enhances the prediction accuracy of GS models. This approach achieved average improvements of 17.52%, 30.36%, 28.94%, and 16.73% in terms of Pearson's correlation, normalized root mean square error, and the percentage of correctly identified lines within the top 10% and 20%, respectively, compared with the conventional integration of genomic and phenomic information (M4 and M5). These findings highlight the potential of this method as a valuable and scalable tool for modern plant breeding programs.