Shaik, A.; Sacks, E.; Leakey, A. D. B.; Zhao, H.; Kjeldsen, J. B.; Jorgensen, U.; Ghimire, B. K.; Lipka, A. E.; Njuguna, J. N.; Yu, C. Y.; Seong, E. S.; Yoo, J. H.; Nagano, H.; Anzoua, K. G.; Yamada, T.; Chebukin, P.; Jin, X.; Clark, L. V.; Petersen, K. K.; Peng, J.; Sabitov, A.; Dzyubenko, E.; Dzyubenko, N.; Glowacka, K.; Nascimento, M.; Campana Nascimento, A. C.; Dwiyanti, M. S.; Bagment, L.; Proma, S.; Garcia-Abadillo, J.; Jarquin, D.

Environmental factors affect crop growth and development thus their consideration across sites and years become essential for genotypic evaluation. Genomic selection (GS) has been broadly implemented to accelerate breeding cycles by skipping field evaluations thus allowing early identification of outperforming genotypes. In this study, 7,740 phenotypic records corresponding to 516 Miscanthus sacchariflorus genotypes evaluated in five locations across three years were considered for analysis. Additionally, environmental data on six weather covariates was implemented to characterize similarities between locations. Different sets of locations of variable sizes were used for model calibration based on two cross-validations (CV00 and CV0) schemes leaving out one location at a time. Predictive ability across locations of the best model varied between 0.45 and 0.90 for both schemes. These results were compared to associate predictive ability in function of weather patterns between training and testing sets to allow model's calibration optimization. We found it is feasible to optimize resource allocation by considering environmentally correlated sets. In most cases, the information from only one and, at most, two locations were enough to deliver better results than using all four locations, reducing training sets by up to 75%. The results obtained shed light on helping breeders make informed decisions considering weather data when designing evaluations.