Improving hybrid breeding efficiency in winter oilseed rape under sparse multi-environment testing.

Balanced, relationship-based omission, multi-environment predictions, and pedigree-genomic matrices improve genomic predictions, highlighting breeder's ability to influence performance through strategic training set design and model choice. Genomic prediction is increasingly applied in plant breeding, yet its robustness under sparse testing in hybrid breeding programs of winter oilseed rape (Brassica napus L.) lacks evidence. We evaluated the effects of data omission strategies, prediction schemes and models on genomic prediction accuracy and selection consistency. Seven omission scenarios were assessed for three traits with contrasting genetic architectures, using the correlation between observed and predicted phenotypes and Cohen's κ as complementary metrics for genomic prediction accuracy and selection consistency. Both declined systematically with increased data omission, although both metrics only correlated moderately. Balanced omission across environments outperformed unbalanced strategies by stabilizing variance estimation and preserving heritability. Our relationship-informed omission strategy further improved prediction performance, yielding higher and more stable prediction accuracies. Models based on combined pedigree-genomic relationship matrices enabled the inclusion of non-genotyped parental lines and further improved prediction performance. Multi-environment predictions using GCA + SCA models were consistently superior to per-environment predictions and GCA-only models; however, responses were trait-specific. Overall, our results demonstrate that informed omission strategies and model choices substantially influence the performance of genomic prediction under data reduction. Relationship-based balanced omission represents a promising selection approach that balances prediction performance with resource efficiency.