Enhancing genomic selection models through manifold data augmentation in wheat and maize

Genomic selection is a powerful tool for accelerating genetic gain in crop improvement; yet its effectiveness is often constrained by limited phenotypic datasets and suboptimal model generalizability across environments and populations. This study introduces and evaluates three data augmentation strategies—Mixup, CutMix, and the newly-developed haplotype-based HapMix—to address these challenges. Using four crop datasets (Wheat599, Wheat487, Wheat2000, and Maize4500) and four base prediction models (CNN, Cropformer, DEM, and LightGBM), we conducted a multidimensional assessment encompassing phenotypic distribution, population structure, predictive accuracy, genotype feature visualization (t-SNE), and model interpretability (SHAP). Across most traits and models, all augmentation strategies significantly enhanced prediction accuracy, as confirmed by paired t -tests. Among them, HapMix consistently achieved the highest number of statistically significant improvements (up to 15 out of 22 comparisons depending on the model) and delivered the largest correlation gains across key agronomic traits. Mechanistically, exploratory t-SNE visualization indicated that HapMix tends to preserve recognizable core genetic clusters while introducing targeted dispersion, thereby enriching feature diversity without erasing biological structure. SHAP analyses further revealed that augmentation reshapes SNP importance patterns, with HapMix promoting a more distributed, multi-locus contribution pattern that enhances model robustness and interpretability. The relative effectiveness of each augmentation method was modulated by trait complexity and genetic architecture. Collectively, these findings establish HapMix as a superior, biologically informed augmentation strategy for genomic selection, and they provide a practical framework for selecting data augmentation methods to enhance prediction accuracy and generalizability in crop breeding.