Stacked enviromic–genomic models improve prediction of genotype performance in new environments

Genotype-by-environment interaction is a major challenge for breeding programs, limiting the predictive ability of genomic selection in untested environments. We propose a Stacked Generalization framework that integrates linear mixed models (factor analytic and genomic best linear unbiased prediction), enviromic reaction norms, and machine learning (Extreme Gradient Boosting) to predict phenotypic plasticity. The framework was evaluated on large multi-environment trials of maize and rice, under scenarios that simulate new environments and seasons. The genetic covariance structures differed between crops, requiring a factor analytic model of order k=6 for continental maize and order k = 2 for the local rice network. Across all scenarios, the Stacking ensemble improved on the genomic baseline (M1), with gains in predictive ability from 10% (r = 0.45$ vs. 0.41 for M1) to 27.5% (r = 0.51 vs. 0.40), and reduced the root mean squared error by 30% to 43% relative to the Enviromic Reaction Norm (M2) when ensembles were selected to minimize error. These gains relied on careful feature engineering. Latent variables from genomic and environmental dimensionality reduction (principal component analysis and PaCMAP) and their interactions were the most important features for the machine learning models, and the first genomic PaCMAP component ranked first for both crops. These results indicate that non-linear dimensionality reduction is a promising tool for genomic and enviromic prediction. By combining the stability of mixed models with the flexibility of machine learning, the framework improves robustness, reduces dependence on any single model, and enhances prediction in new environments, supporting cultivation zone expansion and recommending superior genotypes.