Modeling genotype‐by‐environment interaction for variety recommendation in tropical wheat

Abstract The complexity of genotype‐by‐environment interaction reinforces the need for robust methods in cultivar recommendation. Its unpredictability demands prediction under untested conditions, a challenge intensified by climate change. Breeding programs must anticipate, before harvest, which genotype–management combinations will be most effective. Here, we present a modification of the geographic information system–factor analytic (GIS‐FA) predictive approach that integrates GIS techniques, multi‐management FA models, partial least squares regression, and environmental covariates (ECs) to predict phenotypic performance in unevaluated environments. The dataset comprised 91 inbred lines and 12 commercial cultivars from the Wheat Breeding Program at the Federal University of Viçosa (UFV), evaluated across 6 years and seven locations (25 environments). An environmental dissimilarity map indicated that southern regions were more similar to the experimental network, whereas northern areas exhibited higher dissimilarity values, reflecting stronger environmental divergence from the observed trial conditions. Integrating 83 ECs improved yield predictions. In rainfed conditions, the pixel‐by‐pixel model predicted an average yield of 4183 kg ha , surpassing the best genotype (UFV_19013) selected by the factor analytic selection tools by 9.8%. Under irrigation, the yield gain reached 13.4%, with 4756 kg ha compared to 4195 kg ha for UFV_19042. These results highlight the potential of spatially explicit approaches for more accurate and site‐specific recommendations.