From PLS to Machine Learning: Advancing the GIS-FA Model for Hybrid Performance Prediction

This study presents an innovative extension of the Geographic Information Systems Factor Analytic (GIS-FA) framework, integrating machine learning (ML) to enhance the spatial prediction of tropical maize hybrid performance in untested environments. The GIS-FA model uses Partial Least Squares (PLS) regression to relate environmental covariates to factor loadings, assuming linear relationships. In this study,  Random Forest (RF) and Extreme Gradient Boosting (XGBoost) were employed to capture the complex environmental dependencies more effectively. Data from 60 tropical maize hybrids evaluated in 25 environments from Embrapa multi-environment trials conducted between 2015 and 2017 were analyzed. The genotype-by-environment (G×E) interaction was modeled using a four-factor structure, which explained 77.08% of the total G×E variance. Environmental loadings were predicted using three approaches: PLS (ρ = 0.2008), RF (ρ = 0.2138), and XGBoost (ρ = 0.2381). Compared with PLS, RF increased accuracy by approximately 6% and XGBoost by around 19%. These results demonstrate the ability of ML to model nonlinear and high-order interactions between climatic and soil covariates. Therefore, by combining ML approaches with the GIS-FA, we were able to enhance our predictive model, providing more precise and spatially explicit insights to guide hybrid recommendation and decisions in tropical maize breeding programs.