Transformer-based modeling of maize yield and biomass using integrated agronomic and management traits

Abstract

ABSTRACT Integrating categorical factors such as planting systems, fertilizer-application timing, and other management-related traits into yield and biomass prediction frameworks remains relatively underexplored. This study applied transformer-based architectures to predict maize grain yield and biomass using combined categorical and numerical features, with model hyperparameters optimized through Bayesian optimization. Four levels of planting systems and three levels of foliar nano-fertilizer application were incorporated as categorical variables, alongside multiple quantitative traits, to model yield and biomass performance. The results showed that the Tab-Transformer achieved the most accurate and stable predictions, demonstrating strong generalization across test datasets. For grain yield prediction, the Tab-Transformer produced the highest coefficient of determination (R 2 = 0.93), exceeding the FT-Transformer (R 2 = 0.91) and the ANN (R 2 = 0.77) and Random Forest (R 2 = 0.88). A similar pattern was observed for biomass prediction, where transformer-based models (R 2 = 0.47) outperformed the conventional ANN and Random Forest (R 2 = 0.30 and 0.40 respectively), although the lower accuracy indicates that biomass is a more complex and variable trait that is inherently more difficult to predict. SHAP feature-importance analysis of the best model revealed that biomass-related traits accounted for the largest proportion (45.2%) of predictive accuracy for grain yield, followed by management factors (39.8%), agronomic traits (11.6%), and physiological traits (3.4%). This hierarchy indicates that morphological and yield-associated attributes play a dominant role, while management practices such as fertilizer timing and planting systems significantly influence crop performance. Furthermore, UMAP visualization of high-dimensional phenotypic data displayed clear clustering patterns for yield and biomass groups, confirming that transformer, ANN, and Random Forest models effectively captured the underlying multivariate structures governing maize performance. This study reveals the potential of transformer-based frameworks as decision-support tools for optimizing agronomic management and improving maize productivity.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint