| Texto completo | |
| Autor(es): |
Wei, Marcelo Chan Fu
;
Molin, Jose Paulo
;
Longchamps, Louis
Número total de Autores: 3
|
| Tipo de documento: | Artigo Científico |
| Fonte: | COMPUTERS AND ELECTRONICS IN AGRICULTURE; v. 243, p. 16-pg., 2026-03-01. |
| Resumo | |
Machine learning (ML) has emerged as a transformative tool in agriculture, yet its adoption is hindered by the inherent trade-off between predictive accuracy and interpretability. This study evaluates three ML models-Decision Tree (DT), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)-for sugarcane yield prediction, integrating Explainable AI (XAI) techniques (SHapley Additive exPlanations-SHAP and Local Interpretable Model-agnostic Explanations-LIME) to enhance model transparency. Feature importance was assessed across models using a dataset of similar to 1,600 ha from Brazil, encompassing soil properties, crop characteristics, and yield data. Results indicate that DT offers the highest interpretability but lower accuracy, while XGBoost achieves the highest predictive performance (R-2 = 0.94, RMSE = 9.39 Mg ha(- 1)). Notably, only one feature (high-yielding variety RB985476) was consistently identified as positively influential across all models, whereas aluminum content and production environment C exhibited negative impacts. Mixed effects were observed for nutrients like phosphorus and potassium, highlighting context-dependent interactions. This study advocates for a hybrid approach, combining interpretable models (DT) with high-accuracy ensembles (XGBoost or RF or both) to optimize decision-making in precision agriculture. This framework can bridge the gap between model transparency and predictive power, offering actionable insights for sustainable sugarcane management by looking for convergent outcomes from different methods. (AU) | |
| Processo FAPESP: | 19/07665-4 - Centro de Inteligência Artificial |
| Beneficiário: | Fabio Gagliardi Cozman |
| Modalidade de apoio: | Auxílio à Pesquisa - Programa eScience e Data Science - Centros de Pesquisa Aplicada |