Técnicas de aprendizado de máquina aplicadas na predição da produtividade de soja no estado do Tocantins

Abstract

This study applies and compares machine learning techniques for predicting soybean yield in the state of Tocantins, using categorized variables via k-means derived from climate data from the NASA POWER platform and soybean yield data obtained from experiments conducted in several municipalities of Tocantins from 2013 to 2023. Five algorithms were implemented: Random Forest, Decision Tree, XGBoost, Support Vector Machine, and Bagging. The models were trained using yield data from municipalities and evaluated through cross-validation using the metrics RMSE, MAE, and R2. The results demonstrated the superiority of ensemble methods, with Bagging showing the best performance (RMSE = 498.64 kg/ha, MAE = 380.07 kg/ha, R2 = 0.724). XGBoost and Random Forest also achieved very similar results. The variable importance analysis revealed that climatic factors—especially precipitation during the reproductive period and solar radiation during the vegetative period—are crucial determinants of productivity. The study concludes that machine learning techniques, particularly ensemble methods, are promising tools for predicting soybean yield and can support agricultural planning and decision-making in the sector, contributing to reducing the high inherent risk of this activity.