Prediction of soybean yields from climate data and vegetation indices using machine learning and neural networks models

Abstract

Accurate soybean yield prediction is crucial to support agricultural planning, supply chain logistics, food security strategies and maximize production. In this study, we evaluated the performance of two machine learning models and three neural networks - Random Forest, XGBoost, Multilayer Perceptron (MLP), Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) - for soybean yield forecasting using climate variables and vegetation indices. The data used covers more than 20 years (2001-2020), including municipal soybean yield records from IBGE, meteorological data from NASA POWER and vegetation indices derived from MODIS satellite images. We implemented and compared four forecasting scenarios: a single general predictor, predictors by state, predictors by climate zone and independent predictors specific to each month. Our results reveal that regionalized modeling, especially by climate zones, significantly improves forecasting accuracy. In this prediction scenario, the MLP model obtained the lowest errors (MAE = 102.9 kg/ha, RMSE = 128 kg/ha and rRMSE(%) = 3.88 ) as well as the best coefficient of determination (R² = 0.83).