Machine learning algorithms for soybean yield forecasting in the Brazilian Cerrado

Area of Science:

• Agricultural Science
• Data Science
• Environmental Science

Background:

• Soybean productivity prediction is crucial for agricultural planning in Brazil's Matopiba region.
• Meteorological data from NASA-POWER and yield data from SIDRA/IBGE (2008-2017) were utilized.
• Several machine learning models were assessed for their predictive capabilities.

Purpose of the Study:

• To evaluate and compare various machine learning models for predicting soybean productivity.
• To forecast soybean yield up to one month in advance in the Matopiba agricultural frontier.
• To identify the most effective model for enhancing agricultural decision-making.

Main Methods:

• The study employed five machine learning models: random forest (RF), artificial neural networks, support vector machines with radial basis function (SVM_RBF), linear model, and polynomial regression.
• Cross-validation was used to assess model performance.
• Key performance metrics included R-squared (R²) for precision, root mean square error (RMSE) for accuracy, and mean error of the estimate (EME) for trend.

Main Results:

• The random forest (RF) algorithm exhibited the highest performance, achieving an R² of 0.81, an RMSE of 176.93 kg/ha, and an EME of 1.99 kg/ha.
• The SVM_RBF model showed the lowest performance with an R² of 0.74, an RMSE of 213.58 kg/ha, and an EME of -15.06 kg/ha.
• Predicted average yields by the models fell within the expected range, considering the region's historical average of 2,730 kg/ha.

Conclusions:

• All evaluated machine learning models demonstrated acceptable precision, accuracy, and trend indices for soybean yield prediction.
• These models can be applied to soybean crop yield prediction, considering regional specificities.
• The use of these algorithms serves as a valuable tool for agricultural planning and decision-making in Brazilian Cerrado soy-producing areas.