Machine learning for soybean yield forecasting in Brazil

Brazil supplies half of the world's exported soybeans. Forecasting its national soybean yield before harvest could help mitigate disruptions in food supply. The objective of this study is to develop a national soybean yield forecasting system for Brazil based on machine learning (ML) models. Twenty years (2001–2020) of municipality yield, the Oceanic Niño Index (ONI), remote sensing, and gridded daily climate data across the Brazilian soybean cultivation area were used to train ML models in order to estimate municipality soybean yield. Five different ML approaches and their Ensemble were tested: Linear (Ridge) Regression (LR), Random Forest (RF), Gradient Boosted Trees (XGB), Artificial Neural Network (ANN) and Long Short-Term Memory Network (LSTM). Soybean yield forecasting performance varied according to ML model and location. The best performance in estimating municipal soybean yield was achieved with ANN and an Ensemble model with an average rRMSE of 16%, varying from 4% in central-northern to >30% in southern Brazil. Yield was simulated on the municipality level, and its weighted aggregation was used to estimate the yield on the state and national level. Estimations deviated from the observations by an rRMSE from 4.7% to 18.6% (state level) and 4.8% to 6.7% (national level) with the model Ensemble showing the best results, followed by ANN. National soybean yield is forecasted mid-season with an rRMSE of about 6% by end of December, three months prior to crop harvest in March. The accuracy and uncertainty of such in-season forecasts further improve towards the end of the season. The proposed soybean yield forecasting system is transferable to other countries and could help policymakers and food traders plan strategies ahead of harvest.