Quantitative predictions of crop yields and prices from satellite-based machine learning

Accurate forecasting of agricultural commodity yields and prices is essential for informed decision-making by farmers, traders, and policymakers. This study presents a novel satellite-based framework that integrates satellite-based Gross Primary Production (GPP) data with advanced machine learning (ML) models, including Autoencoders (AE) and Variational Autoencoders (β-VAE), combined with ElasticNet regression. The framework is applied to produce quantitative predictions of US corn and soybean yields and prices throughout the growing season. For yields, the satellite-based ML models achieved high accuracy, with coefficient of determination (R ² ) values reaching up to 0.90 and 0.75 for corn and soybean, respectively. For prices, the satellite-based ML models significantly outperformed standard statistical time-series models—Exponential Smoothing State Space Model (ETS) and Seasonal Autoregressive Integrated Moving Average (SARIMA)—during early to mid-growing seasons (May to August), achieving R ² values up to 0.48 and 0.46 for corn and soybean, respectively. Conversely, standard statistical models exhibited superior performance during the late-growing season (September), with SARIMA achieving R ² values up to 0.85 and 0.69 for corn and soybean. These findings show that satellite-based ML is highly relevant to forecasting crop yields and prices several months before harvest. However, standard statistical models provide better predictions later in the growing season, a few weeks before harvest. By identifying which models offer the best predictive performance at different crop stages, this study contributes to improving the tools used for crop yield and price forecasting.