Forecasting Corn Yield With Machine Learning Ensembles.

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Advancements in big data analytics and high-performance computing enhance crop yield prediction.
• Machine learning (ML) offers faster, more flexible crop yield predictions than traditional simulation models.
• Machine learning ensembles can reduce bias and variance, improving overall prediction accuracy.

Purpose of the Study:

• To develop and evaluate a machine learning-based framework for forecasting corn yields in the US Corn Belt.
• To investigate the impact of using partial in-season weather data on prediction accuracy and timing.
• To compare the performance of various ensemble models and identify key predictive weather features.

Main Methods:

• Developed an ensemble machine learning framework using a blocked sequential procedure for out-of-bag predictions.
• Forecasted county-level corn yields in Illinois, Indiana, and Iowa, aggregating results to agricultural district and state levels.
• Evaluated models using complete and partial in-season weather data, including early-season predictions (e.g., June 1st).

Main Results:

• Optimized weighted and average ensemble models achieved the highest precision (9.5% RRMSE).
• Stacked LASSO provided the least biased predictions (53 kg/ha MBE), with other ensembles outperforming base learners.
• Early-season forecasts (June 1st) using partial weather data yielded accurate results (9.2% RRMSE).
• Ensemble models outperformed individual models and benchmark ensembles at all scales (county, district, state).

Conclusions:

• The proposed ensemble framework accurately forecasts corn yields, with early-season predictions being feasible and reliable.
• Weather features from early-to-mid growing season (weeks 18-24) are critical for accurate yield forecasting.
• Ensemble models, particularly weighted and average ensembles, are superior to individual models for time-series crop yield prediction.