County-scale crop yield prediction by integrating crop simulation with machine learning models.

Area of Science:

• Agricultural Science
• Data Science
• Environmental Science

Background:

• Accurate crop yield prediction is crucial for agricultural decision-making but is complicated by numerous interacting factors.
• Previous research successfully combined crop modeling with machine learning (ML) for maize yield prediction in a limited region.

Purpose of the Study:

• To expand the coupled crop modeling and ML approach for maize yield prediction to the entire US Corn Belt (12 states).
• To develop and evaluate new ML models and ensemble models using diverse input data, including soil, weather, management, and historical yield data.
• To conduct spatial analysis to understand the sources of prediction errors.

Main Methods:

• Developed five new ML models and their ensemble models, with and without crop modeling variables.
• Utilized soil, weather, management, and historical yield data as inputs.
• Performed spatial analysis to identify factors contributing to prediction errors.

Main Results:

• Coupling crop modeling with ML significantly increased prediction accuracy.
• The ensemble model achieved a relative root mean square error (RRMSE) of approximately 9% for test years (2018-2020).
• High prediction errors (RRMSE) were associated with areas having low cropland ratios, and specific regions were impacted by soil data and extreme weather events.

Conclusions:

• The integrated crop modeling and ML approach provides a robust method for large-scale maize yield prediction.
• The models demonstrate potential for county-level and, with sufficient data, field-level yield predictions.
• Understanding error sources is vital for refining prediction models and improving agricultural management strategies.