Coupling machine learning and crop modeling improves crop yield prediction in the US Corn Belt

Area of Science:

• Agricultural Science
• Data Science
• Environmental Science

Background:

• Accurate corn yield prediction is crucial for agricultural management and food security.
• Traditional methods often rely solely on weather data, which may not capture complex crop-environment interactions.

Purpose of the Study:

• To evaluate the effectiveness of coupling crop modeling and machine learning (ML) for enhanced corn yield prediction.
• To identify optimal hybrid model configurations and influential crop modeling features for ML integration.

Main Methods:

• Developed and tested five machine learning models (linear regression, LASSO, LightGBM, random forest, XGBoost) and six ensemble models.
• Integrated variables from the Agricultural Production Systems sIMulator (APSIM) crop model as input features for ML models.
• Utilized feature importance analysis to determine the most influential APSIM variables.

Main Results:

• Hybrid crop modeling + ML approaches reduced yield prediction root mean squared error (RMSE) by 7–20% compared to ML models using only weather data.
• Soil moisture-related APSIM variables were most influential, followed by crop and phenology-related variables.
• Simulated average drought stress and average water table depth were identified as key APSIM inputs for improving ML predictions.

Conclusions:

• Coupling crop modeling with ML offers a significant improvement in corn yield prediction accuracy.
• Hydrological and soil-related variables from crop models are essential inputs for enhancing ML-based yield forecasting.
• Future yield prediction models should incorporate more comprehensive hydrological data beyond basic weather information.