A time-dependent parameter estimation framework for crop modeling.

Area of Science:

• Agricultural Science
• Computational Science
• Machine Learning

Background:

• Crop model performance relies heavily on accurate parameter calibration.
• Estimating time-dependent parameters, like cultivar traits, is complex and often manual, leading to inefficiencies and errors.

Purpose of the Study:

• To develop an automated framework for estimating time-dependent crop model parameters.
• To integrate optimization, machine learning, and agronomic knowledge for enhanced parameter estimation.

Main Methods:

• A parallel Bayesian optimization algorithm was developed for automated parameter estimation.
• The framework was tested using APSIM to simulate historical yield data in the US Corn Belt (1985-2018).
• Performance was compared against standard Bayesian optimization and manual calibration methods.

Main Results:

• The proposed framework reduced prediction error by 11.6% compared to Bayesian optimization and 52.1% compared to manual calibration.
• Machine learning models trained for yield prediction did not outperform the proposed method.
• The framework provided explainable insights into cultivar trait trends over 34 years.

Conclusions:

• The automated framework effectively estimates time-dependent crop model parameters, achieving near-optimal results.
• This approach offers a significant improvement over traditional calibration methods, enhancing crop modeling accuracy and efficiency.