Improving Crop Model Inference Through Bayesian Melding With Spatially Varying Parameters.

Area of Science:

• Agricultural Science
• Computational Science
• Statistical Modeling

Background:

• Crop Simulation Models (CSMs) are crucial for precision agriculture, aiming to explain crop performance variability and inform site-specific management.
• Accurate CSMs require detailed soil, climate, management, and genetic data, which are often prohibitively expensive to obtain at high spatial resolutions.

Purpose of the Study:

• To develop a Bayesian modeling framework that integrates CSMs with sparse yield monitoring data.
• To provide location-specific posterior predicted distributions of crop yield and unobserved spatially varying CSM parameters.
• To facilitate process-based explanations for observed yield variability.

Main Methods:

• A Bayesian melding framework combining a CSM (CERES-Wheat) with sparse yield data was proposed.
• The model incorporates a systemic component from the physical CSM and a residual spatial process to correct bias.
• Multivariate and univariate Gaussian processes model spatially varying inputs and residual components, with dimension reduction via low-rank predictive processes.

Main Results:

• The framework successfully melds CSM outputs with sparse yield data to generate location-specific yield predictions.
• It provides insights into spatially varying model parameters, aiding in understanding yield variability drivers.
• The use of low-rank predictive processes effectively reduced computational burden for large datasets.

Conclusions:

• The proposed Bayesian melding approach offers a viable solution for integrating CSMs with sparse yield data in precision agriculture.
• This method enhances the explanation of spatial yield variability and supports informed site-specific management decisions.
• The framework demonstrates practical application using the CERES-Wheat model and yield data from Foggia, Italy.