|Batchelor, W - MISSISSIPPI STATE UNIV|
|Paz, J - UNIV OF GEORGIA|
|Kaleita, A - IOWA STATE UNIV, AIMES|
|Dejonge, K - COE, OMAHA, NE|
Submitted to: Transactions of the ASABE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: May 25, 2007
Publication Date: July 15, 2007
Citation: Thorp, K.R., Batchelor, W.D., Paz, J.O., Kaleita, A.L., Dejonge, K.C. 2007. Using cross-validation to evaluate ceres-maize yield simulations within a decision support system for precision agriculture. Transactions of the ASABE. 50(4):1467-1479. Interpretive Summary: Agricultural systems models are important tools for understanding the processes occurring within agricultural systems and adjusting management practices to improve productivity and reduce environmental impacts of production agriculture. Researchers have recently developed a decision support system that automates DSSAT crop growth model simulations across management units in agricultural fields. In our work, a cross-validation methodology was develop and tested for evaluating crop growth simulations within the context of this precision agriculture decision support system. Use of the cross-validation methodology to evaluate CERES-Maize for a cornfield in central Iowa demonstrated that the model performed better as more seasons of measured information became available for model calibration. Lack of adequate measured data for model calibration may be a limiting factor in the practical use of this decision support tool for precision management of crop production inputs. The performance of the model was also shown to vary spatially in relation to topography, indicating that the model was limited in its ability to simulate the influence of this soil system property on crop yield. Development of a strategy to evaluate crop growth models within the context of this decision support system is of great benefit to researchers in the precision agriculture area, especially those who are using agricultural systems models as tools for developing precision crop management strategies.
Technical Abstract: Crop growth models have recently been implemented to study precision agriculture questions within the framework of a decision support system (DSS) that automates simulations across management zones. Model calibration in each zone has occurred by automatically optimizing select model parameters to minimize error between measured and simulated yield over multiple growing seasons. However, to date, there have been no efforts to evaluate model simulations within the DSS. In this work, a model evaluation procedure based on leave-one-out cross-validation was developed to explore several issues associated with the implementation of CERES-Maize within the DSS. Five growing seasons of measured yield data from a central Iowa cornfield were available for cross-validation. Two strategies were used to divide the study area into management zones, one based on soil type and the other based on topography. The decision support system was then used to carry out the model calibration and validation simulations as required to complete the cross-validation procedure. Results demonstrated that the model's ability to simulate corn yield improved as more growing seasons were used in the cross-validation. For management zones based on topography, the average root mean squared error of prediction (RMSEP) from cross-validations was 1460 kg ha-1 when two growing seasons were used and 998 kg ha-1 when five years were used. Model performance was shown to vary spatially based on soil type and topography. Average RMSEP was 1651 kg ha-1 on zones of Nicollet loam, while it was 496 kg ha-1 on zones of Canisteo silty clay loam. Spatial patterns also existed between areas of higher RMSEP and areas where measured spatial yield variability was related to topography. Changes in the mean and variance of optimum parameter sets as more growing seasons were used in cross-validation demonstrated that the optimizer was able to arrive at more stable solutions in some zones as compared to others. Results suggested that cross-validation was an appropriate method for addressing several issues associated with the use of crop growth models within a DSS for precision agriculture.