Author
FANG, Q - Qingdao University | |
Ma, Liwang | |
Harmel, Daren | |
Yu, Qingzhong | |
SIMA, M - Duke University | |
Bartling, Patricia | |
Malone, Robert - Rob | |
NOLAN, B - Us Geological Survey (USGS) | |
DOHERTY, J - Department Of Primary Industries |
Submitted to: Agronomy
Publication Type: Peer Reviewed Journal Publication Acceptance Date: 5/6/2019 Publication Date: 5/10/2019 Citation: Fang, Q.X., Ma, L., Harmel, R.D., Yu, Q., Sima, M.W., Bartling, P.N., Malone, R.W., Nolan, B.T., Doherty, J. 2019. Uncertainty of CERES-Maize calibration under different irrigation strategies using PEST optimization algorithm. Agronomy. 9(5):241. https://doi.org/10.3390/agronomy9050241. DOI: https://doi.org/10.3390/agronomy9050241 Interpretive Summary: Selection of experimental datasets for model calibration may affect model performance. In this study, a four-year irrigation study with six irrigation treatments was selected for calibrating RZWQM2 (Root Zone Water Quality Model 2) using automatic Parameter ESTimation (PEST) software to identify the most suitable experimental datasets for model calibration. The twenty-four year by treatment datasets were divided into calibration and validation subsets by either treatments (treatment difference) or years (temporal variability). Simulation results showed that the selection of data subsets for model calibration greatly affected the optimized crop cultivar parameters and model predictions of validation data subsets (e.g., grain yield and biomass). The calibration data subsets of one treatment from all 4-years resulted in less variations in maize cultivar parameters compared with the calibration using data subsets of all treatments in one year. However, validation results after calibrating with data subsets of all treatments in one year showed better simulations of both grain yield and biomass and lower variations across treatments, but higher variations across years, compared to the validation results after calibrating with one treatment for all years. The current calibration exercises showed that calibration with multiple years datasets produced lower uncertainty of crop parameters and model predictions, comparing with the calibration with multiple treatments datasets in one year. Technical Abstract: It is important but less explored for calibrating a model with different measured data subsets to obtain lower uncertainty of predictions for validated data subsets. In this study, a four-year irrigation study with six irrigation treatments was selected for calibrating RZWQM2 (Root Zone Water Quality Model 2) using automatic Parameter ESTimation (PEST) software to identify the most suitable experimental datasets for model calibration. The twenty-four year by treatment datasets were divided into calibration and validation subsets by either treatments (treatment difference) or years (temporal variability). Only plant cultivar parameters are optimized in this study. After calibration for each subset, randomly generated cultivar parameter sets (11-164 cases) that produced objective function (phi) values within 1.1 times the minimum phi were selected for uncertainty analysis, which were then used to run the validation data subsets that not used for calibration. Simulation results showed that the selection of data subsets for model calibration greatly affected the optimized crop cultivar parameters and model predictions of validation data subsets (e.g., grain yield and biomass). The calibration data subsets of one treatment for all 4-years resulted in less variations in maize cultivar parameters among the selected phi cases, but similar variations across the six treatment-calibration scenarios, compared with the calibration using data subsets of all treatments in one year. However, validation results after calibrating with data subsets of all treatments in one year showed lower RRMSE (relative root mean square error) values for both simulated grain yield and biomass and lower variations across treatments, but higher variations across years, compared to the validation results after calibrating with one treatment for all years. The current calibration exercises showed that calibration with multiple years datasets (temporal variability) produced lower uncertainty of crop parameters and model predictions, comparing with the calibration with multiple treatments datasets in one year (treatment difference). |