Understanding Prediction Robustness of the Root Zone Water Quality Model (RZWQM).

The RZWQM is often used to make predictions of environmental behavior such as hydrologic and chemical response, translating research to other locations and climates. Typically, if the calibrated model replicates measured field data “reasonably well” it is assumed that predictions under other conditions will also be reasonably good. Unfortunately, this assumption is false, as we have shown for prediction of nitrate loss from a tile-drained, corn-soybean experiment in Northern Iowa. Predictive uncertainty is only reduced if the information content of the calibration dataset is able to constrain the model parameters relevant to the processes controlling the desired prediction.

Using experimental data over 12 years, we investigated the robustness of RZWQM predictions of crop yield, subsurface drainage flow, and nitrate-N loss for multiple model calibrations using the PEST parameter estimation software. Post-processing analyses provided insights into parameter-observation relationships. Not surprisingly, prediction robustness was related to the range of soil moisture conditions in the calibration data. We test the use of the Palmer Drought Severity Index (PDSI) as an indicator of the information content of calibration data related to soil moisture and suggest its use in evaluating the suitability of calibration data for making predictions about other climate conditions.

We show that data representing a particular range of PDSI allow a calibration able to predict performance in years exhibiting a similar range of PDSI. For example, we show that addition to a five year calibration set of a single year identified by examining the PDSI, improves the Nash-Sutcliffe model efficiency coefficient (NSE) from -0.18 to 0.7, and achieves nearly all of the improvement possible when all available observation are included in calibration. Our work shows how field observations under more variable soil moisture conditions constrain the RZWQM parameters and suggests one way of evaluating the predictive power of a calibration.