Sensitivity-guided reduction of parametric dimensionality for multi-objective calibration of watershed models

Problem complexity for watershed model calibration is heavily dependent on the number of parameters that can be identified during model calibration. This study investigates the use of global sensitivity analysis as a screening tool to reduce the parametric dimensionality of multi-objective hydrological model calibration problems while maximizing the information extracted from hydrological response data. This study shows that by expanding calibration problem formulations beyond traditional, statistical error metrics to also include metrics that capture indices or signatures of hydrological function, it is possible to reduce the complexity of calibration while maintaining high quality model predictions. The sensitivity-guided calibration is demonstrated using the Sacramento Soil Moisture Accounting (SAC-SMA) conceptual rainfall–runoff model of moderate complexity (i.e., up to 14 freely varying parameters). Using both statistical and hydrological metrics, optimization results demonstrate that parameters controlling at least 20% of the model output variance (through individual effects and interactions) should be included in the calibration process. This threshold generally yields 30–40% reductions in the number of SAC-SMA parameters requiring calibration – setting the others to a priori values – while maintaining high quality predictions. Two parameters are recommended to be calibrated in all cases (percent impervious area and lower zone tension water storage), three parameters are needed in drier watersheds (additional impervious area, riparian zone vegetation, and percent of percolation going to tension storage), and the lower zone parameters are crucial unless the watershed is very dry. Overall, this study demonstrates that a coupled, multi-objective sensitivity and calibration analysis better captures differences between watersheds during model calibration and serves to maximize the value of available watershed response time series. These contributions are particularly important given the ongoing development of more complex integrated models, which will require new tools to address the growing discrepancy between the information content of hydrological data and the number of model parameters that have to be estimated.