Local sensitivity analysis provides computationally frugal ways to evaluate models commonly used for resource management, risk assessment, and so on. This includes diagnosing inverse model convergence problems caused by parameter insensitivity and(or) parameter interdependence (correlation), understanding what aspects of the model and data contribute to measures of uncertainty, and identifying new data likely to reduce model uncertainty. Here, we consider sensitivity statistics relevant to models in which the process model parameters are transformed using singular value decomposition (SVD) to create SVD parameters for model calibration. The statistics considered include the PEST identifiability statistic, and combined use of the process-model parameter statistics composite scaled sensitivities and parameter correlation coefficients (CSS and PCC). The statistics are complimentary in that the identifiability statistic integrates the effects of parameter sensitivity and interdependence, while CSS and PCC provide individual measures of sensitivity and interdependence. PCC quantifies correlations between pairs or larger sets of parameters; when a set of parameters is intercorrelated, the absolute value of PCC is close to 1.00 for all pairs in the set. The number of singular vectors to include in the calculation of the identifiability statistic is somewhat subjective and influences the statistic. To demonstrate the statistics, we use the USDA’s Root Zone Water Quality Model to simulate nitrogen fate and transport in the unsaturated zone of the Merced River Basin, CA. There are 16 log-transformed process-model parameters, including water content at field capacity (WFC) and bulk density (BD) for each of five soil layers. Calibration data consisted of 1,670 observations comprising soil moisture, soil water tension, aqueous nitrate and bromide concentrations, soil nitrate concentration, and organic matter content. All 16 of the SVD parameters could be estimated by regression based on the range of singular values. Identifiability statistic results varied based on the number of SVD parameters included. Identifiability statistics calculated for four SVD parameters indicate the same three most important process-model parameters as CSS/PCC (WFC1, WFC2, and BD2), but the order differed. Additionally, the identifiability statistic showed that BD1 was almost as dominant as WFC1. The CSS/PCC analysis showed that this results from its high correlation with WCF1 (-0.94), and not its individual sensitivity. Such distinctions, combined with analysis of how high correlations and(or) sensitivities result from the constructed model, can produce important insights into, for example, the use of sensitivity analysis to design monitoring networks. In conclusion, the statistics considered identified similar important parameters. They differ because (1) with CSS/PCC can be more awkward because sensitivity and interdependence are considered separately and (2) identifiability requires consideration of how many SVD parameters to include. A continuing challenge is to understand how these computationally efficient methods compare with computationally demanding global methods like Markov-Chain Monte Carlo given common nonlinear processes and the often even more nonlinear models.
Additional publication details
Local sensitivity analysis for inverse problems solved by singular value decomposition