An Example of Augmenting Regional Sensitivity Analysis Using Machine Learning Software. Issue 4 (8th April 2020)

• Record Type: Journal Article
• Authors: Spear, Robert C.; Cheng, Qu; Wu, Sean L.
• Is Part Of: Water resources research. Volume 56:Issue 4(2020)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-04-08
• Subjects: Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2019WR026379
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Regional sensitivity analysis, RSA, has been widely applied in assessing the parametric sensitivity of environmental and hydrological models, in part because of its inherent simplicity. In that spirit, this paper reports an example of an augmented approach to improve its utility in ranking parameter importance, beyond reliance solely on the univariate marginal distributions, to include parametric interactions. Both a deterministic and a stochastic model of the transmission of dengue, an important mosquito‐borne disease, were used to explore the effect of interactions to parameter importance ranking using random forests, a commonly used method based on decision trees. The importance ranking based on random forests was generally consistent with the ranking computed from earlier methods that only examined marginal distributions, but with increased importance shown by several interacting parameters. In addition, and building on an earlier application of tree‐structured density estimation, recently developed software was used to map the regions of the parameter space supporting good fits to calibration data. These methods were also found useful in revealing the scale dependence of sensitivity analysis as well as providing a means of identifying alternative explanations for the observed behavior of the system that remain consistent with calibration criteria, a phenomenon known as equifinality. Key Points: Parameter interactions can be included in assessing parameter importance in regional sensitivity analysis via machine learning methods Tree‐structured density estimation is useful in mapping the regions of the parameter space supporting good fits to calibration data Mapping regions of good fits to calibration data may identify alternative explanations of system behavior in many applications