Multivariate Downscaling of Decadal Climate Change Projections over the Sunbelt

Abstract:

Bias Correction and Statistical downscaling (BCSD) of precipitation and temperature are commonly required to bring the large scale variables available from GCMs to a finer grid-scale for ingesting them into watershed models. Most of the currently employed procedures on BCSD primarily consider a univariate approach by developing a statistical relationship between large-scale precipitation/temperature with the local-scale precipitation/temperature ignoring the interdependency between the two variables. In this study, an asynchronous Canonical Correlation Analysis (CCA) approach is proposed for downscaling multiple climatic variables by preserving the temporal correlations among them. The method was first applied on historical runs of climate model inter-comparison project-5 (CMIP5) for the period 1951-1999 and compared with bias corrected dataset using univariate approach from Bureau of Reclamation. Further, the method was applied on decadal runs of CMIP5 models and compared with univariate asynchronous regression results. A metric, fraction bias was defined, and distribution of fraction bias from ensemble was considered for comparing with univariate approach. CCA relatively performs better in preserving the cross-correlation at grids where observed cross correlations are significant, while reducing fraction biases in mean and standard deviation.