Uncertainty in Downscaling Across a Continuum of Methods

Climate downscaling is an inherently uncertain process.  Despite this, many applications of climate data use a single downscaling method to determine locally important effects.  There exist a large number of different downscaling methods, spread across a continuum of general approaches, which range from simple statistical bias corrections to extremely complex regional climate models.  Here we examine a range of methods across a variety of climate projections to begin to understand the relative importance of the choice of downscaling method compared to the choice of climate model, emissions scenario, or chaotic variability. We include some of the simpler statistical rescaling methods (BCSD and Asynchronous Regression), a more sophisticated analog regression model conditioned on atmospheric circulation, a quasi‐dynamical downscaling model (The Intermediate Complexity Atmospheric Research model ICAR), and finally a fully non-hydrostatic regional climate model (WRF) run on convection permitting (4km) and convection parameterized (20 km) grids. These downscaling methods are applied over two major river basins to different climate models and different ensemble members of the same model, with the exception of the high-resolution WRF setup, which is too computationally expensive to apply to more than a few climate projections.  We diagnose the sensitivity to method choice of the mean change in precipitation and temperature, as well as the seasonality and physical interpretation (e.g. elevation dependent warming, and changes in orographic precipitation) of that change.  We further test sensitivity of the change signal in ICAR to model physics choice (e.g. microphysics parameterizations and parameters), and the sensitivity of the analog-regression downscaling model to individual modeling decisions such as the number and weighting of analogs, or mathematical transformations used.  These downscaling method sensitivities are then put in the context of the uncertainty expected from different climate models and chaotic internal variability. While the difference between any two climate projections is greater than the difference between different downscaling methods, when averaged across a larger number of projections, the uncertainty due to downscaling method can be greater than the standard error of the mean change signal.