Evaluating the stationarity assumption in statistically downscaled climate projections: is past performance an indicator of future results?

Abstract

Empirical statistical downscaling (ESD) methods seek to refine global climate model (GCM) outputs via processes that glean information from a combination of observations and GCM simulations. They aim to create value-added climate projections by reducing biases and adding finer spatial detail. Analysis techniques, such as cross-validation, allow assessments of how well ESD methods meet these goals during observational periods. However, the extent to which an ESD method's skill might differ when applied to future climate projections cannot be assessed readily in the same manner. Here we present a "perfect model" experimental design that quantifies aspects of ESD method performance for both historical and late 21st century time periods. The experimental design tests a key stationarity assumption inherent to ESD methods -namely, that ESD performance when applied to future projections is similar to that during the observational training period. Case study results employing a single ESD method (an Asynchronous Regional Regression Model variant) and climate variable (daily maximum temperature) demonstrate that violations of the stationarity assumption can vary geographically, seasonally, and with the amount of projected climate change. For the ESD method tested, the greatest challenges in downscaling daily maximum temperature projections are revealed to occur along coasts, in summer, and under conditions of Climatic Change (2016)

Multivariate quantile mapping bias correction: an N-dimensional probability density function transform for climate model simulations of multiple variables

“…As the focus is on correction of spatial relationships (i.e., downscaling), Can-RCM4 outputs over the region are degraded spatially so that they are representative of a downscaling rather than pure bias correction application. Following Dixon et al (2016), simulated precipitation outputs are "coarsened" by spatially aggregating from the 0.5-deg NAM-44i grid to a coarser 2.5-deg grid and are then interpolated back onto the 0.5-deg grid. Spatial aggregation induces a fundamental mismatch in spatial scales between the model and observational fields.…”

Some Pitfalls in Statistical Downscaling of Future Climate

“…For example, Dixon et al (2016) found that one SD method reduced the raw GCM error in temperature by about 50% during the historical period and 30%-40% during the future. However, lurking beneath the surface of all SD methods is an implicit assumption that the transfer relationships derived from historical data are valid in a future period during which fundamental aspects of the climate system may have changed.…”

“…However, lurking beneath the surface of all SD methods is an implicit assumption that the transfer relationships derived from historical data are valid in a future period during which fundamental aspects of the climate system may have changed. In more technical terms, this is referred to as the assumption of statistical stationarity (Dixon et al 2016).…”

Analysis of Potential Future Climate and Climate Extremes in the Brazos Headwaters Basin, Texas

“…However, it also has some major drawbacks. The major limitation of the statistical downscaling approach is that it assumes the statistical relationship between the historical climate and GCMs will remain stationary for future periods under climate change [24,26,27]. Details of the limitations and strengths of the dynamical and statistical downscaling approaches are discussed in detail by Wilby and Wigely [24].…”