![[logo] US EPA](http://www.epa.gov/epafiles/images/logo_epaseal.gif){kind=logo}
Statistical Climate Downscaling
Research Programs
Climate & Air Quality
Statistical downscaling methods use correlations among observed and modeled meteorological variables to predict regional and/or local patterns and events that are likely to occur based on the broader-scale GCM simulations. Typically, these approaches do not use the same detailed information that is used in dynamical downscaling, such as physical equations, orographic data, or extensive land-use information. The advantages of statistical downscaling methods lie in their efficiency and speed, and these methods could be particularly attractive if numerous climate scenarios need to be investigated. Statistical methods are not limited by the resolution achievable by the nested regional dynamical model. Thus, statistical methods possibly could be used to gain a better understanding of fine-scale variability, even down to point locations, given high-resolution training data.
It has been reported in the literature that the performances of dynamical and statistical downscaling are comparable for current climatic conditions. However, it is questionable whether statistical models can perform as well under future conditions (Wilby et al., 2002) because statistical downscaling methods rely on associations among meteorological variables. These relationships do not explain all of the inherent variability in atmospheric phenomena; in fact, the choice of variables to be used as the "predictors" in such approaches can be a difficult part of the statistical downscaling process. Once a statistical model has been developed for a particular time period (e.g., using current climate), it is unclear whether the relationships it incorporates will remain the same under different climatic conditions (e.g., in future decades). (See Schmith, 2008 for more discussion.) However, statistical downscaling does make this "stationarity" assumption as it extrapolates to future conditions.
During the past year, EPA has been reviewing, improving, and testing statistical downscaling methods, with an emphasis on evaluation and calibration. Most of this research is being done using a regression-based downscaling technique proposed by Hoar and Nychka (2008), which models the response variable (i.e., surface temperature) as a function of fine-scale temperature estimates obtained by applying a thin-plate spline interpolation technique to the output of a GCM (or other coarse-resolution model). This allows us to bypass the difficulties associated with predictor selection (a difficult issue with many regression-based methods), so that we can focus on issues associated with calibration, training period selection, and quality and quantity of training data. If the efficiency and accuracy of this method proves suitable, this method could be adapted to use different response variables or different predictors. In addition, EPA is planning future work to better understand the relative strengths and weaknesses of statistical and dynamical downscaling for various applications. Research questions of particular interest include:
- Assessing whether regression relationships change substantially over time (Do the regression relationships estimated based on training data yield accurate estimates for years immediately following the training window? How about for 30 years or so into the future?)
- Investigating whether the choice of training window has a substantial impact on the estimated relationships (Does the training period of 1930-1959 yield different estimates for the year 1960 than does a training period of 1961-1990?)
- Accounting for the impact of autocorrelation (temporal dependence) in the training data on the statistical estimates. (Substantial autocorrelation can affect the validity of our error estimates if not identified and adjusted for.)
- Developing at least a rough understanding of how the uncertainty may change when applied to future-year GCM simulations.
- Identifying the relative strengths/weaknesses of the dynamical and statistical approaches to downscaling.
- Determining whether hybrid downscaling approaches may be able to capitalize on the strengths of both dynamical and statistical methods.
Contact: Jenise Swall
References:
Hoar, T., Nychka, D. (2008),
Statistical downscaling of the community climate system model (CCSM) monthly temperature and precipitation projections
Schmith, T. (2008),
Stationarity of regression relationships: Application to empirical downscaling, Journal of Climate, 21, 4529-4537.
Wilby, R.L., Dawson, C.W., Barrow, E.M. (2002),
SDSM - a decision support tool for the assessment of regional climate change impacts, Environmental Modelling & Software, 17, 147-159.