Improving regional climate change projections of temperature for Halifax, Nova Scotia via statistical downscaling
Lee Titus, Environment Canada/Dalhousie University, Dartmouth, NS, Canada; and R. Greatbatch, I. Folkins, and J. Sheng
In order to best assess the expected climate change impacts on a species, ecosystem or natural resource in a region, climate variables and climate change scenarios must be developed on a regional or even site-specific scale (Wilby et al, 2002). To provide these values, projections of climate variables must be ‘downscaled' from the GCM results, utilizing either dynamical or statistical methods (IPCC, 2001).
In this study the Canadian general circulation model version 3 (CGCM3) run on the IPCC A2 emission scenario was statistically downscaled to Halifax, NS. Observed TMAX (daily maximum temperature, the predictand), taken from Shearwater airport (used as a proxy for Halifax) was selected as the variable to be downscaled in winter (DJF). The predictand and predictors were turned into Z scores. The seasonal cycle was removed from each and a three step data reduction was employed to remove predictors that are not useful or redundant. Then the principal components (PC's) of the predictors from NCEP were calculated for the historical period (1961-2000). The regression was trained on the PC's from 1961-1990 and then validated by predicting 1991-2000 TMAX. Predictability of the regression was assessed via gamma squared (Thompson & Sheng et al, 1997) which is the variance in the errors of prediction divided by the variance of the observations. Gamma squared was found to be 0.21 which suggests good predictive skill. An explained variance of nearly 80 percent was obtained by the developed regression. Once the best regression was created from NCEP, the PC's from the historical CGCM3 predictors were created via projecting the predictors onto the NCEP created eigenvectors. The CGCM3 PC's were then used to hindcast 1961-2000 TMAX. Hypothesis testing showed that both the NCEP PC's and the CGCM3 PC's were able to capture the observed mean and variance of historical TMAX.
To ensure physical realism, the PC that had the highest correlation with TMAX was examined. The parts of that PC with the largest weighting (largest coefficients) are the predictors that play the largest role in governing TMAX. During winter, temperature advection is the dominant forcing on TMAX which explains nearly forty percent of the variation in TMAX.
Finally, the future predictor PC's were created and used to make future projections of TMAX using two different methods (standard method and detrending). The standard technique consists of allowing the trend in the predictors to shift the mean and change the variance. The detrending method detrends all the predictors and allows the mean to shift via the trend in the grid box from the CGCM3 and the regression itself handles the variance. A Gaussian distribution was produced via the mean and variance for three projection periods (2011-2040, 2041-2070, 2071-2100). The 2071-2100 projection shows a mean shift of about 3.5 degrees celcius compared to the historical distribution using the standard technique. The projection for the same period via the detrending method gives a mean shift of near 5 degrees celcius. The variance prediction of the detrending method was slightly larger than that of the standard technique projection.
Extended Abstract (436K)
Session 13A, Regional climate modeling, especially with urban applications
Thursday, 15 January 2009, 11:00 AM-12:00 PM, Room 129A
Previous paper Next paper
Browse or search entire meeting
AMS Home Page