A Comparison of Statistical and Model-Based Downscaling Techniques for Estimating Local Climate Variations

Abstract

The respective merits of statistical and regional modeling techniques for downscaling GCM predictions have been evaluated over New Zealand, a small mountainous country surrounded by ocean. The boundary conditions were supplied from twice-daily European Centre for Medium-Range Weather Forecasts analyses at 2.5° resolution for the period 1980?94, which were taken as the output of a ?perfect? climate model. Daily and monthly estimates of minimum and maximum temperature and precipitation from both techniques were validated against readings from a network of 78 climate stations. The statistical estimates were made by a screening regression technique using the EOFs of the regional height fields at 1000 and 500 hPa, and local variables derived from these fields, as predictors. The model interpolations made use of the RAMS model developed at Colorado State University running at 50-km resolution for 1990?94 only. The model values at the nearest grid point to each station were rescaled using a simple linear regression to give the best fit to the station values. The results show both methods to have comparable skill in estimating daily and monthly station anomalies of temperatures and rainfall. Statistical estimates of monthly departures were better obtained directly from monthly mean forcing than from a combination of daily estimates; however, daily values are needed if one wishes to estimate variability. While there are good physical grounds for using the modeling technique to estimate the likely effects of climate change, the statistical technique requires considerably less computational effort and may be preferred for many applications.