1 استادیار اقلیم شناسی/ دانشگاه گیلان
2 دانشگاه تربیت مدرس
3 دانشگاه تهران
عنوان مقاله [English]
Two single and multi-site statistical downscaling methods of Statistical Downscaling Model–Decision Centric (SDSM–DC) for daily temperature and precipitation are evaluated at nine stations located in the mountainous region of Iran’s Midwest. SDSM is best described as a single-site model, but it can be extended to multi-site applications via conditional resampling (CR-SDSM, Wilby et al. 2003; Harpham and Wilby 2005).
SDSM–DC (Wilby and Dawson 2013) is a hybrid of the stochastic weather generator and transfer function methods. Predictor selection is based on empirical relationships between GCM-scale predictors and single-site predictand variables. (Farajzadeh et al. 2015).
Applying SDSM to multi-site daily rainfall downscaling includes two steps: (1) the daily rainfall and temperature at a “marker” site (in this study, the area average amounts) is first downscaled by the single-site SDSM; (2) Daily rainfall amounts are then “resampled from the empirical distribution of area averages, conditional on the large-scale atmospheric forcing and the stochastic error term. The actual daily rainfall is determined by mapping the modeled normal cumulative distribution value onto the observed cumulative distribution of amounts at the marker site” (Wilby et al. 2003; Liu et al. 2013). Ultimately, the marker site rainfall is resampled to the constituent amount falling on the same day from each station in the multi-sites array (Harpham and Wilby 2005). Thus, if the marker series is based on an unweighted average of all sites, the conditional resampling will preserve both the areal average of the marker series and the spatial covariance of the multi-site rainfall (Wilby et al. 2003). Additionally, using area average, instead of individual sites as the marker series, reduces the risk of employing a nonhomogeneous/non-representative record and increases the signal to noise ratio of the predictand (Wilby et al. 2003; Liu et al. 2013). To downscale temperature, the same steps are applied but unconditionally using transfer function methods.
For statistical downscaling, two sets of data are generally required: (1) observational data for model calibration and validation, as predictands; and (2) synoptic-scale climate data from GCM and/or reanalysis, as predictors. In order for a better assessment of climate variability and change on local and regional scale, long-term time series of reliable climate data at fine-scale resolution are required (Vincent et al 2002; Mekis and Vincent 2011; Menne et al 2012). As mentioned before, for the Midwest of Iran, we selected nine synoptic stations with nearly complete data coverage for 1981–2010.
We used station data from two decades (1981–2000) for calibration and from one decade (2001–2010) for validation of daily values of minimum and maximum temperature, and total daily precipitation. To assess the accuracy and homogeneity of the observational data, we used different methods for quality control: the R packages RHtestsV3 (Wang and Feng 2010) and RHtests_dlyPrcp (Wang et al. 2010), based on penalized maximal t and F tests (Wang et al. 2007; Wang 2008b) that are embedded in a recursive testing algorithm (Wang 2008a); the R package Climatol (Guijarro 2012), which applies a type II linear regression model; and SDSM (Wilby and Dawson 2012, 2013) based on reanalysis predictor variables. Missing values are filled in by using the sequential k-nearest neighbor imputation method (Kim and Yi 2008) and homogeneity tests are applied both before and after infilling to assess infilling performance.
Predictor fields are extracted from the National Centers for Environmental Prediction (NCEP) Reanalysis (Kalnay et al. 1996) archives at resolutions of 2.5°×2.5°. As mentioned earlier, SDSM has its own methodology for predictor selection in which EOFs of NCEP reanalysis data over the domain (30° N, 42° E) and (40° N, 52° E) are screened separately for temperature and for precipitation. (Farajzadeh et al. 2015)
Results indicated that the methods are of widely varying complexity, with input requirements that range from single point predictors of temperature and precipitation to multivariate synoptic-scale fields. The period 1981-2000 is used for model calibration and 2001–2010 for validation, with performance assessed in terms of 27 Climate Extremes Indices (CLIMDEX). The sensitivity of the methods to large-scale anomalies and their ability to replicate the observed data distribution in the validation period are separately tested for each index by Pearson correlation and Kolmogorov–Smirnov (KS) tests, respectively. Combined tests are used to assess overall model performances. Single (multi)-site method of SDSM, passing 76%(81%), 16%(7%) and 14% (5%) of the Kolmogorov–Smirnov (KS), the Pearson correlation and the combined tests, performed well in terms of temperature and precipitation downscaling. Single-site method performed better than multi-site one at single sites; however, multisite method performance is better at regional downscaling. Correlation tests were passed less frequently than KS tests. Both methods downscaled temperature indices better than precipitation indices. Some indices, notably R20, R25, SDII, CWD, and TNx, were not successfully simulated by any of the methods. Model performance varied widely across the study region.