Comparison of Multi-site and Single-site Daily Precipitation and Temperature Extremes Downscaling (Case Study: Southern Coast of the Caspian Sea)

To characterize the linkage between large-scale climate conditions and local climate variability, statistical downscaling techniques have been frequently used in climate-related studies. Different single-site and multi-site approaches to downscaling methods are applied in this regard. Most of the studies, however, have been mainly dealing with downscaling of climatic processes for a specific site or watershed average, but few studies are concerned with the multi-sites downscaling techniques because of the complexity in accurately describing both observed at-site temporal persistence and spatial dependence between different variables and locations (Khalili et al. 2013; Chen et al., 2017).
In this study, in order to comparison of multi-site and single-site daily precipitation and temperature extremes downscaling, two methods of Single-site Quintile Mapping (SSQ) and multi-site Modular Expanded Downscaling (ModExDs) (Cannon, 2013) were applied to a set of 5 synoptic stations located within Southern Coast of the Caspian Sea, Iran. The SSQ downscaling technique is based on application of the quantile mapping bias correction step from the Bias Correction Spatial Disaggregation (BCSD; Wood et al. 2002) downscaling algorithm directly to daily GCM data, i.e., without spatial and temporal disaggregation (Bürger et al. 2013). In this study, quantile mapping algorithm with delta method extrapolation for nonlinear bias correction is applied. Expanded Downscaling (XDS) is a perfect prognosis technique which maps large-scale atmospheric fields to local station data. The XDS method is based on defining a multivariate linear regression between predictors and predictands, extended by the side condition that the local co-variability between the variables and stations is preserved (Sunyer et al.). The ModExDs, as a modular of XDS which is implemented in R by Cannon (2013), is performed here.
The predictands include daily time series of precipitation and temperature extremes for the 1961-2013 period that leads to create a training set consisting of the first 30 years data, and a test set consisting of the remaining observations. Same variables of the NCEP/NCAR (National Centers for Environmental Prediction/National Centre for Atmospheric Research) reanalysis dataset were considered as climate predictors. Missing values of observed data are filled in by using the sequential k-nearest neighbor imputation method (Kim and Yi 2008) and homogeneity tests, of 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), are applied both before and after infilling to assess infilling performance.
The methods sensitivity to large-scale anomalies and their skill in replication of the observation data distribution during the validation period (1991-2013) are tested, according to the 27 Climate Extremes Indices (CLIMDEX), using Pearson correlation and Kolmogorov–Smirnov (KS) tests, respectively. Combined tests are used to assess overall model performances.
The results showed that the multi-site method of ModExDs was able to pass 66.7, 48.9 and 33.3 percent and the single-site method of SSQ, passed 63, 48 and 29.6 percent of the Kolmogorov–Smirnov (KS), the Pearson correlation and the combined tests respectively. Therefore, both methods performed well in terms of temperature and precipitation downscaling. However, multi-site method performed better than single-site overlay.
Correlation tests were passed less frequently than KS tests. Both methods downscaled temperature indices better than precipitation indices. According to the Tables 3, 4 and 5, some indices, notably FD, GSL, TN10p, TN90p, TR and DTR, were passed correlation test successfully. Most of the indices related to the precipitation, especially, Rx5day and R10 were not successfully simulated by any of the methods in the region. Model performance varied widely across the study region. Methods performance, however, were better in the Anzali station regarding to the test 1 (corr). More indices were able to pass the test 2 (KS) throughout the region. However, indices such as DTR, FD, TN10p, TN90p, TNN, were not successfully downscaled or appeared to be fairly weak in all stations except for Gorgan in this regard. The indices of GSL and TX90p could pass the combined test throughout the study area. As mentioned before, methods performance varied across the region. So that the methods performed well in Gorgan station, while both performed worse in Rasht station.