| ABSTRACT | In the context of statistical downscaling of future climate change in the Mediterranean area, regression models are often affected by nonstationary behavior. For this reason a new approach was developed in order to take those varying predictor-predictand-relationships explicitly into account. By means of a Three-step Censored Quantile Regression (TSCQR) precipitation extremes (Ï„=0.9, 0.95, 0.99) were linked to the large-scale circulation, using the Censored Quantile Verification Skill Score (CQVSS) to assess the quality of the regression models.
The Mediterranean area was divided into various seasonal (summer excluded) precipitation regions by means of s-mode principal component analysis (s-mode PCA). For each region and season, based on different selection criteria, a reference station was determined and used as predictand. The large-scale predictors are represented by different circulation- and thermodynamic variables of the NCEP/NCAR reanalysis dataset, which were subjected to an s-mode PCA in order to obtain centers of variations. The PC scores of the centers of variations were used as predictor time series.
By means of the CQVSS, first the best variable combination for each reference station was determined and, subsequently, the significant centers of variations of the respective variables. A combination of at most two predictor variables was most suitable. The introduction of a third variable offered only a small gain of skill. The centers of variations of the respective variables were only taken into account as predictors for further analysis, if they were considered as significant (α=0.01) for at least 95% of 100 31-year random samples.
For each reference station and season, regression models for 31-year running sub-periods, each shifted by one year, were established and, subsequently, validated by means of the remaining years. Here, the running CQVSS of the validation periods indicates the range of variability of the circulation-precipitation relationships. In order to obtain information about possible reasons for non-stationarities within these relationships, composites of the predictor fields were calculated and analyzed for the different sub-periods.
Within the scope of assessing future changes of heavy precipitation events, the composites of the calibration periods were compared with the composites of 31-year running sub-periods of the 21st century for each predictor separately. Subsequently, a multi-correlation coefficient was calculated based on the different states of the atmosphere within the composites of the calibration and the future periods. The regression model of the calibration sub-period, which has the highest correlation coefficient, was then used to assess the quantiles of precipitation in the future. |