**4. Discussion**

The results confirmed our supposition that bias corrections have a significant effect on climate classification. The effect of the bias correction varied among the models and the regions of the model domains. Table 11 shows the differences between the observed and simulated Köppen−Geiger climate zones in each region. The results were received by the calculation of the number of grid points where the simulated and observed K-G zones were different in each region, and then this number was divided with the number of grid points of the regions. The eQM and gpQM resulted in the largest differences between the RCMs. These differences stemmed from the correction of precipitation. Simulated precipitation is very sensitive to the properties of a model, e.g., physical parameterization, surface properties, and resolution; hence, the distribution of precipitation varied among the RCMs. In the HIRHAM model, eQM and gpQM produced drier negative bias, i.e., dryer zones in almost the entire studied area. The dominance of these dry climate classes originated from the surface properties in the HIRHAM model, the 1 mm threshold value of precipitation and the correction method. Unlike the other RMCs, HIRHAM has only one soil moisture layer [50], which results in a smaller water-holding capacity, which probably causes a negative feedback effect on precipitation formation. Owing to the threshold value, most of the daily mean precipitation values were less than 1 mm, which were resized to zero. This threshold value also caused negative precipitation bias in the JJA season. Moreover, the eQM corrected the ranked category, but not the value of the variable. Hence, the precipitation (or temperature) values transformed into "very high" values correspond to what observations tell us about actual "very high "values [15]. Notwithstanding that the eQM is expected to be the best method according to some literature [15,51,52], but according to some studies, the distribution-based methods improve the RCMs [31,44,53]. The remaining large biases may originate from the weakness of linear extrapolation of the cumulative distribution of parameters.

**Table 11.** Disagreement between observed and simulated K-G zones in eight different regions: the Alps (AL), the British Isles (BI), Eastern Europe (EA), France (FR), the Iberia Peninsula (IP), the Mediterranean (MD), Mid-Europe (ME) and Scandinavia (SC) and in the whole study area in DJF and JJA in the case of eQM-eQM, gpQM-eQM, power transformation of precipitation and variance scaling of temperature and LOCI and variance scaling of temperature bias correction combination. The values are in %.


The model results corrected by the gpQM resulted in a similar climate classification to the eQM corrected simulations, regardless of the gpQM using gamma and generalized Pareto distributions. The remained bias can be explained by the fact that daily precipitation cannot be adequately expressed by gamma distribution for every region of Europe [54].

The power transformation of precipitation and the local intensity scaling of precipitation combined with the variance scaling of temperature performed correct K-G zone distribution with a negligible difference from the observed one. Furthermore, they resulted in very similar values in each of the RCMs. Their independence on the model and regions of the model domain can be explained by the fact that these are distribution-free correction approaches. Furthermore, they are also able to adjust the variance statistics of the precipitation time series, the simulated wet-day intensity, the wet-day frequency of precipitation and the variance and the mean values of temperature.

The bias correction methods were validated through a split-sample test by calculating the K-G zones in the 1981−2000 time period, except for the local intensity scaling of precipitation. According to the climate classification, the power transformation of precipitation and the variance scaling of temperature combination performed best in terms of K-G zones, despite the fact that the eQM bias correction methods had a smaller residual bias value in some RCMs, e.g., in ALADIN in the JJA season.

The bias correction methods were tested by the differential split-sample test in [44]. According to the statistical evaluation of the bias corrections in the test period, they found that the best method was distribution mapping based on gamma distribution, which was able to correct statistical moments other than means and standard deviations. Their findings presumably stemmed from their decision to choose smaller sized domains, in which only one European region was taken into account. We found the eQM and gpQM of precipitation had great limitations in the larger sized pan-European domain and produced incorrect climate classification in each RCM.
