Open Access

Gridded maps of meteorological variables are needed for the evaluation of weather and climate models and for climate change monitoring. In order to produce them, values at locations where no observing stations are available need to be estimated from point-wise observations. For the interpolation of meteorological observations deterministic and stochastic methods are often combined. Deterministic methods can account for ancillary information such as elevation, continentality or satellite observations. Stochastic methods such as kriging reproduce observed values at the station locations and also account for spatial variability. In the first two studies of this thesis, a flexible interpolation method for the gridding of locally observed daily extreme temperatures is developed that also provides an optimal estimate of the interpolation ncertainty. In the third study, an observational dataset is created using this interpolation method and then applied to evaluate a climate simulation for Africa. In the first study, the Regression-Kriging-Kriging (RKK) method is tested for the interpolation of daily minimum and maximum temperatures (Tmin and Tmax) in different regions in Europe. RKK accounts for elevation, continentality index and zonal mean temperature and is applicable in regions of differing station density and climate. The accuracy of RKK is compared to Inverse Distance Weighting, a common deterministic interpolation method, and to Ordinary Kriging, a common stochastic interpolation method. The first step in RKK is to use regression kriging, in which multiple linear regression accounts for topographical effects on the temperature field and kriging minimizes the regression error, to interpolate climatological means. In the second step daily deviations from the monthly climatology are interpolated using simple kriging. Owing to the large climatological differences across the investigation area the interpolation is performed in homogeneous subregions defined according to the Köppen-Geiger climate classification. Cross validation demonstrates the superiority of RKK over the simpler algorithms in terms of accuracy and preservation of spatial variability. The interpolation performance however strongly varies across Europe, being considerably higher over Central Europe (highest station density) than over Greenland (few stations along the coast line). This illustrates the strong impact of the station density on the accuracy of the interpolation result. Satellites provide comprehensive observations of climate variables such as land surface temperature (LST) and cloud cover (CC). However, LST is associated with high uncertainty (standard error ~ 1-2°C), preventing its direct application in meteorology and climatology. The second study investigates the usefulness of LST and CC as predictors for the gridding of daily Tmin and Tmax. The RKK algorithm is compared with similar interpolation methods that apply LST and CC in addition to the predictors used with the RKK algorithm. The investigation is conducted in two regions, Central Europe and the Iberian Peninsula, which differ strongly in average cloud cover (Central Europe is approximately 30% cloud free and the Iberian Peninsula approximately 60 % cloud free). RKKLST (in which monthly mean LST is used as an additional predictor) yields for Central Europe no clear improvement over RKK, yet it reduces the interpolation error over the Iberian Peninsula. This finding can be explained by the higher percentage of cloud free pixels over that region in summer which enables a more robust determination of monthly mean LST. Adding a regression step for daily anomalies (using the predictor CC) yields the RKRK method and improves the preservation of spatial variability over the Iberian Peninsula. Moreover, a successive reduction of the station number (from 140 to 10 stations) reveals an increasing superiority of RKKLST and RKRK over RKK in both regions. The application of a gridded observational dataset for climate monitoring or climate model validation requires knowledge of the uncertainties associated with the dataset. The estimation of the interpolation uncertainty, here the inter quartile range is the used uncertainty measure, is therefore an important issue within the frame of this thesis. By means of cross validation it is shown that the largest uncertainties occur in regions of low station density (e.g. Greenland), in mountainous regions and along coastlines (in these regions model evaluation results should be interpreted carefully). The magnitude of the interpolation error mainly depends on the station density, while the complexity of terrain has substantially less influence. On average over all regions and investigation days the target precision of the uncertainty estimate is reached. However, on local scales and for single days it can be clearly over- or underestimated. The application of satellite-derived predictors (LST and CC) yields no noteworthy improvement of the uncertainty estimate. In the last study two regional climate simulations for Africa using the ERA-Interim driven COSMO-CLM (CCLM) model at two different horizontal resolutions (0.22° and 0.44°) are validated. It is assessed whether observed patterns and statistical properties of daily Tmin and Tmax are correctly represented in the model. The ERA-Interim reanalysis and a specially created observational dataset are used as reference. The observational dataset is generated by applying the RKRK algorithm (developed within the second study). The investigations show an occasionally large bias in Tmin and Tmax. The hemispheric summers are generally too warm and the temporal variability in temperature is too high, particularly over extra tropical Africa. The diurnal temperature range is overestimated by about 2°C in the northern subtropics but underestimated by about 2°C over large parts of the African tropics. CCLM reproduces the observed frequency distribution of daily Tmin and Tmax in all African climate regions, and the extreme values in the lower percentiles (5, 10, 20%) for Tmin are well simulated. The higher percentiles (80, 90, 95%) for Tmax are however overestimated by 2-5°C. For both Tmin and Tmax the 0.22° simulation is on average 0.5°C warmer than the 0.44° simulation. Additionally, the higher percentiles are about 1°C warmer for both Tmin and Tmax in the higher resolution run, while the lower percentiles in both runs match very well. Although the temperature pattern is represented in more detail along the coastlines and in topographically complex regions, the higher resolution simulation yields no qualitative improvement. To summarize, the choice of the appropriate algorithm mainly depends on the interpolation conditions. In cases where the station density is high across the target region and the predictor space is adequately covered by observing stations, the computationally less demanding RK algorithm should be preferred. In regions where the station density is low the more robust RKRK algorithm should be the first choice. Due to the strong physical relation of both CC and LST to Tmin and Tmax the missing information is at least partially compensated for. The estimation of the interpolation uncertainty could be improved by applying a normal score transformation to the data prior to a kriging step. This is because the kriging assumption that the increments of the variable of interest are second order stationary can be approximately met by a normal score transformation.