Eine Methode zur Validierung von Klimamodellen für die Klimawirkungsforschung hinsichtlich der Wiedergabe extremer Ereignisse
U. Böhm (September 2000)
A method is presented to validate climate models with respect to extreme events which are suitable for risk assessment in impact modeling. The algorithm is intended to complement conventional techniques. These procedures mainly compare simulation results with reference data based on single or only a few climatic variables at the same time under the aspect how well a model performs in reproducing the known physical processes of the atmosphere. Such investigations are often based on seasonal or annual mean values. For impact research, however, extreme climatic conditions with shorter typical time scales are generally more interesting. Furthermore, such extreme events are frequently characterized by combinations of individual extremes which require a multivariate approach. The validation method presented here basically consists of a combination of several well-known statistical techniques, completed by a newly developed diagnosis module to quantify model deficiencies. First of all, critical threshold values of key climatic variables for impact research have to be derived serving as criteria to define extreme conditions for a specific activity. Unlike in other techniques, the simulation results to be validated are interpolated to the reference data sampling points in the initial step of this new technique. Besides that fact that the same spatial representation is provided in this way in both data sets for the next diagnostic steps, this procedure also enables to leave the reference basis unchanged for any type of model output and to perform the validation on a real orography. To simultaneously identify the spatial characteristics of a given situation regarding all considered extreme value criteria, a multivariate cluster analysis method for pattern recognition is separately applied to both simulation results and reference data. Afterwards, various distribution-free statistical tests are applied depending on the specific situation to detect statistical significant differences between the most similar clusters from both data sets concerned. Then, a newly developed module is utilized to calculate distance functions between simulation results and reference data using the results of these statistical tests. To assess a model’s performance in a comprehensive way, overall measures are finally computed combining the individual test results.
Preceding sensitivity experiments were carried out to give evidence that it is possible to identify realistic patterns of drought incidence in Northeast Brazil when using the minimum distance cluster analysis technique as implemented in the validation method. As a first test case, the general risk of an agricultural drought and the potential yield loss for the predominating crops in semi-arid Northeast Brazil as derived from the sensitivity of a crop model to precipitation deficits are then investigated for the year 1983, which was one of the driest in this region. For the observations, a high sensitivity of the cluster analysis results as to the number of extreme value criteria used could be detected, even in case of pronounced correlations between them. On the other hand, the minimum distance method turned out to be relatively robust in respect of variations in the number of sampling points. In the main part of the investigations presented here, the new method was employed to validate global and regional climate model simulation results, global analysis data and gridded observations which have been related to observed time series. Compared to the observed time series, remarkable differences regarding the representation of extreme conditions could be detected for all of these data sets. The defined overall quality measures allowed to quantify both, discrepancies between the most similar clusters in test and reference data sets, and structural errors. The method was sensitive enough to provide useful results not only for regional climate model results, but also for gridded observations with relatively small differences to the reference data. On the other hand, deficiencies could even be identified for the coarse-resolution data sets. Altogether, the algorithm could be successfully applied to the tested data sets under the specified conditions. If, however, distinct differences became visible, only a restricted identification of the most similar clusters in both data sets to be compared was possible in some cases. Therefore, the distribution features of the reference data should be used exclusively when relating clusters to each other in future versions of the method. Another transformation algorithm for normalizing data of different scale level should also be tested for applicability to avoid any correlations in advance.
