A Combination of Cluster Analysis and Kappa Statistic for the Evaluation of Climate Model Results

Abstract

The authors present a combination of different statistical methods for the validation of climate simulation data with respect to observational data of the same spatial and temporal coverage. It is assumed that simulated data and observed data are both given as time series at locations such as grid cells or station locations. The aim of this approach is to quantify the agreement between the two spatial structures of observed and simulated data. These spatial structures consist of the spatial distributions of clusters (obtained from a cluster analysis) that contain climatologically similar locations. If the spatial distribution of clusters were identical for the observed and the simulated data, the simulation would describe the spatial structure of the observations perfectly. Differences from this ideal situation can be objectively quantified using the ? statistic. If the simulation data have shortcomings, the different ? variants can be used to diagnose where these are located. The method is demonstrated using simulation data from the Statistical Regional Model (STAR) for Germany. The combination of cluster analysis and ? statistic proves to be an excellent tool for quantifying the spatial correctness of climate models that can be extended to multimodel comparisons. It can thereby serve as a standard measure for climate model evaluation.