A new combined statistical method for bias adjustment and downscaling making use of multi-variate bias adjustment and PCA-driven rescaling

Abstract

One major concern of climate modeling is the projection of future extreme events as they have the most severe impact on society and environment. This is a challenging task for modeling and due to the low occurrence rate of extreme values. Furthermore, the local-scale characteristics of extreme events demand for high resolution model data. In the framework of the EURO-CORDEX initiative, climate model ensemble data on 0.11 degrees grid resolution are produced. In order to provide climate data on a higher resolution in an eff cient and reliable way, a statistical downscaling method has been developed, which combines bias adjustment and downscaling. With this method, an ensemble of climate model data on a target resolution of 5 km has been built and it was established as a reference ensemble for Germany. The ensemble consists of the three scenarios RCP 2.6, 4.5 and 8.5, 44 members in total. The method is comprehensible and of minimum complexity. It involves objective predictor selection and it can be applied for different areas, horizontal resolutions, target variables and predictor data sets, and, thus, providing high fl xibility. While the methodology imposes ref ned structures onto modeled data, it does not affect the models data range and therefore allows for extrapolation beyond observed values. The raw model data show for threshold-based indices a rather large spread and bias, which was tremendously improved in the bias adjustment step. Downscaling is challenging as local terrain features can introduce unpredictable residual variation without localized information from e.g. observations. In particular, temperature-based extreme values were well captured by the downscaling algorithm, as temperature is strongly elevation dependent.