Abstract
As greenhouse gas concentrations continue to rise and the world enters uncharted climate conditions, climate projections at the regional scale are becoming more and more crucial for planning adequate adaptation and mitigation strategies. In particular, in mountainous regions, high spatial resolution of climate information is crucial, as climate variables often exhibit strong variability over short spatial scales. Here, we (1) present a novel downscaling method based on principal components analysis (PCA), (2) evaluate the influence of separating bias adjustment from downscaling, (3) compare several methods for downscaling, including quantile delta mapping and a lapse rate adjustment and (4) assess the impact of multivariate bias adjustment using a snowfall climate index. We test all our objectives on an ensemble of 11 regional climate models driven by reanalysis for the period 1989–2008 in a cross‐validation approach using two folds of 10 years. Our study area is Trentino‐South Tyrol, a mountainous region in the Southeastern European Alps. Here, we exploit an existing high‐resolution (1 km) gridded observational dataset for daily precipitation and temperature minima and maxima. Our results show that the novel PCA‐based downscaling method performs well, especially for precipitation and accurately captures the observed spatial patterns. Separating bias adjustment from downscaling allows for combining different approaches; however, if quantile delta mapping is used for both bias adjustment and downscaling, then the separation increases errors. Among different downscaling methods, quantile delta mapping performs best overall; however, for temperature, simple lapse‐rate adjustments also work reasonably well under average conditions. The multivariate bias adjustment shows minor benefits by accounting for inter‐variable correlations, especially when the climate models did not reproduce observed correlations between temperature and precipitation. Overall, this research provides novel insights for the post‐processing of climate model output in complex terrain, for example, when performing climate change impact models.