Local scale simulation of air temperature by a two-step hybrid downscaling approach using regional climate modeling and artificial neural networks

The influence of microscale and mesoscale meteorology on the local scale variation of air temperature cannot be correctly simulated by the coarse resolution Global Climate Models. The scope of this work is to develop a hybrid dynamic-statistical downscaling procedure and quantify its predictive ability to estimate air temperature variability at finer spatial scales. The study focuses on the warm period of the year (June – August) and the method is applied to eight sites located in Greece with different topographical characteristics. The two-step methodology initially involves the dynamic downscaling of coarse resolution climate data via the RegCM4 regional climate model and subsequently the statistical downscaling of the modeled outputs by training site-specific artificial neural networks (ANN). The RegCM4 model is employed to enhance the representativity of the dataset, while the ANNs are used as function approximators to model the relationship between a number of atmospheric predictor variables and the observed air temperature time series. An insight of the ANN transfer function is obtained by examining the relative contribution of each input variable. The performance of the methodology is evaluated and the results indicate significant improvement from the inclusion of the ANN models in downscaling air temperature.