Statistical and Machine Learning Downscaling Methods to Assess Changes to Rainfall Amounts and Frequency in Climate Change Context- CMIP 6

Abstract. General Circulation Models (GCMs) simulations result on grids ranging from 50 km to 600 km, and, therefore, this coarse spatial resolution requires data processing, whereby the application of downscaling techniques has become a standard procedure. The main approaches employed are Statistical DownScaling (SDS) and Dynamic DownScaling (DDS). The former SDS consists of Linear Methods (LM), Stochastic Weather Generators, and Artificial Intelligence DownScaling techniques (IADS). Being computationally less demanding and highly portable, most studies apply LM, and IADS approaches to develop the downscaling. However, it is needed to evaluate whether these approaches allow obtaining representative, in the development of rainfall frequency analysis (RFA), in the estimative of the total precipitation (TP) and the number of rainy days (RD) both water year and multiannual level, as well as identify whether any of these approaches provide better results for the last generation of GCM’s made available for CMIP 6. On this basis and considering only the models with a horizontal resolution of 100 km that participated in the SSP1-2.6 and/or SSP5-8.5 scenarios of CMIP6, the present study aim to evaluate the performance of Delta Method (DM), Quantile Mapping (QM) and Regression Trees (RT) to develop RFA, estimate the TP and RD, based on rainfall series obtained by DownScaling, respect to estimative developed with historical records. The results show that the application of DM, RT and QM does not guarantee a temporal correlation between the TP and RD estimated with DownScaling and historical series, likewise, it is observed that in the estimation of RFA, the application of RT generates better results than QM. Finally, it is evident that not applying any DownScaling technique and applying QM generates similar results.