Climate Model Downscaling in Central Asia: A Dynamical and a Neural Network Approach

Abstract

Abstract. To estimate future climate change impacts, usually high-resolution climate projections are necessary. Statistical and dynamical downscaling or a hybrid of both methods are mostly used to produce input datasets for impact modelers. In this study, we use the regional climate model (RCM) COSMO-CLM (CCLM) version 6.0 to identify the added value of dynamically downscaling a general circulation model (GCM) from the sixth phase of the Coupled Model Inter-comparison Project (CMIP6) and its climate change projections' signal over Central Asia (CA). We use the MPI-ESM1-2-HR (at 1° spatial resolution) to drive the CCLM (at 0.22° horizontal resolution) for the historical period of 1985–2014 and the projection period of 2019–2100 under three different shared socioeconomic pathways (SSPs): SSP1-2.6, SSP3-7.0 and SSP5-8.5 scenarios. Using the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) gridded observation dataset, we evaluate the CCLM performance over the historical period using a simulation driven by ERAInterim reanalysis. CCLM's added value, compared to its driving GCM, is significant over CA mountainous areas, which are at higher risk of extreme precipitation events. Furthermore, we downscale the CCLM for future climate projections. We present high-resolution maps of heavy precipitation changes based on CCLM and compare them with CMIP6 GCMs ensemble. Our analysis shows a significant increase in heavy precipitation intensity and frequency over CA areas that are already at risk of extreme climatic events in the present day. Finally, applying our single model high-resolution dynamical downscaling, we train a convolutional neural network (CNN) to map the low-resolution GCM simulations to the dynamically downscaled CCLM ones. We show that applied CNN could emulate the GCM-CCLM model chain over large CA areas. However, this specific emulator has shortcomings when applied to a new GCM-CCLM model chain. Our downscaling data and the pre-trained CNN model could be used by scientific communities interested in downscaling CMIP6 models and searching for a trade-off between the dynamical and statistical methods.