Deep Learning Downscaling of Precipitation Projection over Central Asia

Central Asia, as a chronically water-stressed region marked by extreme aridity, faces significant environmental challenges from intensifying desertification and deteriorating ecological stability. The region’s vulnerability to shifting precipitation regimes and extreme hydrometeorological events has been magnified under anthropogenic climate forcing. Although global climate models (GCMs) remain essential tools for climate projections, their utility in Central Asia’s complex terrain is constrained by inherent limitations: coarse spatial resolution (~100–250 km) and imperfect parameterization of orographic precipitation mechanisms. This investigation advances precipitation modeling through deep learning-enhanced statistical downscaling, employing convolutional neural networks (CNNs) to generate high-resolution precipitation data at approximately 10 km resolution. Our results show that the deep learning models successfully simulate the high center of precipitation and extreme precipitation near the Tianshan Mountains, exhibiting high spatial applicability. Under intermediate (SSP-245) and high-emission (SSP-585) future scenarios, the increase in extreme precipitation over the next century is significantly more pronounced compared to mean precipitation. By the end of the 21st century, the interannual variability of mean precipitation and extreme precipitation will become even larger under SSP-585, indicating an increased risk of extreme droughts/floods in Central Asia under high greenhouse gas emissions. Our findings provide technical support for climate change impact assessments in the region and highlight the potential of CNN-based downscaling for future climate change studies.