Exploring machine learning approaches for precipitation downscaling

Accurate precipitation has great significance in hydrological, climatological, and meteorological studies. Numerous efforts have been devoted to developing global satellite-derived precipitation products. However, their coarse spatial resolution typically prevented their applicability in regional flood predictions and agricultural management. To achieve reliable and finer-scale precipitation data, many techniques and frameworks have been employed to improve the resolution of the satellite-derived precipitation data. This study critically reviewed existing spatial downscaling approaches, specifically focusing on machine learning (ML)-based algorithms. Insights into the accuracy of these downscaling techniques were provided based on findings from published validation studies. Additionally, the environmental variables utilized in these approaches and the post-processing of residual correction and calibration after downscaling were categorized and analyzed, in which meticulous comparisons of their performance in various study areas were conducted. This study emphasized the importance of generating high-resolution precipitation, systematically evaluated the strengths and limitations of ML-based methods, aiming to identify existing research gaps and potential inconsistencies with previous studies, and ultimately highlighted future research trends and challenges.