Applicability of machine learning-based downscaling method to climate change prediction

The precipitation characteristics that cause water-related disasters strongly depend on local factors such as topography. Therefore, high-resolution climate change projection data is needed to accurately assess regional flood disaster risk. Climate models generally have low resolution and are insufficient to reproduce observed precipitation distributions. Downscaling techniques are usually applied to estimate detailed precipitation distributions. In recent years, machine learning techniques have been widely adopted for downscaling to improve accuracy. However, data-driven machine learning methods have been criticized for issues such as an inability to make appropriate extrapolations when predicting climate change, and there are currently very few examples of their application in this context. In this study, a machine learning–based downscaling bias correction method that recognizes hourly weather patterns in past and future climates was applied and its validity was examined. This method enables temporal and spatial downscaling and bias correction of multiple variables related to hydrological processes, while adequately reproducing climate change characteristics in climate models that are difficult to achieve using conventional methods. Although each variable was estimated independently, the temporal changes were highly correlated with reanalysis values, indicating that the variables were interrelated. Therefore, this simple method of recognizing temporal and spatial distribution patterns can also be applied to hydrologically relevant climate model output variables, allowing downscaling and bias correction while accurately reflecting the climate change characteristics predicted by global climate models.