Precipitation downscaling based on ensemble learning for climate risk assessment

Extreme precipitation is projected to intensify under climate change, yet global and regional climate model outputs are typically provided at resolutions of several to tens of kilometers, limiting their ability to represent localized precipitation structures and extremes. This study aims to develop an ensemble-learning framework for downscaling coarse precipitation fields to high-resolution fields. The proposed framework ensembles a generative adversarial network (GAN), a convolutional encoder–decoder architecture (U-Net), and a diffusion model to avoid single-model bias and to quantify downscaling uncertainty through ensemble spread. High-resolution gridded precipitation data from the Korea Meteorological Administration (KMA) serve as a reference for ensemble learning. Performance is evaluated through a reconstruction experiment in which high-resolution precipitation fields are artificially coarsened, downscaled, and compared with the original data using root mean squared error, bias, and an extreme-focused metric (the 95th percentile). The trained framework is applied to 25 km regional climate projections under Shared Socioeconomic Pathway (SSP) scenarios, generating 1 km precipitation projections for the Republic of Korea through 2100. Results show improved representation of spatial patterns and extreme statistics relative to individual models, while providing uncertainty bounds for projected extremes. Future work will extend the framework so that the downscaled precipitation data are compatible with geological data (e.g., terrain) at tens-of-meters resolution, enabling analyses of how climate risks influence geohazard risks.