A computational framework for high-resolution drought assessment using conditional GAN downscaling of GPM IMERG precipitation over Jeju Island

Study Region Jeju Island, South Korea, is a groundwater-dependent volcanic island with steep orography, coastal exposure, and sparse high-elevation gauges, where precipitation variability strongly influences drought diagnostics. Study Focus This study develops a supervised conditional generative adversarial network (cGAN) framework to downscale GPM IMERG precipitation from 0.1° to 0.01° and evaluate the resulting fine-grid (0.01° gridded) precipitation reconstruction for Standardized Precipitation Index (SPI)-based drought assessment. A co-kriging (CK) precipitation field derived from ground observations and DEM elevation served as both training target and primary reference; therefore, the evaluation quantifies CK-referenced reproducibility rather than independent precipitation accuracy. The cGAN used upsampled IMERG, elevation, elevation-gradient, and precipitation-intensity channels and was applied to the 2004–2023 IMERG archive. SPI-3, SPI-6, SPI-9, and SPI-12 were evaluated against CK-SPI at grid, watershed, and elevation-band scales, with bilinear interpolation as a baseline. New Hydrological Insights for the Region The cGAN reproduced CK-SPI temporal variability, with mean CC decreasing from 0.858 at SPI-3–0.685 at SPI-12. Errors were largest in the high-elevation interior and western coastal watershed, where CK LOOCV indicated greater reference uncertainty. Relative to bilinear interpolation during 2022–2023, the cGAN reduced RMSE, MAE, and absolute bias across all SPI windows and improved CSI for SPI-6 to SPI-12, while bilinear retained higher correlation. Thus, the cGAN added value mainly through CK-oriented error reduction, not stronger temporal co-variability.