Assessing the risk of extreme precipitation in Japan through GEV distribution and spatial modeling

Japan, characterized by diverse climatic zones and complex topography, has experienced increasing frequency and severity of extreme precipitation events in recent decades. Rural areas, in particular, face heightened flood risks due to sparse meteorological observation networks and challenges in implementing uniform design standards across complex terrains. This study analyzed hourly precipitation data from 752 meteorological stations across Japan (1981–2020). First, return levels were calculated for each station using the Generalized Extreme Value (GEV) distribution. Subsequently, the INLA-SPDE method, incorporating covariates such as annual precipitation and distance to the coastline, was employed to spatially predict extreme precipitation in unobserved areas. The predictive accuracy and stability were rigorously compared with ordinary kriging interpolation. The INLA-SPDE method outperformed kriging in predictive stability, yielding significantly smaller standard deviations for long return periods. Results reveal that spatial variability intensifies with extreme levels: while precipitation intensity generally decreases from south to north, a notable northward expansion of high-risk zones was observed for 50- and 100-year return periods. Specifically, localized extremes emerged in southern Hokkaido and parts of Tohoku, regions traditionally considered lower risk. These findings highlight the limitations of current hazard maps and provide a robust scientific basis for updating hydrological management strategies under climate change. These findings contribute to improved understanding of regional precipitation extremes and provide a scientific basis for flood risk assessment, hydrological management, and disaster prevention planning in Japan. • Assessed the hazard of extreme hourly rainfall in Japan based on latest data (1981–2020). • Estimated precipitation recurrence intervals at 752 observation stations across Japan. • Used six models for spatial interpolation based on INLA-SPDE and kriging methods. • INLA-SPDE model with annual precipitation covariate had the best predictive performance. • Developed return-level maps for 2- to 100-year return periods across Japan.