Researchers Develop High-Resolution Rainfall Models to Predict 100-Year Flood Risks Across Japan’s Rural Gaps

Osaka Metropolitan University researchers use the INLA-SPDE method to predict extreme rainfall and 100-year flood risks in Japan's rural and unobserved regions.

By: AXL Media

Published: Feb 27, 2026, 3:59 AM EST

Source: The information in this article was sourced from Osaka Metropolitan University

Addressing the Statistical Blind Spots in Disaster Planning

Japan’s complex topography and diverse climate zones make it uniquely vulnerable to catastrophic flooding and heavy rainfall. While urban centers are well-monitored, large swathes of rural Japan remain statistical "gaps" where comprehensive weather stations are sparse. This lack of localized data presents a major hurdle for future-proofing infrastructure against the intensifying precipitation events driven by global warming. To resolve this, researchers from Osaka Metropolitan University (OMU) and Yantai University have developed a robust spatial modeling framework capable of predicting extreme rainfall even in unobserved regions.

The Shift from Traditional Kriging to INLA-SPDE

The research team, led by Associate Professor Jihui Yuan, sought a more reliable alternative to traditional spatial prediction methods. In environmental research, the "kriging" method often underestimates extreme values, while Bayesian hierarchical models using the Markov Chain Monte Carlo (MCMC) method are computationally burdensome. The team instead utilized the Integrated Nested Laplace Approximation - Stochastic Partial Differential Equation (INLA-SPDE) method. This approach offers a more efficient and stable alternative for spatiotemporal analysis in complex topographies, providing higher prediction accuracy for long return periods.

Analyzing Forty Years of Precipitation Data

The study analyzed hourly precipitation data from 752 meteorological stations across Japan’s four main islands, spanning from 1981 to 2020. The researchers divided the archipelago into four distinct climate zones and estimated the Generalized Extreme Value (GEV) distribution at each station. By calculating return levels for events ranging from 2-year to 100-year intervals, the team applied the INLA-SPDE method alongside traditional kriging methods. They utilized covariates such as annual precipitation, distance from the coast, and population density to forecast risks in areas currently lacking sensors.