Modeling extreme precipitation across Japan using stationary and non-stationary generalized extreme value distributions with different estimation methods

Extreme precipitation has intensified under global warming, and Japan is especially vulnerable due to its complex topography and climate. This study examines annual maximum daily precipitation recorded at 51 synoptic stations across Japan for 1901–2020. Annual maxima were modeled within an extreme-value framework using both GEV and Gumbel distributions. To represent possible temporal change, we considered both time-invariant and time-varying formulations, where the location parameter, the scale parameter, or both were allowed to change over time after stationarity assessment. The main objective of the study is to compare three estimation approaches for these models: maximum likelihood estimation (MLE), ordinary least squares (OLS), and weighted least squares (WLS). A supplementary Monte Carlo experiment under stationary GEV settings further indicates that OLS and WLS generally outperform MLE in terms of overall parameter error when contamination becomes stronger. Based on multiple goodness-of-fit criteria, the full GEV is preferred over the Gumbel model at 45 of 51 stations, and the non-stationary models provided the best fit at 42 stations. For estimation, least-squares approaches are selected more often than MLE at 43 of 51 stations (OLS: 24; WLS: 19; MLE: 8), indicating improved robustness to highly irregular rainfall. Return levels for 10-, 20-, 50-, and 100-year periods reveal a clear southward increase; for example, the 100-year return level is 593.1 mm at Naze versus 171.4 mm at Akita. These findings support region-specific risk assessment and mitigation planning across Japan.

Text 125878.pdf - Published Version Restricted to Repository staff only Download (2MB) Official URL or Download Paper: https://link.springer.com/10.1007/s13762-026-07273...

Actions (login required)

View Item