In recent years, record-breaking heavy rains have occurred in Kyushu, Japan, resulting in severe flooding. Considering the effects of climate change, it is necessary to examine the non-stationarity of data when analyzing extreme precipitation values. This study used a non-stationary GEV (Generalized Extreme Value) model to calculate the return level with a 100-year return period for various cities in Japan. The results indicated an increasing trend of severe rainfall events in Kyushu. In addition, the maximum daily precipitation for each season was obtained by dividing each year into 3-month intervals. Increasing trends were observed along the shores of the Sea of Japan in spring, Kyushu in summer, and western Japan in autumn. Subsequently, the parameters were compared between seasons, and the results showed that autumn had the highest probability of extreme rainfall. Major disasters occur when an unexpected value (outlier) is observed, and these outliers may affect the estimation of the distribution. Therefore, an approach for outlier detection in non-stationary extreme value data was proposed, and the influence of these outliers on the analysis was examined. The results confirmed that a single outlier significantly affected the estimation of the shape parameter and the return level. In some cases, an outlier in the distribution of annual maximum daily precipitation is not considered an outlier in the seasonal maximum daily precipitation. This suggests that it may be better to use the distribution parameters of each season when calculating the return level. Despite a vast amount of research on trend analysis, studies on outliers for extreme weather data are scarce. Further research is required to refine the definition of an outlier in an extreme value distribution.
View full abstract