Under global climate change

spatial clustering variations in extreme rainfall time series and their responses to climate drivers are critical to storm risk assessment for a river basin. In this study

we first take the Xiang River basin as the study area

and divide its extreme rainfall series into three clustering regions by applying the partitioning around medoids (PAM) algorithm that is based on the special variogram F-madogram. Then

for each clustering region

the climate drivers are identified by testing their significant correlation with extreme rainfall series from most rainfall stations. Finally

we take the extreme rainfall events of clustering center stations as a representative of the clustering region

and construct a non-stationary extreme rainfall frequency model based on the Bayesian inference. The modeling results reveal the extreme clustering algorithm gives better predictions of the extreme value series than the K-means clustering algorithm. The time-varying models using climate drivers as covariates have the best modeling performance and lowest uncertainties. We demonstrate that the probability of extreme rainfall events in this basin will be increased effectively

especially its rainfall intensity

if three conditions occurred in previous year－the values of North Atlantic Oscillation Index were negative

the sea level pressure in western Pacific Ocean rose

and the sea surface temperature in eastern Pacific Ocean rose. The results help evaluate and forecast the risks of extreme rainfall events in the Xiang River basin.