Impact of thresholds on nonstationary frequency analyses of peak over threshold extreme rainfall series in Pearl River Basin, China

Nonstationary frequency analysis of peak over threshold (POT) extreme rainfall series is crucial in hydrology. Most previous studies on the nonstationary frequency analyses of POT extreme rainfall series used only a single threshold, ignoring the differences in the statistical characteristics of POT extreme rainfall series extracted with different thresholds. This study investigated the impact of thresholds on the nonstationary frequency analyses of POT extreme rainfall series using a three-step method. First, the non-stationarities in POT extreme rainfall series extracted with different thresholds were assessed using the Iterative Mann-Kendall test and the Mood test. Second, the nonstationary POT extreme rainfall series were modeled using the generalized Pareto distribution (GPD) with the scale parameter to be linked with physical covariates. Third, the uncertainty in the parameters and return level of the optimal model for the POT extreme rainfall series were evaluated using the Markov chain Monte Carlo method. This method was applied to the daily rainfall at 48 stations in the Pearl River Basin (PRB) from 1979 to 2020. By comparing the optimal models of POT extreme rainfall series extracted with different percentile thresholds, we can draw the following conclusions. (1) As the threshold increases from the 90th percentile to the 99.7th percentile, the percentage of nonstationary POT extreme rainfall series gradually decreases from 92% to 13%, with most stations with a significant increasing trend changing to most stations with a significant decreasing trend, especially when the threshold is greater than the 98th percentile. (2) The uncertainty in the scale parameter, shape parameter, and return level increases as the threshold increases, and increases significantly when the threshold is greater than the 98th percentile, especially for the scale and shape parameters. (3) The 98th percentile is suggested as the optimal threshold for the PRB; (4) For the 98th percentile threshold, the total column water vapor in the convective indices is the most significant covariate for the entire PRB.