Nonstationary Frequency Analysis of Censored Data: A Case Study of the Floods in the Yangtze River From 1470 to 2017

Censored data (CD) of floods, that is, the combination of systematic data (SD) and historical data, can help improve the robustness of flood frequency analysis, due to its temporal information expansion. However, in nonstationary flood frequency analysis, the approach to utilize the CD has rarely been investigated. In this study, a covariate-based nonstationary flood frequency analysis framework based on various likelihood functions using the generalized extreme value (GEV) distribution was built to utilize the CD, with uncertainty considered. This framework was applied to the study of the annual maximum flood frequency of the Yichang gauging station 44 km downstream of the Three Gorges Dam over the period from 1470 to 2017. A summer precipitation anomaly and a reservoir index were used as covariates to explain the variation of the distribution parameters. The results show that for either the SD or CD, the nonstationary models are preferred to the stationary ones by the deviance information criterion, and these nonstationary models may prove to be practical in engineering application, due to the acceptable uncertainty range in flood quantiles derived from covariates. Compared to the stationary or nonstationary models based on the SD, the corresponding model based on the CD results in a higher posterior mean and a smaller posterior standard deviation for the shape parameter of the GEV distribution. It is concluded that the use of historical information under the nonstationary frequency analysis framework may be remarkable in reducing design flood uncertainty, especially for the very small exceedance probability at the tail. Key Points The nonstationarity and insufficient length of data impact the accuracy of design flood estimation The stationary and nonstationary models considering data type are developed by a covariate method and the likelihood principle Temporal information expansion by censored data corrects the underestimation of probability for rare flood events in the original models