Abstract. Traditional stationarity strategy for extrapolating future design floods requires renovation in response to the possible nonstationarity caused by changing climate. Capable of tackling such problem, the expected-number-of-events (ENE) method is employed with both Annual Maximum (AM) and Peaks over Threshold (POT) sampling schemes expatiated. The existing paradigms of the ENE method are extended focusing on the over-dispersion emerged in POT arrival rate, for which by virtue of the ability to account, the Negative Binomial (NB) distribution is proposed as an alternative since the common assumption of homogeneous Poisson process would likely be invalid under nonstationarity. Flood return levels are estimated and compared under future climate scenarios (embodied by the two covariates of precipitation and air temperature) using the ENE method for both sampling schemes in the Weihe basin, China. To further understand how flood estimation responds to climate change, a global sensitivity analysis is performed. It is found that design floods dependent on nonstationarity are usually but not necessarily more different from those analyzed by stationarity strategy due to the interaction between air temperature and precipitation. In general, a large decrease in flood projection could be induced under nonstationarity if air temperature presents dramatically increasing trend or reduction occurs in precipitation, and vice versa. AM-based flood projections are mostly smaller than POT estimations (unless a low threshold is assumed) and more sensitive to changing climate. The outcome of the biased flood estimates resulting from an unrestricted use of the Poisson assumption suggests a priority to the NB distribution when fitting POT arrival rate with significantly larger variance than the mean. The study supplements the knowledge of future design floods under changing climate and makes an effort to improve guidance of choices in flood inference.
Comparative study of flood projections under the climate scenarios: links with sampling schemes, probability distribution models, and return level concepts