With climate change and intensification of human activities, the flood generation mechanism has been influenced which results in the nonstationarity of flood time series. In this study, in order to use historical flood data for flood frequency analysis, a nonlinear AutoRegressive exogenous (ARX) model was employed to complement missing flood data between the historical and recorded data. The key statistics of the complemented time series agreed well with those of the documented record. Using EWT (the expected waiting time until an exceedance) and ENE (the expected number of events in m year is one), nonstationary flood return periods were estimated by the complemented and recorded annual maximum flood peak series, respectively. The results showed that the return levels under nonstationarity are smaller than those under stationarity due to the construction of small hydraulic structures and many check dams after 1980s. This study implies that it is necessary to refine the flood return levels and the corresponding flood control measures for nonstationary flood series due to environmental change.