Climate change and human activities have led to nonstationarity in hydrological series. To systematically consider nonstationarity in regional frequency analysis (RFA), the features accounting for temporal variability of data series were developed and a nonstationary index flood model considering the trend and jump mutations was proposed in this study. The features extracted by empirical mode decomposition (EMD) were regarded as attributes to identify homogeneous regions. The fuzzy c-means clustering (FCM) and the combination of the self-organizing feature map and Ward’s agglomerative hierarchical clustering (SOM+Ward) were compared. Then the complete nonstationary RFA was applied to the annual maximum daily precipitation (AMDP) of Jiangxi province, China. The results indicate that the regionalization with the attributes reflecting temporal variability of the data series is more detailed. Moreover, the performance of SOM+Ward is better than FCM. The comparison results of precipitation quantiles, which were estimated by stationary and the proposed nonstationary index model, indicate that ignoring nonstationarity in RFA affects the choice of the best-fit distribution and the determination of index flood. In addition, the complete framework of nonstationary RFA developed in this study can provide more proper information when stations with trend and jump mutations exist in the region.