Abstract: A terrain perturbation scheme (ter) has been first incorporated into an ensemble prediction system (EPS) and preliminarily tested in the simulation of the extremely heavy rain event occurred on 21 July, 2012 in Beijing. Along with other three perturbation schemes (i.e., initial condition (ic), multi-physics (phy), and one mixed forms (icphy)), similarities and differences in ensemble spread among these schemes have been compared mainly in terms of spatial correlation, total perturbation energy, and scale decomposition. The results show that:(1) the ensemble spread and probabilistic forecasts have been slightly improved by incorporating terrain uncertainty while the ensemble mean of precipitation forecasts remains similar in quality; (2) the spatial structure evolution of the ensemble spread is closely related to the weather system. Quite different initial patterns rapidly evolve into a similar pattern (the correlation coefficient reaches 0.6 or above) in the first 6 hours of model integration for all the schemes. Therefore, little extra information was gained about the spread structure by the mixed icphy scheme compared to either the ic or the phy single-source scheme; (3) differences in the ensemble spread magnitude are, however, large among these schemes, i.e. the magnitude of phy is the largest, that of ter is the smallest, and the magnitude of ic is in between. The mixture of ic and phy perturbations can noticeably boost the spread magnitude for upper level variables but not for near-surface variables, which implies that it is more challenging to increase ensemble spread for near-surface variables including precipitation in an EPS; (4) the scale decomposition of ensemble spread shows that the similarity of ensemble spread structure among the schemes generally increases with increases in spatial scale and forecast hours, but the differences remain large over small scales (<448 km) and for very short forecast range (<12 h); over small scales the differences in the spread magnitude are also obvious among the schemes. Therefore, the selection of perturbation methods in an EPS is more important for small scale and very short-range forecasts than for large scales and long-range forecasts. This study can be used as a guideline to design an effective EPS.