A large number of railways and highways in the western part of China are located in near-fault mountain-canyon sites. The bridges and tunnels account for a large proportion due to the complex topography, and many important projects are faced with severe seismic risk. The ground motions in the near-fault mountain-canyon sites are very complex. On the one hand, velocity pulses and large vertical amplitudes are typical characteristics of the ground motions in near-fault regions; on the other hand, the topography of mountains and canyons leads to amplification and non-uniformity effects on ground motions. For example, the 1992 Hualien earthquake records show that the peak ground acceleration on the sidewalls of the Feitsui canyon in Taiwan is 2.69 times than that of the canyon bottom. In the 2008 Wenchuan earthquake, the peak ground acceleration in the east-west direction at the top of Xishan Mountain in Zigong is 1.77 times than that at the foot of the mountain. Theoretical and numerical analyses reveal that the physical essence of the amplification and non-uniformity effect is the scattering and local focusing of seismic waves by the topography of mountain-canyon sites.

In addition, the current technologies of geophysical prospecting make it difficult to finely determine the physical parameters of faults, interface slip characteristics, etc. It means that the fault rupture process has uncertainty. Based on the previous studies, the uncertainty of source existed objectively and had a significant impact on the characteristics of near-surface ground motions. In this study, it is an issue of quantifying uncertainty in ground motion parameters at near-fault mountain-canyon sites. Monte Carlo simulations and logic trees are commonly used to quantify the uncertainty in this problem. The main purpose is to construct different seismic scenarios, focusing on comparing the standard deviation of the spatial distribution of ground motions in the actual regional site with the standard deviation in the ground motion prediction model. It is worth pointing out that the Monte Carlo simulation has low efficiency to carry out the multidimensional uncertainty analysis. Besides, the simulation of ground motions in near-fault mountain-canyon sites needs to take into account near-fault and topography effects. Meanwhile, the uncertainty of the seismic source will cause random scattering of the seismic waves in mountain-canyon sites, which will lead to the variability of the ground motion parameters at various surface locations. However, the existing studies have not explored the propagation mechanism in depth.

In this paper, the multiplicative dimensional reduction method （M-DRM） is applied to solve the ground motion variability of complex sites near-fault considering the uncertainty of the source. The uncertainty analysis problem is converted into a finite deterministic analysis to obtain statistical moments of ground motion parameters consistent with Monte Carlo simulation. The deterministic analysis uses the boundary element method to simulate the entire physical process. Based on this method, the mountain-canyon site near a strike-slip fault was discussed as an example. The spatial distribution variability of the peak acceleration （PGA） and peak velocity （PGV） under the coupling of near-fault effect, site effect and source epistemic uncertainty was analyzed, as well as statistical values of spectral acceleration （SA） at some surface points.

The results indicated that the M-DRM is applied to the ground motion variability problem of near-fault mountain-canyon sites caused by seismic source uncertainty, which has higher computational efficiency compared with the conventional Monte Carlo simulation. This method can be used for the stochastic ground motion simulation of the complex sites based on the phylsical model and considering the uncertainty of seismic source. When there is a mountain-canyon topography in the near-fault region, the coupling of the near-fault effect and the local site effect causes a significant amplification of the mean values of the PGAs at the sites, which shows significant spatial variations, especially in the canyon. It can be up to 2.69 times that of the result without the local topography. The mean values of the variability of the PGVs at the different surface points are smaller than those of the PGAs. The structural periods corresponding to the maximum values of the surface ground motion acceleration response spectrum are basically the same under the conditions with and without mountain-canyon topography. The seismic source uncertainty is propagated through the site, which is finally manifested in the spatial distribution variability of PGAs and PGVs. Due to the different energy distributions of ground motion acceleration and velocity, there are differences in the variability of PGAs and PGVs. PGAs have larger coefficients of variation. The variability of PGAs and PGVs is different from that of a single parameter under different rupture scenarios when both the asperity intensity and the rupture velocity uncertainty are taken into account. However, the results of the acceleration response spectrum are more complicated. The variability of the structural response at different locations may be lower or higher than the superposition of single-parameter uncertainty variability, and it is affected by the location of the asperity.