During an earthquake, the seismic movement recorded at a given site may be strongly dominated by the response of the subsurface geology. Indeed, certain geological configurations (e.g. soft layer on a hard bedrock, topography, sedimentary basin) are conducive to the trapping of seismic waves leading to an amplification of the seismic motion at certain frequencies and an extension of the duration of the signal, we speak in this case of site effects. Moreover, the small-scale heterogeneities (a few meters to tens of meters) of the geological layers can redistribute the energy of the trapped waves, thus increasing the variability of the seismic motion recorded at the surface. The effects of these heterogeneities are very often neglected when calculating the response of a site from numerical simulations because the limit of resolution of classical geophysical imaging methods does not allow to resolve them well. To overcome the resolution limits of knowledge of the environment, it is common to use stochastic methods in which the spatial variation of the heterogeneities is assumed to be random. Thus they can be characterized by a spatial autocorrelation function (ACF). The main objective of this thesis is to study the effect of velocity heterogeneities in a sedimentary basin on the seismic motion recorded at the surface. For this purpose, 2D numerical simulations of seismic wave propagation are carried out. Two basins are considered: the Nice basin, characterized by previous studies and instrumented with several stations of the permanent accelerometric network (RESIF-RAP), and a canonical basin. The heterogeneities of the basins are modeled as a random field characterized by a Von Karman ACF. In a first step, numerical simulations are performed in the Nice basin by considering a simple source function and by varying the parameters of the ACF (correlation length and coefficient of variation) from values taken from the literature. The results of this sensitivity study show that the coefficient of variation controls the variability of surface motion (maximum value of velocity, transfer function, response spectrum) at first order, ahead of the correlation length. In a second step, we use available drilling data from the Nice sedimentary basin to constrain the ACF parameters, in particular the coefficient of variation and the coefficient of variation in the vertical direction. Different correlation lengths are tested in the horizontal direction. New numerical simulations are thus carried out, this time using as source function a seismic signal recorded at the Nice rock. The variability of the motion in this case is studied for a more complex incident wave field. A comparison between the transfer function (FT) calculated from the results of our numerical simulations and the empirical FT calculated at an NLIB station located on the basin profile is also performed. These results show that the response of the Nice basin is controlled at first order by the reference velocity model, and heterogeneities govern the variability of this response. Finally, in a third step, numerical simulations are performed in a wider and deeper canonical model, in order to evaluate the influence of the correlation length and the Hurst coefficient for waves propagating longer in the basin. They show that the longer the correlation length, the greater the variability of the basin response. These results highlight the importance of environmental heterogeneities in understanding the variability of seismic motion in sedimentary basins.