Thesis: Numerous studies have demonstrated the impact of surface geological layers on ground motion during an earthquake. The modification of the incident wavefield by these layers is known as "site effects". Certain geological configurations are prone to significant site effects, which can lead to an increase in the amplitude of seismic movement at the surface and a lengthening of its duration. It is therefore important to take this phenomenon into account when assessing the seismic hazard, as stipulated in the Fundamental Safety Rule RFS 2001-01 for nuclear facilities.The IRSN is conducting studies to improve the estimation of site effects using empirical and numerical approaches. These approaches require a large number of earthquake recordings. In areas of low-to-moderate seismicity, as in mainland France, the return period of moderate to strong earthquakes is long, and therefore limiting. It is therefore of great interest to exploit the ambient noise. Ambient noise corresponds to weak vibrations of the Earth's surface. Generated by natural or anthropogenic activities, its characteristic feature is that it is permanent. This makes it a valuable source of data, even in areas of low seismicity.This thesis uses empirical approaches to study the potential of ambient noise for characterizing site effects. It makes use of ambient noise recorded on a 400 nodes array deployed in the Tricastin Valley (French Rhône Valley) as part of the DARE project (Dense ARray for seismic site effect Estimation, 2020-2023). This valley is prone to generate important and complex site effects. The objectives are to: – (1) study the potential of ambient noise and the contribution of a dense array of sensors in estimating site effects – (2) estimate the amplification associated with the Tricastin Valley in the event of an earthquake, based on ambient noise – (3) propose new methods for estimating amplification.The amplification associated with the valley was first estimated through SSRn and SSRh (hybrid method combining ambient noise and earthquake data, Perron et al. 2018). The results were compared with SSR measurements. At low frequency (f < 1 Hz), the estimated amplification is consistent with the expected geology. Estimation of amplification at high frequency (f > 1 Hz) is more critical due to the influence of local anthropogenic sources of ambient noise (heavily industrialized area). SSRh, which uses a reference sensor at the site, nevertheless provides an amplification factor comparable to SSR up to 4 Hz.To limit the influence of local sources, a clustering algorithm was used to select the least impacted data. Estimation at sensors impacted by transient sources was thus improved, but not in areas impacted by permanent sources. These results illustrate the difficulty of implementing a method based solely on ambient noise amplitude in highly industrialized areas. We therefore worked on estimating amplification from deconvolutions calculated between sensor pairs. The idea is that the phase information contained in the deconvolution can help filter the ambient noise wavefield and thus limit the impact of local sources, particularly permanent ones. Deconvolutions were calculated using 7 virtual sources distributed around the area. By processing monochromatic signals and selecting the part of the deconvolutions least affected (temporal selection), we show that it is possible to attenuate the impact of permanent sources up to a few Hertz. The results allow us to discuss the advantages and limitations of the different approaches, and open up prospects for future work to improve the study of site effects from ambient noise in industrialized areas.