The investigation and monitoring of shallow hazards due to the presence of underground cavities remain a challenge for geophysical approaches. Thus, seismic surface waves have been tested in several recent research projects in order to detect and localize voids as well as to determine their geometries. Among these works, numerous numerical studies have proved the feasibility of Rayleigh waves to detect cavities. However, most imagery processes adapted to R waves are faced with difficulties when applying them to real data. This limitation points to a major problem: the interactions between Rayleigh waves and a cavity are complex, particularly in the case of dispersing and attenuating surrounding media. Here, a combined approach based on numerical and experimental data obtained in a reduced-scale measurement bench is conducted to better understand the seismic wave propagation phenomena involved in the presence of a cavity and define robust observables that can be used in field measurements. The observables bearing the cavity signature are studied qualitatively and quantitatively on numerical and experimental recordings. The latter take into account all the propagation phenomena involved. The observations are carried out on the vertical and horizontal component of the Rayleigh wave displacement. The selected observables are studied depending on nondimensional cavity's parameters versus the frequency, that is the wavelength-to-size ratio and the wavelength-to-depth ratio. The effects of the cavity's parameters on the observables show particularities as a function of these components, such as a higher rate of the amplitude on the horizontal component as well as a perturbation of the direct seismic surface wave amplitude above the cavity, also higher on the horizontal component. This latter feature is particularly visible on the variation of the elliptical particle motion recorded at the surface. It can be linked to the mode conversions that occur in the vicinity of the cavity and which predominate on the horizontal component when the signal is normalized.