We report on a laboratory experiment that investigates the effect of soil surface roughness on the identification of the soil electromagnetic properties from full-wave inversion of ground-penetrating radar (GPR) data in the frequency domain. The GPR system consists of an ultrawide band stepped-frequency continuous-wave radar combined with an off-ground monostatic horn antenna. Radar measurements were performed above a rectangular container filled with a loose sandy soil subject to seven water contents and four random surface roughnesses, including a smooth surface as reference. Compared to previous studies, we have reduced the modeling error of the GPR signal for the smooth surface case thanks to improved antenna transfer functions by solving an overdetermined system of equations based on six model configurations instead of only three. Then, the continuously increasing effect of surface roughness on the radar signal with respect to frequency is clearly observed. In close accordance with Rayleigh's criterion, both the radar signal and the inversely estimated parameters are not significantly affected if the surface protuberances are smaller than one eighth of a wavelength. In addition, when this criterion is not respected, errors are made in the estimated parameters, but the inverse solution remains stable. This demonstrates the promising perspectives for application of GPR for noninvasive water content estimation in agricultural and environmental field applications.