In this paper, we present an inverse scattering approach based on a neural network technique for the detection of dielectric targets buried in a lossy half-space. In particular, we focus the attention on the analysis of the neural network capability to localize a given cylinder starting from the known values of the electromagnetic scattered field evaluated at a number of points near the half-space interface. The robustness of the developed technique is firstly tested by considering a target and/or an electromagnetic scenario characterized by dielectric parameters different from those used during the training phase and secondly by using noisy input data.