On the effectiveness of kernel-based learning-by-examples techniques for electromagnetic subsurface sensing