In this paper we present a resonance-based classification technique for the identification of plastic-cased antipersonnel (AP) land mines buried in lossy and dispersive soils under rough surfaces by a stepped-frequency ultra-wideband (UWB) downward-looking ground penetrating radar (GPR) with an array of receivers. For this application the multistatic ground probing sensor is positioned just above the ground surface and operates from UHF to C-Band frequencies. Novel physics-based models based on the finite difference frequency domain (FDFD) technique simulate the characteristic resonating multi-aspect target frequency responses for several realistic buried land mine detection scenarios. Matched filter detection results are presented which assess the GPR's performance in identifying a simulated mine buried under a rough surface at varying depths in dry sand and a dispersive clay loam soil from other false targets such as buried rocks.