Application of system identification and neural networks to classification of land mines

The ability to discriminate or classify land mines is critical. One methodology includes various resonance-based approaches, whereby the Singularity Expansion Method (SEM) parameters are estimated by some method, often a Prony or Maximum Likelihood approach. The natural resonance approach is very appealing for two reasons: firstly, SEM theory states that the natural resonances are invariant with target aspect angle and polarization. For application to land mines, this is appealing since the mines are probably oriented randomly. Secondly, one need store only a limited number of identification parameters in a library for each object. It is for these reasons that we have chosen to investigate the natural resonance approach, as applicable to an impulse ground penetrating radar (GPR). We have restricted the scope of the paper to a subset of conducting bodies of revolution (BOR). In particular, we seek to decide if the radar return from a buried BOR is either a “Cylinder Class” or a “Cone Class”. From the perspective of land mine classification, the cylinder could represent a mine, and the cone a nonhazardous object. The methods described are applicable to a wideband incident waveform (or stepped frequency waveform) which excites natural resonance modes of the BOR. Exact resonance parameters of canonical BOR have been calculated by S.R. Vechinski (1989), and those results are used in this paper