Automatic buried mine detection using the maximum likelihood adaptive neural system (MLANS)

A new approach is described for the automatic detection of buried mines using ground penetrating radar (GPR). The main component of the approach is the maximum likelihood adaptive neural system (MLANS), which is a model-based neural network combining the adaptivity of a neural network with the a priori knowledge of signal models. The MLANS technique is designed to adapt to unknown and changing soil conditions, while incorporating signal models based on the physics of electromagnetic scattering to combat the relatively low signal-to-noise ratio (SNR) of typical GPR data. The approach uses a mixture of competing submodels, each designed to model the energy scattered by a specific object or clutter type. The parameters of these submodels are estimated in an iterative fashion by maximizing a log-likelihood function associating the measured data with the submodel mixture. The submodel parameters are subsequently used to compute features in an automatic classifier for the detection of mines. Results are presented in which several types of buried objects are detected from experimental GPR data