Detection and discrimination of landmines and UXO

We consider the use of electromagnetic sensor measurements for discriminating unexploded ordnance and landmine responses from those due to background noise or clutter. The data stream is potentially very large depending on the time or frequency channels used, the number of receiver elements, and the spatial sampling density. Sensor classification performance will typically improve if the raw input data is reduced to its intrinsic dimensionality. Compression may be achieved in a number of ways but we concentrate on (1) model-based and (2) data-based approaches. In the model-based approach we estimate the model parameters of a physical model that represents the target. Key to this approach is the availability of an accurate and rapid method to compute the electromagnetic scattering from a high contrast and possibly complex-shaped object. We have developed a mean field theory to address this precise need. In the data-based approach we seek the coefficients of an efficient basis set (spanning the dataspace) to represent the original data. We use the so-called best-basis paradigm to discover the optimal wavelet basis. Approach (1) explains the data since it uses a physics-based model to fit the recorded measurements. Approach (2), on the other hand, represents the data in terms of a basis that may or may not have physical significance. The parameters estimated from either approach may then be used to compose a feature vector for input into a conventional classifier that has embedded in it a data bank of parameters for known targets. Each approach has complementary advantages, and it is likely that their joint use is optimal. Finally, an understanding of the discrimination methodology to be used can be coupled into the sensor design process for improved performance