Object detection in clutter with learning maps

The nonparametric classification scheme called learning vector quantization, LVQ, is applied to the detection of an object in clutter with SAR imagery. The LVQ structure self-organizes into a decision mapping via unsupervised training which can require a relatively large data set. A physical model can generate a training set for the LVQ prior to further training with an actual data set. The approach is especially attractive when the physical model is too complex to yield an optimal, or near optimal, decision rule. Here a two-scale electromagnetic scattering model is used derive a SAR image model that includes obscuration, shadowing and range inversion. The LVQ performance achieved is good and comparable to the optimum in cases permitting analysis.