The Linear Separability of Multiple-Frequency Radar Returns,With Applications to Target Classification.

摘要

Methods are presented for distinguishing and identifying object classes from a limited number of radar measurements. The radar measurements are the backscattered signal amplitudes at up to twelve harmonically related frequencies in the lower portion of the objects spectrum. The amplitudes of n radar returns of an object are defined as an n-dimensional vector. A class of vectors is defined for each object by a representative set of vectors which cover the various aspects of the objects. The classes of vectors corresponding to different objects are linearly mapped into a beta-space of reduced dimension. The separability of these points in beta-space is demonstrated,indicating a correct object classification. The maximum and average probabilities of misclassification are derived for classes immersed in gaussian noise. Mapping methods are derived which minimize classification errors. The probability of misclassification is shown to increase,but non-uniformly as the dimensionality of the n-space is reduced,indicating that the returns at higher frequencies do not contribute as significantly to the reduction of the probability of error as the lower frequencies. Finally it is shown that the techniques used for improving the noise resistance of pairwise classifiers apply also to multiclass classifiers. (Author)