Sensitivity of Bayes and Fisher Classifiers in Radar Target Detection

摘要

The detection and classification of stationary ground targets using polarimetric radar sensors is presently an important area of research. This reprint considers the use of feature vector methods for detection and/or classification of targets in severe ground clutter. The radar sensor collects polarimetric signatures of the target and its surrounding clutter environment. From these signatures, training data is obtained and the performance of the Bayes quadratic and Fisher linear classifiers is evaluated under the assumption of jointly Gaussian features. Although the Bayes quadratic classifier is theoretically optimal, this algorithm is shown to be extremely sensitive to such factors as erroneous training data, radar system calibration, etc.... The Fisher linear algorithm is shown to have many desirable properties from the standpoint of radar system design. (Author)