A machine learning approach to distribution identification in non-Gaussian clutter

Abstract

We consider a set of non-linear transformations of order statistics incorporated into a machine learning approach to perform distribution identification from data with low sample support with the ultimate goal of determining the appropriate detection threshold. The set of transformations provide a means with which data may be compared to a library of known clutter distributions. Several common non-Gaussian distributions are discussed and incorporated into the initial implementation of the library. This approach allows for the addition of empirically measured clutter distributions, which may not have a known analytic form. The adaptive threshold estimation reduces the probability of false alarm when non-Gaussian clutter is present.