Learning Strategies for Radar Clutter Classification

Abstract

In this paper, we address the problem of classifying clutter returns into statistically homogeneous subsets. The classification procedures are devised assuming latent variables, which represent the classes to which each range bin belongs, and three different models for the structure of the clutter covariance matrix. Then, the expectation-maximization algorithm is exploited in conjunction with cyclic estimation procedures to come up with suitable estimates of the unknown parameters. Finally, the classification is performed by maximizing the posterior probability that a range bin belongs to a specific class. The performance analysis of the proposed classifiers is conducted over synthetic data as well as real recorded data and highlights that they represent a viable means to cluster clutter returns with respect to their range.