Classification of training data with reduced-rank generalized inner product

Abstract

Selection of training data for space-time adaptive processing in radar systems remains one of the critical problems to be solved. The practical application of optimal detection theory relies on a large number of i.i.d. training samples. The required homogeneity is typically assumed to be satisfied by range cells adjacent to the cell under test. This is typically not valid in real-world applications. The generalized inner product has previously been proposed to assist in training data selection. This paper introduces two innovations: (1) the generalized inner product in the data-adaptive reduced-rank subspace of the multistage Wiener filter; and (2) classification of the available data into distinct, self-homogenous sets. Injected targets in recorded data from the MCARM program are used to assess performance. Training with data classified within the multistage Wiener filter subspace, also known as the Krylov subspace, is shown to outperform the conventional technique of selecting adjacent training cells.