T. Li, N. D. Sidiropoulos, G. B. Giannakis
University of Virginia
Charlottesville, VA  22903


Abstract Unlike linear arrays, the UESA circular array under development by the Navy affords 360 degree coverage without requiring mechanical rotation. One of the drawbacks of circular arrays is that near/mid-range clutter statistics are range-dependent, due to elevation dependence. This means that the sample support available for estimating meaningful clutter statistics from adjacent range bins is limited, which can significantly affect the performance of linear STAP algorithms that build on Wiener filtering ideas.  This was a prime motivation behind the ONR CSTAP program.

Our approach builds on the following key idea. Unlike low-rank matrix (two way array) decompositions which are inherently non-unique, low-rank three-way array decomposition, known under the common name PARAllel FACtor (PARAFAC) analysis is inherently unique, under mild conditions. This allows us to model clutter in the neighborhood of a certain range gate of interest using PARAFAC, and blindly extract Doppler, spatial, and range profiles for the clutter patches in the vincinity of the given range gate. This is achieved by joint least squares (LS) PARAFAC fitting of the 3-D radar data in the neighborhood of interest. The overall algorithm consists of: (i) LS beamforming in the look direction (this serves to reduce clutter rank and improve target-to-clutter ratio); (ii) LS PARAFAC fitting and extraction of clutter Doppler and Spatial components; (iii) clutter estimation for the range gate of interest by LS projection of the received range gate data onto the Spatial-Doppler span of the clutter; and (iv) clutter removal and target Doppler estimation.

PARAFAC affords reliable blind target detection down to -40 dB target-to-clutter, using realistic circular array clutter data (courtesy of Dr. Michael Zatman). Compared to PRSTAP, PARAFAC does significantly better in terms of performance. An adaptive implementation of the proposed PARAFAC CSTAP algorithm has also been developed, and it brings complexity down to PRSTAP-like levels.

Presentation in PDF format