Performance Analysis of Two-Dimensional Parametric STAP for Airborne Radar using KASSPER Data YURI I. ABRAMOVICH, Fellow, IEEE Defence Science and Technology Organisation MURALIDHAR RANGASWAMY, Fellow, IEEE AFRL Sensors Directorate BEN A. JOHNSON, Senior Member, IEEE Lockheed Martin Australia Electronic Systems Pty Ltd PHILLIP M. CORBELL, Member, IEEE Air Force Electronic Systems Center NICHOLAS K. SPENCER Defence Science and Technology Organisation We analyze the performance of a recently introduced class of two-dimensional (2-D) multivariate parametric models for space-time adaptive processing (STAP) in airborne radars on the DARPA airborne side-looking radar model known as KASSPER Dataset 1. Investigation of the impact of linear uniform antenna array errors on techniques that exploit spatial smoothing is demonstrated using a complementary phenomenological clutter model developed at the AFRL. Signal-to-interference-plus-noise ratio (SINR) degradation with respect to the optimal clairvoyant receiver is studied for different parametric models, antenna errors, and training sample volumes. We also analyze the impact of KASSPER training data inhomogeneity on STAP performance. For an extremely small number of training-data samples, we demonstrate that a properly selected parametric model and an accompanying covariance matrix estimation technique should achieve efficient performance for practical STAP applications. Manuscript received September 24, 2008; revised February 16, 2009; released for publication July 7, 2009. IEEE Log No. T-AES/47/1/940019. Refereeing of this contribution was handled by P. Willett. This work was partially funded under DSTO/LMA R&D Collaborative Agreement 290905. N. Spencer is supported by DSTO. Authors’ addresses: Y. I. Abramovich and N. K. Spencer, 200 Labs, ISR Division, Defence Science and Technology Organisation, PO Box 1500, Edinburgh SA 5111, Australia; M. Rangaswamy, RYHE, AFRL, 80 Scott Drive, Hanscom Air Force Base, MA 01731-2909; B. A. Johnson, Lockheed Martin Australia Electronic Systems Pty Ltd, 82—86 Woomer Ave., Edinburgh SA 5111, Australia, E-mail: (Ben.A.Johnson@ieee.org); P. M. Corbell, AWACS Blk 40/45 Sensors, 636th Electronic Systems Squadron, Hansom Air Force Base, MA 01731-2909. 0018-9251/11/$26.00 c ° 2011 IEEE I. INTRODUCTION AND BACKGROUND Optimal space-time (2-D) processing of airborne radar data collected by an M-sensor antenna array over N pulse repetition intervals (PRIs) is known to be efficient at detecting targets masked by ground clutter and active interference (jamming) [1]. Space-time adaptive processing (STAP) is one practical technique that aims to approach the performance of the optimal (clairvoyant) receiver. STAP uses training data as a substitute for clairvoyant knowledge of the model parameters that are required for optimum receiver design; the training data is used to estimate parameters of the interference and/or ground clutter. Unfortunately, practical STAP implementation is hampered in many cases by a severe shortage of appropriate training samples, amongst other reasons. Indeed, the number of sufficiently homogeneous ground-clutter returns is very limited due to both geometric considerations and terrain variations [1—3]. Hence the way in which the training data is selected and used for parameter estimation is crucial for the success of any practical STAP scheme for mitigating interference/clutter. For the given (small) number of appropriate training range bins, the aim is to find the most efficient STAP design in terms of its performance degradation relative to the clairvoyant receiver. Usually the number of training samples necessary to achieve an efficient filter design increases as the number of model parameters increases, so the “best STAP design” means finding a trade-off between the following two extremes. The first extreme approach does not rely on any (potentially inaccurate) a priori information or assumptions about the clutter covariance matrix, and so the estimated MN -variate Hermitian clutter covariance matrix is formed just from the training data. However, the required number of samples is excessive. For example, if the usual maximum-likelihood (ML) (“sample” or “direct-data”) covariance matrix estimate is used, then the famous Reed-Mallet-Brennan (RMB) result [4] is that the number ¿ of independent identically distributed (IID) Gaussian training samples must exceed approximately 2MN in order to achieve approximately 3 dB average SINR losses relative to the clairvoyant optimal filter. In most practical applications, this large number 2MN of sufficiently homogeneous training samples simply does not exist. At the other extreme, there is a holistic knowledge-based (deterministic) approach that relies on complete and accurate information about the radar front-end and the backscattering surface (e.g., from digital terrain maps). This approach assumes that antenna arrays can be precalibrated with sufficient accuracy to permit efficient clutter mitigation. This 118 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 47, NO. 1 JANUARY 2011