interference contribution can be great [23]; the detection
performance of RR STAP algorithms degrades relatively
rapidly as the number of eigenvectors retained is
decreased below the rank of clutter.
Minimizing the degrees of freedom (DoFs) with little
performance degradation is the goal of RD STAP.
However, none of the beamspace RD STAP algorithms
can provide a method for selecting best auxiliary channels.
Ward in [5] suggests different approaches to select the
angular-Doppler bins. One selection strategy is to choose a
rectangular block of filters that is centered on and includes
the target filter, and this adjacent filter selection strategy is
the approach suggested by Cai and Wang [11]. Ward in [5]
also suggests choosing the interference cancellation filter
to be those filters with maximum interference power,
selecting the set of filters that have maximum cross-
correlation with the target filter, or selecting those filters
whose center frequencies are closest to the expected
location of the clutter ridge. However, the best
performance can’t be assured by all of these approaches
when the number of auxiliary channels is fixed. The
problem at hand is that, if the number of total auxiliary
channels is fixed, which set of channels should be selected
as auxiliary channels to cancel the interference
components in the main channel to maximize the output
SINR? In other words, given the performance degradation
requirement, the aim is to minimize the DoFs of STAP.
In the present paper we develop a beamspace
post-Doppler RD STAP algorithm named BCM (best
channel method) based on evaluating the contributions of
individual 2-D angular-Doppler channels to the output
SINR. The proposed algorithm can provide a method of
selecting the best channels to maximize the output SINR
when the total number of auxiliary channels is fixed. The
simulations demonstrate that only 3∼5channelscan
achieve comparative SINR performance rather than 8
channels of JDL3 ×3. Consequently, the requirements of
secondary data samples which are used to estimate the
local covariance matrix are reduced. Moreover, the
proposed algorithm allows the dimensionality of the STAP
to be reduced much less than the rank of clutter without
significant SINR loss. The proposed method exhibits
faster convergence to the performance bound set by the
fully adaptive STAP than the well-known RR STAP
algorithms, such as PC, CSM, and MSWF.
In summary we propose to select the best channels to
achieve near-optimum performance using a RD matrix
constructed by some steering vectors which correspond to
the angular-Doppler channels in the transform domain. Of
course one can use any basis for this purpose, such as
eigenvectors. However, using the best set of angular-
Doppler channels leads to very good performance, which
is much better than using, for instance, columns of
eigenvectors. Unlike the eigenvector-based approaches,
the choice of the channels depends on the Doppler bin.
Note that eigenvectors form one set of basis vectors, while
the angular-Doppler channels form a different set of basis
vectors. The point of our analysis is to show that it is better
to use the latter basis in choosing the appropriate subspace.
The significance of each angular-Doppler channel to
output SINR is evaluated and taken into consideration for
the auxiliary channels selection in RD STAP. Thus, the
proposed algorithm can achieve the near-optimal output
SINR by selecting the best channels as auxiliary channels
when the total number of auxiliary channels is fixed.
This paper is organized as follows. Section II introduces
the airborne radar signal model, beamspace STAP, and
generalized sidelobe canceler (GSC) form. The proposed
BCM STAP algorithm is derived in Section III. The sim-
ulation and performance analysis of the BCM is presented
in Section IV. Conclusions are provided in Section V.
II. SPACE-TIME ADAPTIVE PROCESSING AND
GENERALIZED SIDELOBE CANCELER FORM
PROCESSOR
A. Airborne Radar Signal Model
The radar consists of N elements and transmits a burst
of M identical pulses at a constant pulse repetition
frequency (PRF) during the coherent processing interval
(CPI). The function of radar is to ascertain whether targets
are present. Thus, radar detection is a binary hypothesis
problem, where hypothesis H
0
corresponds to target
absence and H
1
corresponds to target presence. Under
hypothesis H
0
,theNM × 1 received vector x consists
only of clutter x
c
and white noise n contributions.
H
0
: x = x
c
+n
H
1
: x = αs + x
c
+ n (1)
where α =|α| e
jϑ
is a complex gain whose random phase
ϑ is uniformly distributed between 0 and 2π,ands is the
target vector. For airborne mounted radar that moves with
velocity v,receivedgroundclutterisviewedasa“clutter
ring” with the same radius as the desired target signal. The
width of this ring is the range resolution of the radar, i.e.,
%R = c/2B,wherec is speed of light and B is the array
operating bandwidth. Assume that there are N
C
clutter
patches in the range ring of interest. Assuming that the
p
th
clutter patch at angle ψ
p
in the iso-range ring, the
NM-dimensional space-time steering vector for the p
th
clutter patch v
p
is [4] [5]
v
p
= b
p
⊗ a
p
∈ C
NM×1
(2)
where a
p
is the spatial steering vector, and b
p
is its
corresponding M-dimensional Doppler steering vector.
The total space-time clutter return from a given iso-range
is thus an NM-dimensional random vector x
c
of the form,
x
c
=
N
C
!
p=1
˜γ
p
v
p
∈ C
NM×1
(3)
where ˜γ
p
is a complex scalar that accounts of the amplitude
and phase of the p
th
clutter patch, and N
C
= 360 is the
number of evenly distributed clutter patches in azimuth.
At present we assume that the underlying clutter is
homogeneous and can be modeled as a wide sense
ZHANG ET AL.: METHOD FOR FINDING BEST CHANNELS IN BEAM-SPACE POST-DOPPLER RD STAP 255