Page 520 - Fundamentals of Radar Signal Processing

CFAR for the case of a linear detector is provided in Pace and Taylor (1994).

The SOCA and GOCA CFARs estimate the threshold using only half of the
reference cells. Consequently, they exhibit a higher CFAR loss above than the
conventional CA CFAR. This additional loss is less than 0.3 dB for the GOCA
CFAR over a wide range of parameters (Hansen and Sawyers, 1980). The
additional loss for the SOCA CFAR is greater, especially for small values of N.
It is necessary to use N > 32, approximately, to ensure that the additional loss is

less than 1 dB over a wide range of values (Weiss, 1982). A mitigating
influence for either approach is that the use of a split window may allow a
larger value of N than would normally be used in a conventional cell-averaging
CFAR. The reason is that the window size in CA CFAR is often limited by
concern over nonhomogeneous clutter. Since it is known that only half the
window will actually be used, a larger value of N can be tolerated.
Still another way to combat the target masking problem is censored or

trimmed mean CFAR (Ritcey, 1986). In these techniques, the M reference cells
(M < N) having the highest power are discarded and the interference power is
estimated from the remaining N – M cells. In some versions of trimmed mean
CFAR both the highest and lowest power reference cells are discarded.
Consider an example where M = 2. If an interfering target is present, but is
confined to only one or two cells (or if two interferers are present, each

confined to one cell), the censoring process will completely eliminate their
elevating effect on the estimate of interference power. There will, however, be
a small additional CFAR loss due to the use of only N – M cells instead of N
cells (Ritcey and Hines, 1989). Proper selection of M requires some knowledge
of the maximum number of interferers to be expected, as well as whether they
will be confined to one cell or will be distributed over multiple cells.
Typically, one-quarter to one-half of the reference window cells are discarded

(Nathanson, 1991). In addition, implementation of the technique requires logic
to rank order the reference cell data, sometimes a significant implementation
consideration at the speeds at which real-time CFAR calculations often must be
done.
Many additional variations on the approaches described previously can be

used. For example, censoring can be combined with any of the CA, SOCA, or
GOCA techniques. A more elaborate approach attempts to examine the behavior
of the interference in the lead and lag windows and then choose an appropriate
CFAR algorithm. One version of these ideas computes the mean and the
variance in each of the lead and lag windows. If the variance in a window
exceeds a certain threshold, it is assumed that the data in that window are not
homogeneous Rayleigh interference, most likely due to target contamination. A
series of logical decisions then determines whether to combine the windows for

a CA CFAR using the data from both windows, use CA CFAR using only one
window of data, or use GOCA or SOCA CFAR (Smith and Varshney, 2000).
For example, if the means differ by less than a specified threshold and the