Performance characterisation of hybrid STAP architecture incorporating elevation interferometry

T.B.Hale, M.A.Temple, M.G.Wicks, J.F.Raquet and M.E.Oxley

Abstract: Radar space-time adaptive processing (STAP) techniques have classically focused on azimuth-Doppler adaptivity while placing minimal emphasis on elevation. Elevation beamform- ing offers significant clutter suppression improvement, allowing further suppression of inter- ference sources having identical Doppler and azimuth. This work incorporates elevation adaptivity through an interferometric approach, greatly improving clutter suppression while providing an often overlooked target height discrimination capability. A mathematical construct encapsulating the multistage processing framework is developed for the proposed technique. This framework allows extension of the traditional factored time-space (FTS) technique into the azimuth- Doppler-elevation hypercube and represents a subclass of more generalised hybrid methods. The proposed concept is validated through results based on simulated airbome radar data. Target detection improvement of the order of 25 dB, when compared to standard two-dimensional FTS processing, is demonstrated for an 8 x 8 nonuniform rectangular array. Elevation pattem data are provided to illustrate achievable null width/depth capabilities. These data also indicate target height discrimination is inherently provided and further developmcnt is warranted.

1 Introduction

Target detection improvement within the azimuth-Doppler plane has been clearly demonstrated using STAP, with limited work addressing the 'benefits of elevation adaptivity. A previous contribution introduced the concept of using elevation processing to suppress range ambiguous interference [ I ] . The formulation of the elevation-based interferometric preprocessor published in [2] extended STAP principles into the azimuth-Doppler- elevation hypercube. This factored approach demonstrated constant false alarm rate (CFAR) test statistic improve- ments on the order o r 15 dB using measured airborne radar data. The gencrel premisc of the proposed hybrid FTS method is to improve azimuth-Doppler processing perfor- mance by first suppressing clutter in clevation.

Using simulated data generated from a physical radar environment model, this work extends previous results by improving interferometric elevation processing and demonstrating capabilities for larger vertical degrees of freedom than available in the multichannel airborne radar measurement (MCARM) radar. Extremely deep and rela-

0 IEE, 2002 IEE Pmmeding.s onlinc no. 20020077 DVI: IO. 1049Iip-rsn:Z0020077 Paper first received 2lst August and in revised form 20th December 2001 T.B. Hale. M.A. Temple and J.F. Raquct are with thc Air Force Institute of Tcchnology, Dcpartinent of Electrical and Computer Engincering (AFITJENG), Wright-Pattersou AFB. OH 45433. USA M.C. Wicks is wifh the Air Force Research Laboratory, Sensors Direcrorate. Radar Signal Processing Branch (AFRLJSNRT), Rome, NY 13440, USA M.E. Onley is with the Air Force Institute of Technology. Department of Mathematics and Statistics (AFITIENC), Wright-Palterson AFB, OH 45433, USA

IEE Aoc.-Rud~~r Sonor Nwig. , Vu/. 14Y. Nu 2, Aprii 1V02

tively wide clutter nulls are achieved while maintaining an elevation mainbeam in the target direction. Detection improvement on the order of 25dB, when compared to standard two-dimensional factored time-space (FTS) processing, is demonstrated for an 8 x 8 rectangular array using a coherent processing interval (CPI) of eight pulses. Furthermore, target height information as a function of scan angle is inherently available with resolution depen- dent on the number of elevation channels.

2 Adaptivity synopsis

The proposed approach follows a hybrid framework. Generally, this framework reaps benefits of dissimilar adaptive processing techniques while avoiding the draw- backs of each. The original idea [3-61 provided substantial improvement within non-homogeneous signal environ- ments. Thc hybrid concept proposed here cssentially reduces to a factored approach (a special case of hybrid techniques) as shown in Fig. I . Array element data are first filtered using interferometric elevation beamforming, Le. a vertical element beamforming approach similar to 3-D synthetic aperture radar methods (elevation angular depen- dence on range cell location). The goal is clutter suppres- sion while avoiding target nulling. This filtering is adaptive on a range cell basis with the clutter null centred at the clutter elevation angle. Doppler filtering follows and the final stage is statistical adaptive processing. The final two filtering stages represent traditional FTS [7], a subclass of STAP. The goal is to show significant performance improvements using (i) a relatively simple STAP technique having marginal stand-alone detection capability and (ii) smaller array/CPI sizes generally thought to provide insufficient suppression. Any STAP technique can follow

77

elevation. Doppler. azimuth

data A

-t --t _f

-+ ----t _t

framework of the defined data structure. The identity matrix role is seen by simply expanding we @I, , , into

the interferometric elevation processing with expected performance improvements over those presented here.

Because the proposed hybrid FTS method attempts to suppress interference over an entire range ring, it is suboptimum. An optimum processor is capable of placing nulls at point(s) in the three-dimensional space, i.e. at a particular azimuth angle, Doppler frequency, and elevation angle. Achieving optimum processing capability is the subject of ongoing research.

The proposed adaptive processing method can he repre- sented by expanding the traditional structure of the incom- ing space-time snapshot x to incorporate array elevation data. This expansion yields a space-time steering vector of up @ b @ aN, where up and ah. represent the elevation (P x 1) and azimuth (N x 1) steering vectors, and b is the Doppler (M x 1) steering vector. This formulation drives the form of x. yiclding a PMN x I vector. Using this structure, the proposed adaptive processing technique is easily encapsulated as

(EBW,)"~ = dx (1)

where w., E , and E are defined in the following subsec- tions. The complexity of this mathematical construct is primarily a result of casting a three-dimensional problem into a one-dimensional, vectorised framework. Although a more compact format may exist, the framework presented here offers strong similarity to existing STAP architectures.

2.7 Elevation adaptivity Each weight vector component is derived sequentially in the factored approach. First, elevation adaptivity is applied in (1) through

E = we@fMAr (2)

a PMN x M N matrix with I,,,,, an MN x M N identity matrix and we a P x 1 elevation weight vector (described below). The Kronecker product with the identity matrix is a direct result of the incoming data format. The goal is elevation heamforming with the set of weights contained in we, yet the operation is a P x 1 weight vector filtering an incoming space-time snapshot x of dimension PMN x 1 ..After eleva- tion beamforming, the data vector should contain only MN elements. Examining the stmcture of the space-time snap- shot shows the mathematical form of this operation to he

EHx = (we @ I M N ) ~ X ( 3 )

The Kronecker product provides a compact, succinct notation easy to read and understand within the overall

78

From (4), it is clear this method (after the Hermitian operation in the weight application) extracts the correct elements of the incoming space-time snapshot for eleva- tion beamforming. For this portion of the processing, the identity matrix has dimension MN since each set of vertical channel/array samples is separated by M N samples.

The range cell under test (range ring) defines an eleva- tion angle from array horesight to the ground. Elevation weights w ~ , are chosen to suppress clutter at this angle while simultaneously focusing a beam towards the target's elevation location. Clutter suppression in elevation reduces interference, allowing improved performance of partially adaptive methods in azimuth and Doppler. Using an operation parallcling Wiener filter theory, we and C, (an artificially generated interference covariance matrix designed to place null(s) at the unambiguous clutter ring) are calculated from the filter look direction angle H and the angle to the clutter ring 0, under consideration

we = c;'a,(e) ( 5 )

where

I J-I

J i=i c e = - z : a P (Hj)ap(Oi)H + u2fP (6)

H i = m,Oc where c ( ~ E 111 (7)

The role of steering vector up(@) is mainbeamltarget preservation. The projection of matrix C;' onto the steer- ing vector results in strong clutter nulls at locations described by H,, given the look direction H does not lie within the range of Hi used in ( 6 ) . For simulations presented in this paper, C, contains J = 3 significant clutter sources with the clutter elevation angle bracketed Le. B j = {0.950,, H,, 1.050,). These angles and the degree of separation between them act as a null width control para- meter.

IFF P m . - R a d a r Sonar Novig.. 1/01. 149. ,\,a. 2. Apnl 2002

One design consideration, or optimisation, is of interest at this point. It may be beneficial to vary the clutter null width based on the range to the clutter source/ring. For shorter ranges, the angular extent of the range ring is considerably more than at longer ranges. In this case, optimisation may be possible by merely choosing an appropriate balance of J and angular separation in 0; (a constant of 0.05 in this development) to fully cover the range ring angular extent. A second point of considerable importance involves range ambiguous data. The elevation weights, as presented here, are designed to suppress inter- ference only within the unambiguous range ring. This choice ignores the potentially significant performance improvements gained by suppressing range ambiguous .~ interference [I]:

The final term of (6), C J ~ I , , , injects noise into the process ensurine C,. is full rank and invertible. Here. C J ~ was set

I -

equal to thc noise power spectral density multiplied by the receivcr bandwidth. Equation ( 5 ) projects c' onto an elevation steering vector, ensuring the target is not nullcd in the clutter suppression process. Sidelobe levels are of little conccrn, since the range gating process effectively constrains harmful clutter to a small angular extent in elevation.

2.2 Doppler filtering Consistent with conventional FTS processing, the second wcight vector component in (I)

B = b t3 I,,r (8)

provides nonadaptivc Doppler filtering to the target bin. A Kronecker product is applied to a DFT matrix column, essentially the temporal steering vector given in [7]. Again, the Kronecker product extracts the appropriate elements of the elevation beamformed space-time snapshot EHx. Since pulses are separated within the data structure by N samples, the identity matrix dimension is N.

2.3 Azimuth adaptivity The final weight vector component in ( I ) provides azimuth adaptivity and is given by

wm = R,$U,\, (9)

whcre RN is thc (N x N) covariance matrix estimate and a," is an (N x 1) azimuth steering vector. This ( N x I) wcight vector w,, is statistically based and derived from a true Wiener filter. In conjunction with the previous Doppler filtering step, this operation completes the FTS portion of the processing. This method is only adaptive within the angular domain since the Doppler filtering is done with the standard tyansfonnation/steeriug vector. Covariance matrix estimate R," is estimated by averaging K = 2 N secondary

covariance estimate must reflect the additional transforma- tion on the incoming data vector and is given by

2.4 Test statistic Results presented use the modified sample matrix inversion (SMI) test [Y] for estimated covariance. The SMI test statistic is identical to the adaptive matched filter (AMF) test [IO] and both exhibit embedded CFAR. Therefore, any improvement directly improves output signal-to- interference-plus-noise ratio (SINR) and detection prob- ability (Pd); The CFAR claim is valid for the SMI test provided KRN is complex Wishart distributed [9], assumed true after elevation beamforming.

3 Results

Radar data were simulated using an extension of a physical cnvironment model [7] and simulation parameters consis- tent with the MCARM array (Table I ) . Originally presented as a two-dimensional physical model of radar returns in the azimuth-Doppler plane [7, 1 I], the model was extensively modified to include returns from a rectan- gular array. This expansion of the physical model permits generation of data (ambiguous and unambiguous) within an azimuth-Doppler-elevation hypercube. In this work, the radar parameters of Table 1 result in ambiguous data consisting of the unambiguous clutter response plus the clutter response from four range ambiguous regions. For validetion, work in [2] was duplicated using simulated data and compared with MCARM results; simulated data results were identical. For comparing the proposed hybrid techni- que with a conventional FTS approach, a simulated target was injected into range cell 80 (36.8 km) with 496 Hz Doppler shift at an azimuth angle of 0" and elevation angle of 45". Reported results correspond to an 8 x 8 planar array (side-looking) using a CPI of eight pulses (resulting in eight Doppler bins).

Figs. 2 and 3 present CFAR test statistic results for FTS using an azimuth array with the proposed hyhrid/factored technique incorporating elevation adaptivity. Received SINR (per element, per pulse) was set to -53.0dB, making the target virtually undetectable by standard FTS processing. As illustrated in the range profile (at the target Doppler) of Fig. 2, the target's relative peak sidelobe level (RPSL) across range improved by 24.5 dB to -27.0 dB and relative average sidelobe level (RASL) across range improved by 23.2 dB to -36.6 dB. Fig. 3 offers a similar comparison of the two output measures across Doppler in

data vectors oriented symmetrically about the target range cell [7]. These secondary data vectors are taken from the Tab,e ,: Radar Simulation variables

Doppler filtered data, not the raw temporal samples. For K parameter Value . ~~~

equal to twice the degrees of freedom,-performance predic- tions are within 3 dB of optimal for this stage of processing Aircraft altitude (Reed's rule) [PI. For the standard FTS algorithm, RN is Transmit frequency given by [7] Pulse repetition frequency (PRF) 1984 H r

Pulse width 50 psec

Ar ( N ) & El (0 channels

9 km

1.24 GHz

1 K-1

K i = i R,v = ~ B ~ ~ , x ; B = BHRB (10) Az (dx) & El (d,) channel spacing 0.1092 mJ0.1407 m

8/8 However, an additional level of factored processing has Pulses per CPI (M) 8 been introduced by the proposed method. Thus, the new

IEE Pm.-Rudoiar Sv,iu,- Nmig., Pf~l. 149. N o 2, April 2002 79

range, um

13.8 18.4 23.0 27.8 32.2 36.9 41.5 46.1 50.7 55.3 59.9

40-

50-

6 0 ~

I 70- 0

L

e 90- 100-

110-

1 2 0 ~

-50 I , , ! , , , , I 30 40 50 60 70 80 90 100 110 120 130

31.18

-28.78 I I I ~26.38 I I -23.96 I

-21.59 E I

19.19 d I I .16.79 !! I - I -14.39 I I -11.99 I I

30+ I

_.

1 -- - - - - - e

- -9.59

range cell

FTS-linear array - proposed method ~ ~ ~ . .

Fig. 2 N o r n d m d CFAR ie.si .siororitiic against range ai die rorger Doppler

Thc proposed method achieves alniosl 25 d B RPSL improvement and 23 dB RASL improvement

the target range cell. Here, the target’s Doppler RPSL and RASL improved by 20 dB and 17 dB, respectively.

An improvement surface illustrates points where the method degrades pcrformance over standard FTS (using a linear array). The surface of Fig. 4 is generated by subtracting the proposed hybrid CFAR test statistic (in dB) from the standard FTS CFAR test statistic (in dB) with both normalised by the target peak response (not necessarily the highest peak in the test statistic surface). The dashed line intersection corresponds to the target location and represents 0 dB improvement by definition since the data are normalised to the target peak. The black regions of the surface represent degradation of the output CFAR test statistic when using the proposed method. White areas represent varying levels of improvement.

As Fig. 4 shows, there is minimal degradation over the entire range-Doppler surface under consideration. Further- more, the small amount of indicated degradation (only five range-Doppler regions) is proven inconsequential upon closer examination, i.e. the black regions merely represent

relative velocity. mileih

-216 -162 -108 -54 0 54 108 0 , P

-600 -600 -400 -200 0 200 400 Dopplerfrequency, Hz

600

Fig. 3 Normalired CFAR resf ararirtic ugoinsr Doppler in ihr iurgei rnnge cell

The proposed method achieves almost 20 dB RPSL improvement and 17 dB RASL improrcmcnt

80

130’ , , , , , . , , 1 , , 1 7.20 -1000-800 -600 -400 -200 0 200 400 600 600

Doppler frequency. Hr Fig. 4 hiorion only in llw block him ahen ari,rg elevoiion heanzJbming

Range-Doppler- inii~ovemenr .sw/bm rhonn perfirmance degro.

areas of deep nulls in the standard FTS approach that are only raised slightly when using elevation interferometry in conjunction with FTS. Thcse areas still represent consider- able nulls and not false target detections. For example, consider the black pixel in the upper right-hand corner (range cell 30 and Doppler bin containing 800Hz). A Doppler cut of Fig. 4 at this range cell, shown in Fig. 5, shows a degradation of approximately 8 dB near SO0 Hr, yet the test statistic is still 27 dB below the target peak.

Fig. 6 compares the proposed method to FTS over a linear array using a detection probability metric. The proposed method exhibits 230 dB Pd improvement, consistent with previous CFAR test statistic improvements. All results presented thus far include range ambiguous interference. As observed by Klemm [I ] , researchers seldom address the range ambiguous interference case. For completeness, range ambiguous clutter was removed from the data and further comparisons made with a two- dimensional fully adaptive matched filter (AMF). Using K = 2MN secondary data samples for the AMF, the unam- biguous data results show the proposed technique offers

Each curve in Figs. 6 and 7 is the result of a Monte Carlo simulation for 100 000 and I0 000 realisations, respec- tively. For each ambiguous and unambiguous case consid-

significant detection improvcment, as evidenced by Fig. 7.

relative Yelocity, milelh

-216 -162 -108 -54 0 54 108 162 I~ I \ I~

5 0 0 4 0 0 -400 -200 0 200 400 600 Doppler frequency, Hz

FTS-linear arrav I i ..... ,) 1 - proposed method I I

ercd, identical realisations were used to determine f ,k and Pd for each algorithm shown. This approach removes any data dependencics within the curves and represents the most consistent approach for detection analysis. For each case? f f i was held constant allowing a side-by-side comparison of the data shown. Although higher than practical, a P," of IO-' was used in Fig. 7 primarily for illustrative purposes and pennits reliable comparative analysis within computational constraints. Results prescnted in Fig. 6 are for Pfa = IO-'.

For clarification, the proposed hybrid technique (FTS with elevation adaptivity) shows improved performance over the AMF operating on a two-dimcnsional data sct, i.e. over a lincar array with constant pulse sampling. The hybrid technique developed i n this paper operates over three dimensions, i.e. a two-dimensional rectangular array with constant pulse sampling. Therefore, the idea that a matched filter provides optimum performance over an additive white Gaussian noise (AWGN) channel has not been violated. A three-dimensional matched filter frame- work has not been investigated at this point in the research.

Unfortunately, the MCARM program only provides data for two vertical channels. Therefore, the only practical validation technique for the proposcd multi-dimensional processing technique is a Montc Carlo analysis of the RPSL and RASL performance measures. The physical modcl used for data generation demonstrated reliable results for comparison with MCARM two-channel vertical data. Table 2 summarises results for the RASL/RPSL Monte Carlo analysis. Thc tabulated statistics imply that

the results presented are statistically significant. The proposed hybrid method consistently offered an average of 22.3 dB RPSL improvement (directly impact P,&) with a standard deviation of 3.7 dB (just over 10%). Likewise, an average of 22.3 dB RASL improveincnt was achieved. Thesc findings illustrate the significant perfomiance gains realised through the introduction of elevation adaptivity.

The data in Table 2 also illustrate a previously mcntioned point regarding FTS. By examining the FTS surface RPSL measurements for the target, one sees an average of only -0.1 dB of separation in the test statistic between the target and next highest peak. Thus, the detec- tion threshold would have to be set relatively close to spurious on-target responses to detect the relatively small target response, likely resulting in unacceptable P,". However, the proposed method offers 22.3 dB average separation, allowing realistic threshold values and accep- table Pd/P,; performance. These concepts are reinforced by inspection of the P,, data presented in Fig. 6.

3.1 Elevation patterns The calculated elevation pattern (45" mainbeam), shown in Fig. 8, illustrates the achievable depth and width of vertical clutter nulls. Given a spherical Earth model and radar parameters of Table I , the ground location angle to the clutter ring corresponding to the target range cell is approximately -14". The Figure clearly shows the deep clutter null suppressing returns from this clutter ring. Note that the indicated suppression level o f nearly l lOdB is

Table 2: Average kt standard deviation) performance measures using Monte Carlo analysis

Range cut (Fig. 2) RPSL RASL RPSL RASL RPSL RASL

Doppler cut (Fig. 31 Surface (Fig. 4)

FTS-linear array

-2 .6z t2 .8 -15.5i 2.3 -9.6zt3.2 -15.7 sk3.3 -0.1 +2.7 -15.1 f 2 . 3

Proposed hybrid method

-25.8+2.6 - 3 8 . 1 i 1 . 6 -30.9+3.0 -37 .5 i2 .7 -22.4+2.5 -37.5zt 1.5

Improveinem

23.2 + 4.0 2 2 . 6 i 2 . 8 21.3 i 4.4 21.8 zt 4.2 22.29zt3.7 22 .312 .8

2 m[ m 4 -70

E g -80 -90

-100

-110 -80 -60 -40 -20 0 20 40 60 80

elevation angle, 8, deg

Fig. 8 Simabred rievatior~ p a r i m .for Ihr x x 8 mu). wiik mninheani fonned 10 45"

elevation exten[ or the m y The pattrm was cillciilated by moving a constant amplitude target through the

primarily attributable to computational capability and is not likely to he realisable in practical systems. This figure also provides a simulated and calculated elevation pattern comparison. As indicated, grating lobes encroach upon thc angular extent of the array; this result is merely a conse- quence of under sampling along the vertical axis of the array. Because of the range gating process, the only source of unambiguous interfering clutter occurs in the clutter ring located at -14". This condition strictly applies to the simulated data case since the Earth surface is not truly spherical in real world applications due to terrain varia- tions.

Fig. 8 validates thc pattern calculation methodology and shows the algorithm is performing as predicted. Similar plots can he generated for Doppler and azimuth dimen- sions. However, these are not of primary concem to this development. For example, the Doppler filtering operation is a simple nonadaptive filter. As such, its pattern exhibits the typical -13 dB sidelohe pattern commonly found in textbooks. (The only distinguishing characteristic i s i t s shifting to the target Doppler bin.) Similarly, the factored approach taken in this development allows the azimuth pattern to he generated independently of the other two dimensions. The focus of this development was on the proposed FTS hybrid approach and achieving enhanced clutter suppression through elevation processing; the idio- syncracies of basic FTS processing were not addressed.

4 Conclusions

Limitations of classic STAP techniques include large computational load and large secondary data sets. Commu- nity response has focused on developing computationally efficient reduced dimension algorithms, from which FTS evolved. Although FTS provides the desired reduction in degrees of freedom, a performance penalty is incurred. The interferometric hybrid FTS method proposed here, operat- ing within the azimuth-Doppler-elevation hypercube,

effectively mitigates the performance penalty by adding elevation processing without increasing secondary data support. Improved performance over a two-dimensional fully adaptive algorithm operating within the azimuth- Doppler plane is dcmonstrated for the range unambiguous clutter case; secondary data support remains the same as FTS. Furthermore, target height discrimination is inher- ently provided through elevation beaniforming. However, extension into the elevation domain dictates the establish- ment of a new upper performance bound, i.e. the perfor- mance provided by a three-dimensional fully adaptive processor.

As included in the proposed method, elevation adaptivity fosters other concepts for providing potentially significant pcrformancc improvements. Heterogeneous clutter has long been the nemesis of many STAP techniques, given the inherent inability to obtain accurate interference esti- mates. Heterogeneous clutter impact can perhaps he dimin- ished through elevation processing since interference over an entire range ring can be reduced. There is also potential for improved jammer suppression and enhanced perfor- mance in environments containing hot clutter regions, given the localised (angular) nature of such interference and the ability to form elevation nulls.

5 Disclaimer

It should be noted that the views expressed in this article are those of the authors and do not reflect official policy of the United States Air Force, Department of Defense or the US Govemment.

6 References

I KLEMM, R.: 'Space-time adaptive processing principles and applica- tions'. IEE Radar, Sonar. Nnigation and A\,ionics Series 9 (IEE, UK. 1998). ISBN 0852969465 HALE, T.. TEMPLE, M., and WICKS, M.: 'Clunrr suppression using elwation intrrfcrometry fused with space-time adaptive processing', Electron. Lex , 2001, 37, pp. 793-794 ADVE. R.S., HALE. T.B.. and WICKS, M.C.: 'A hvo-stage hybrid space-timc adaptive processing algorithm'. Proceedings of the 1999 IEEE National Radar Conference. April 1999, pp. 279-284 HALE, T., ADVE, R.. and WICKS, M.: 'Two-stage hybrid space-time adaptive processing in radar and communication systems'. United States Patcnt, June 2001. Air Force Invention No. RL10.028. U.S. Patent No. 6,252,540 A W E . R.S., HALE. T.B., and WICKS, M.C.: 'Practical joint domain localiscd adaptive processing in homogeneous and nonhomogeneous environments. Part I : Homogrncous etwironmentr', IEE Pmc., Radar Sonar Nmig., 2000. 147. (2), pp. 57-65 ADVE, R.S.. HALE, T.B., and WICKS, M.C.: 'Practical joint domain localised adaptive processing in homogeneous and nonhomogeneous environmcntr. Part 2: Nonhomogcneous ewironmcnts', IEE Proc., R u d m Somr Novig., 2000. 147. (2) pp. 66-14 WARD, 1.: 'Space-time adaptive proccssing for airbome radar'. Contract F19628-95-C-0002, Lincoln L~boratory, Massachuscrrs lnstihlte of Technology. Lexington. Massachusetts, December I994 REED, I.S., MALLETT, I., and BRENNAN, L.: 'Rapid convergence rate in adaptive arrays'. IEEE Trans. Aerorp. Elrctrori. Sysi., 1974, AES-IO.

2

3

4

5

6

7

8

:FAR detection test for radar',

, D.R., KELLY, E.J., and NITZBERG, R.: 'A CFAR adaktive matched filter detector'. IEEE Emi. Aeros~. Elrc-

4. pp. 198-214

82