CLASSIFICATION USING NULTIFREGUENCY V/2 … · ad-r162 026 ground vehicle classification using nultifreguency v/2 nultipolarization resonance radar(u) ohio strte uuiy columbus electroscience

AD-R162 026 GROUND VEHICLE CLASSIFICATION USING NULTIFREGUENCY V/2NULTIPOLARIZATION RESONANCE RADAR(U) OHIO STRTE UUIYCOLUMBUS ELECTROSCIENCE LAB N CHAMBERLAIN JUL 85

UNCLASSIFIED ESL-714199-1S NBS14-82-K-0037 F/B 1/9 M

1-611- 11111 -6

NATIONALMJEU OF STANDARDm~wOOOy fEsoLuTION TEST 01WT8

71C A

The Ohio State University

GROUND VEHICLE CLASSIFICATIONUSING MULTIFREQUENCY MULTIPOLARIZATION

RESONANCE RADAR DT1..__

DTICi Nell Chamberlain JANO8 1988

_. . - .o

The Ohio State University :"'-::":"

ElectroScience LaboratoryDepartment of Electrical Engineering

Columbus, Ohio 43212

Progress Report 714190-10

Contract Number N00014-82-K-0037

July 1985

C-, _ _

DISTRIBUTON STATEMMN ALj Approved for public jolonol_- *j' Distribution Unlimited

Department of the Navy

Office of Naval Research800 North Quincy Street

Arlington, VA 22217

11 19-85 201/.. . .Il... . . . . .. . . ...

-7 r Vl- -- -1 V...'-U k.V~

NOTICES

When Government drawings, specifications, or other data areused for any purpose other than in connection with a definitelyrelated Government procurement operation, the United States

whatsoever, and the fact that the Government may have formulated,

* furnished, or in any way supplied the said drawings, specifications,or other data, is not to be regarded by implication or otherwise asin any manner licensing the holder or any other person or corporation,

or conveying any rights or permission to manufacture, use, or sell .

any atened iventon hat ay i anywaybe rlate theeto

S0272-101Lo o

REPORT DOCUMENTATION W gPoM? No. i l J. '-AU" do IPACE 749-10

Ground Vehicle Classification Using Multifrequency

Multipolarization Resonance Radar a

7 *~ ., .Auermuws Organixto Rept. No.

N. Chamberlain 714190-10. s.iorming Organit.i. Name and Address 14. P .ec. -T-k/Wo Unit No

ElectroScience LaboratoryThe Ohio State University 13. c..ec or ..nffG No

1320 Kinnear Road (C) NOOO14-82-K-0037- Columbus, OH 43212

12. Sonsoring organization Name and Address M2 Type ot RePowt & Period Covered

Department of the Navy ProgressVii Office of Naval Research

800 North Quincy Street 1I

Arlington, VA 22217 ___

AL Suptlfmentary Notes I .

W Aetrict (Limit: 20 warls)

Experiments investigating the classification of ground vehicles using processed radar

returns are described. The calibrated and scaled backscatter measurements of scale-model

vehicles at several azimuth angles are used to establish a catalog representing the VHF

resonance region radar returns of actual vehicle targets. The performance of both the N

nearest neighbor algorithm, (using frequency domain data), and a correlation algorithm .(using time domain data), is investigated. The effects of wave polarization, azimuthangle, and other key parameters are examined. The consequences of introducing forced

errors into the estimates of aspect angle are studied. A novel feature set employing the

ratio of vertically and horizontally polarized radar returns is described, and its

N' classification performance is examined. '

In general, classification is found to be very much dependent on the particularalgorithm, azimuth angle and polarization of interest. The Nearest Neighbor alborithm,

k" using Radar Cross Section Amplitudes as features, was found to perform quite well,

yielding classification rates of about 90%, depending on target azimuthal angle and wave

polarization. Increasing the number of classification frequncles improves performance, V,'

but only to a limit. Errors in aspect angle are found to significantly degradeclassification performance../

17. Document Aalysis a. Doeeripter

. a

b. Idenlfler,/OpefI nded Terms

. ..COSAT:.i-.:.IL Aellb~liy eemenbfoen' approvd 29. Setmnty Class inThis Report) 121. No of Pages

for ." -d _ it"

See Ietruetion en Revere OPTIONAL FORM 272 (4-771(L Fi ormerly NTIS-3S)

Department of Commerce

, -*. .- *

ta.

TABLE OF CONTENTS

rs PAGE

LIST OF TABLES iv

LIST OF FIGURES v P

CHAPTER

I INTRODUCTION 1'',

A. VHF RADAR 3

II DATABASE AND CLASSIFICATION ALGORITHMS 7L.

A. INTRODUCTION 7 .

B. CALIBRATING THE DATA 9

C. CHECKING CALIBRATED RESPONSES 10

D. DATA SCALING 12

E. TARGET CLASSIFICATION 14

F. NEAREST NEIGHBOR CLASSIFICATION 15

G. CORRELATION CLASSIFICATION 18

II EXPERIMENTAL CONSIDERATIONS 20

A. INTRODUCTION 20

B. EXPERIMENTAL FREQUENCIES 23

C. NOISE MODEL 23

D. ESTIMATION OF THE PROBABILITY OF MISCLASSIFICATION 26

E. POST PROCESSED SIGNAL TO NOISE RATIO 27

iii "..'---.-

.* *%~~ %d. ~ -. - ~ . *%***-..*.,-..*...:*.-:

IV EXPERIMENTS 29

A. INTRODUCTION 29

B. FREQUENCY BAND 36

1. INTRODUCTION 36

2. RESULTS AND CONCLUSIONS 36

C. NUMBER OF FREQUENCIES AND ALGORITHM 39

1 . INTRODUCTION 392RESULTS AND CONCLUSIONS3

0 PAECTON4

1.INTRODUCTION 42RESULTS AND CONCLUSIONS4

ERRI ASPECT ANGLE4

1.INTRODUCTION 4

2.RESULTS AND CONCLUSIONS 4

V CONCLUSIONS a.51

APPENDICES 54

A CLASSIFICATION RESULTS 54

B AMPLITUDE AND PHASE RETURNS 135

iv

LIST OF TABLES

Table Page

3.1 Range of Frequencies Used in Experiments 24

4.1 Definitions 30

4.2 Interpretation of Headers in Classification Results 34

Acceslon For

NTIS CRAMI~DTIC TAB 9Uriannourcecj 7

B...................Di.,A. ib Aio I~-iAvailabi;ity Codes

Dist vi rdo

Sp..;c4a'

v .

LIST OF FIGURES1

Figure Page

1.1 Block diagram of target classification system. Thecatalog contains returns of some preselected targets (Ae).The output of the signal processor is a set of amplitude andphase returns of the unknown target. (from [2]) 2

2.1 Silhouettes of the 5 ground vehicles used in classificationstudies. 8

2.2 A tank scaled 1:1 with its impulse response in order to k r~check a calibration. 12

2.3 Basic process of a pattern classification system. 14

3.1 A flow chart of the experimental process of classification. 21

3.2 The distribution of the noise on an 1-0 plane (from [2]). 25

4.1 Classification performance of various sub-bands in theavailable 25-200 MHz band, as a function of polarization. 38

4.2 Classification performance of various sub-bands in theavailable 25-200 MHz band, as a function of aspect zone. 38

4.3 Classification performance of various numbers ofSfrequencies as a flction of polarization. 41

4.4 Classification performance of various numbers offrequencies as a function of aspect zone. 41

4.5 Classification performance of various numbers ofrfrequencies as a function of algorithm. 42

4.6 Classification performance of various polarizations as afunction of aspect zone. 45

*4.7 Classification performance of varlou! polarizations as afunction of algorithm. 45

4.8 Classification performance of various aspect zones as afunction of polarization. 47

4.9 Classification performance of various aspect zones as afunction of algorithm. 47

Fi gure Page

A.11 Misclassification percentage versus post-processing SNR,

comparing the performance of 4 values of NF. -66

A.12 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF. 67



fA.15 Misclassification percentage versus post-processing SV ,comparing the performance of 4 values of NP. 70




*A.19 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NP. 74

A.20 Misclassification percentage versus post-processing SNR,Vcomparing the performance of 4 values of NF. 75

A.21 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NP. 76

A.22 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NP. 77 :



A.25 Misclassification percentage versus post-processing SNR,-comparing the performance of 4 values of NF. 80


Viii *-

W.4

Fi gure Page

4.10 Classification performance of various aspect zones with41 :forced errors in aspect angle, as a function of

polarization. 50

4.11 Classification performance of various aspect zones withforced errors in aspect angle, as a function ofalgorithm. 50

A.1 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200MHz band. 56



A.4 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200 .

MHz band. 59::



A.7 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200MHz band. 62 '" "-'




vii

'.-

Figure Page




A.30 Misclassification percentage versus post-processing SNR, - -comparing the performance of 4 values of NF. 85 - --


A.32 Misclassification percentage versus post-processing SNR, ..... comparing the performance of 4 values of NF. 87


A.34 Misclassification percentage versus post-p.ocessing SNR,r-, comparing the performance of 4 values of NF. 89



A.37 Misclassification percentage versus post-processing SNR.comparing the performance of 4 values of NF. 92







ix

I ;- ° " -' -

0 . ..

Fi gure Page



A.46 Misclassification percentage versus post-processing SNR, ""comparing the performance of various polarizations. 101

A.47 Misclassification percentage versus post-processing SNR,comparing the performance of various polarizations. 102



A.50 Misclassification percentage versus post-processing SNR, .comparing the performance of various polarizations. 105







A.57 Misclassification percentage versus post-processing SNR, ,-

comparing the performance of various polarizations. 112 ..

A.58 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones. 113


A.60 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones. 115 r

x

X. . . . . . . .- . -

Figure Page

A. 61 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones. 116





A.66 Misclassification percentage versus post-processing SNR,*comparing the performance of various aspect zones. 121

A.67 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones. 122-



A.70 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle. 125

A.71 Misclassification percentage versus post-processing SNR,Ucomparing the performance of various aspect zones, with

a forced error in aspect angle. 126

A.72 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle. 127b




xi

Fi gure fPae


*A.77 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle. 132



*A.80 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle. 135

A.81 Misclassification percentage versus post-processing SNR, *

comparing the performance of various aspect zones, witha forced error in aspect angle. 136

B.1 RCS magnitude and phase response for Vehicle A, at 00aspect zone using vertical polarization. 138

B.2 RCS magnitude and phase response for Vehicle A, at 450aspect zone using verticatl polarization. 139

*B.3 RCS magnitude and phase response for Vehicle A, at 900aspect zone using vertical polarization. 140

*BA4 RCS magnitude and phase response for Vehicle A, at 00aspect zone using horizontal polarization. 141

B.5 RCS magnitude and phase response for Vehicle A, at 450aspect zone using horizontal polarization. 142

B.6 RCS magnitude and phase response for Vehicle A, at 900aspect zone using horizontal polarization. 143

xi i

1-~4

CHAPTER I

INTRODUCTION

The conventional radar problem has been one of finding the spatial

location, or velocity or both of some target. This information can be

extended to include a knowledge of the target's identity, if the

operating frequencies and wave polarizations are properly chosen. More

specifically, if the wavelength of the radar energy is comparable to the

maximum dimension of the target (i.e., in the resonance region), then

certain information, relating to the target's dimensions and shape, will

be imbedded in the radar return [3].

Radars operating in the VHF band (30 to 300 MHz) have wavelengths

ranging from 1 to 100 m; hence tarks, trucks, jeeps, etc., are potential

candidates for resonance region target identification. Figure 1.1 shows

a diagram of a VHF radar system for target classification. Typically,

* the radar platform might be airborne, with the radar operating in the

line-of-sight mode.

The problem addressed here is one of studying the classifiability

of ground vehicles using VHF radar in a representatively error prone

(noisy) environment, using simulated multiple-frequency, multiple-

polarization resonance radar returns. The work presented here is closely

related to that of Technical Report 714190-9, "Surface ship

* . Classification Using Multipolarizatlon, Multifrequency Sky-Wave

Resonance Radar", and consequently, that report will be referenced

frequently., 1

CATALOG

A A

A,8

TARGETCLASSIFIER DCSO

____DECISION

4,5. L-,

Figure 1.1 Block diagram of target classification system. Thecatalog contains returns of some preselected targets (A,B). ..

The output of the signal processor is a set of amplitude andphase returns of the unknown target. ~::.:

2

Later in this Chapter, the general nature of the VHF radar system

will be briefly discussed, along with the measurement of amplitude and

phase returns. Chapter II summiarizes the key points in generating a

database, and outlines the classification algorithms used in the

experiments. Experimental procedure and other considerations are

discussed in Chapter 111, along with a presentation of experimental

results. Chapter IV presents a summiary of the work, emphasizing the

more important findings of Chapter III.

A. VHF RADAR

The purpose of the VHF Radar System is to make a set of

measurements pertaining to the radar cross section of a target, which

can then be processed in order to classify the target.

Typically the radar energy will be pulsed to allow estimation of

range and to reduce clutter levels by range gating. If the targets are

moving, doppler filtering may be used to further reduce clutter levels.

The amplitudes of the radar cross sections, A, are key target

features. They can be calculated directly from Equation (1.2). *

Scattered Power0= Incident Power Density at the Tar-get11

p r (4) R4 L 2 L

2 (1.2)~T GT GR X A

3

where:

PT = Transmitter power

GT = Transmitting antenna gain

GR = Receiving antenna gain

X = Wavelength of propagating energy

GA = Receiver power gain

pr = Power received

R = Range of target

L = One-way propagation loss -pLs = System loss

Parameters such as L the one-way propagation loss, can be,:

estimated with a useful degree of accuracy. Generally speaking, the

atmosphere, as far as direct line-of-sight propagation In the VHF band

is concerned, is a non-dispersive and homogeneous medium. (An exception ' ,

to this is the Troposphere, which can cause scattering of

electromagnetic energy). Consequently, the range delay and hence the

range of a target can be estimated quite accurately, with respect to the

measurement of amplitudes.

Unfortunately, the intrinsic phase of a target's radar cross-

section cannot be estimated as accurately as the amplitude. Skolnik

[4] gives the R.M.S. error in range delay, 6TR, for a simple leading and

trailing edge pulse detector as

6TR 4 BE/No. (1.3)

4

4 "4 "'"""

* Pulse width of Radar Energy (seconds)

E - Energy in a single pulse (Joules)

No - Noise power per unit bandwidth (Watts/Hz)

B = Bandwidth of IF amplifier (Hz)

If it is assumed that the bandwidth of the IF amplifier and filters X

is roughly equal to the bandwidth of a pulse, and that the pulse is

roughly rectangular, then B = I/T.

The R.M.S. range error SR is calculated as

c. STR (1.4)

(meters) (1.5)•r -.-- "

where N =No B (Watts) (1.6)S E T. (Watts)

(1.7)

The term rc is the actual length of the pulse, in meters. Assuming thatthe pulse width is at least 20 wavelengths lcity, SR is given as "'

6R= x (1.8)- -

Such an error might be quite admissible in the estimation of amplitudes,

however, its effect on phase is more serious. The error in phase

resulting from this inaccuracy is given by

le 7 (1.9)

Substituting Equation (1.8) into Equation (1.9) yields

, . . - . . .. .. . . . .*. . . .~

.. - ."

el 201 . (1.10)

A typical signal-to-noise ratio (S/N) might be on the order of 10 dB,

resulting in an R.M.S. error, 8e, of about 6w radians. In view of the

fact that intrinsic phase values will be distributed between 0 and 2W

radians, this error precludes any meaningful measurement of intrinsic

phase. C ...

One method of utilizing phase information which obviates the

problems incurred by range errors is to use the parameter W (1],

defined as Wi = e i - i+I i +1l, where

Wi = 47r(Ri-Ri+ 1) + (tili - ,i+ii+1) (1.11)

and

Ri = range to target at wavelength i

Ri+ = range to target at wavelength Xi+1 "

.= intrinsic phase of target at wavelength

,. 1 = intrinsic phase of target at wavelength Xi+1 . -

8i = measured phase of target at wavelength Xi

= measured phase of target at wavelength Xi+.-

Since the propagating medium is non-dispersive, Ri = Ri+I over a

wide range of frequencies (several MHz), and consequently, (1.11)

reduces to .-

S-'-p.:W= - 1 1 (1.12)

This parameter, or more properly, feature, is used in subsequent - ..

classification algorithms.

6

. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . .

CHAPTER II

DATABASE AND CLASSIFICATION ALGORITHMS

A. INTRODUCTION

In order to build a catalog of reference vehicle responses, a large

amount of experimental data collection and processing using scaled model

ships is necessary.

Firstly, the phase and amplitude returns of each model vehicle are

measured at a set of frequencies, polarizations, aspect angles and

elevation angles of interest. The raw data are calibrated to remove

unwanted background and system response effects, and are converted into

absolute radar cross section magnitude and phase. The calibrated data

are then scaled in magnitude and frequency so that the responses are

representative of real vehicle returns measured in the VHF band. A more - .

comprehensive discussion of these procedures is given in [1], the main

points of which are summnarized below.

Measurement of Data

The measurement techniques employed in this study are virtually the

same as those used for ship targets, as described in Technical Report

714190-9 [l]. However, for ground vehicles, polarization was restricted

to vertical and horizontal, measurements were made using an elevation

angle of 270, and aspect angles of 0* through 900 in 10* increments

Y (including 150 and 450). A large flat circular groundplane was used to

simulate the surface of the Earth.

7

p

Fiur 2. ihute fte rudvhce se ncasfcto

z studies

PK

8~

%..&..-1..'..'..2..

i.-

B. CALIBRATING THE DATA

The purpose of calibrating frequency data is to remove the imbedded .-'

system characteristic. Calibration techniques are discussed by Kimball

[7], and (with particular respect to targets on a ground plane) in

714190-9 [1]. In summary, the calibration procedure is as follows. ,..-.'.

1. Remove the range delay from the measured data.

2. Subtract backgrounds from target and calibration target. ,

3. Remove invalid points from resulting subtractions.

4. Filter (in range) the subtracted files to further reduce ..

background terms.

5. Calibrate the target of interest according to Equation (2.1).

6. Filter (in range) the calibrated data again if necessary.

The calibration equation is given by

- E (T -T)TC s (2.1)

where ic, E, T, BT, S, and B are complex phasors for each frequency

defined as:

iC' the calculated radar cross section (RCS) of the calibrated

target

E, the computed (exact) RCS o: in units of meters and absolute

phase (deg.), from the calibration target.

9

* '*.*.***.*.......'-..:*

T, the signal voltage measured with the target installed.

BT' the signal voltage measured for the background (no target

installed) associated with the target.

S, the signal voltage measured with the calibration target U

installed. .

the signal voltage measured for the background (no target

installed) associated with the calibration target.

The term T in Equation (2.1) is a combination of backscatter from

the vehicle and groundplane, and the term BT is the backscatter from the

groundplane alone. Hence, a background (groundplane) subtraction should

yield the vehicle backscatter. Unfortunately, owing to target- . -

groundplane edge interactions and the unavoidable positional disturbance

of the groundplane when the targets were installed or moved, some

residual response remains at the location of the groundplane edges after

the subtraction and calibration.

Post-calibration time domain windowing was used to remove these

residuals as they represent a distortion of the desired vehicle data,

and would be particularly disruptive in the correlation algorithm

discussed later.

C. CHECKING CALIBRATED RESPONSES

Once a particular vehicle had been calibrated, its frequency

response and impulse response were then generated as plots. The

frequency responses were examined for 'glitches', i.e., large spikes of

'A 10

,- Io :*.* 'A*. *., . :'\ ' *f . .:'*.-. :-S

about 10-100 MHz bandwidth and 10 to 30 dB in ex~ent, caused by receiver

hardware problems. Generally a glitch is hard to deal with because the

phase and amplitude responses affect up to 10 points. If the glitch

exists in a background or calibration target file then an alternative

1P data file might be used. A bad glitch in a ground vehicle data file

might require a new set of measurements. Small glitches, both in

amplitude and bandwidth, at frequencies lower than 4 GHz can be

tolerated.

{:. The time response was the main tool for checking the validity of

calibrations because the transient response gives an intuitive geometric

guide to the mechanisms which cause scattering. Figure 2.2 shows a

typical response, scaled 1 to 1 with the vehicle overlaid. Using

templates in this way shows whether the main response confines itself to

the length of the vehicle and if structures likely to cause large

amounts of scattering are indeed doing so. The bandwidth of the .- ~

responses is sufficiently large to provide the necessary resolution to

make these judgements. Resolution in time is given by

T =1/B (2.2)

For B =16 GHz, then T 62.5 p seconds, which at the speed of light

corresponds to 1.875 cm. The average length of vehicles used in

experiments was about 7 cm (including gun barrels of tanks) and the

average width was 2.5 cm.

K

-"-" '.'.'- T-Z IT2-:

Figure 2.2 A tank scaled 1:1 with its impulse response in order tocheck a calibration.

D. DATA SCALING

Data collected and calibrated in the microwave region on scale

models must be scaled before being used in classification algorithms. A

data point is scaled in 2 ways. First, its amplitude, /, is multiplied-

by a scale factor, and secondly, the frequency it represents is divided -

by the same scale factor. Data measured in the 2 to 18 GHz microwave

band was collected for models having a scale factor of 1:87. The

resulting VHF band extends from 23 to 206 MHz.

12

•. "Since processing frequencies may have been selected which were not -

actually represented by data points, it was necessary to interpolate

between data points by means of a Hamming window.

A frequency increment of 0.5 MHz was selected for the 23 to 206 MHz

band, with a Hamming Window width of 5 MHz. Representative scaled

I.. frequency and phase returns are shown in Appendix B. Since the validity

of data near the band edges is uncertain, these frequencies are avoided

in subsequent classification experiments, resulting in a net useable

frequency range of 25 to 200 MHz.

A SUMMARY OF SCALED DATA GENERATION

1. Measured Data

Amplitudes Am = ''3 CPhases(at) frequencies fm 2-18 GHz, 10 MHz steps 1601

data points

2. Calibrated Data Ac

ec

fc 2-18 GHz, 10 MHz steps

* Smoothing - 3 ns (first nulls) Hanning Windowequivelent to 65 point smoothing,gives a 1:25 window to bandwidthratio.

3. Scaled Data As = Ac SFs = c (SF = Scale Factor)

fs fs/SF

SF = 1:8723-206 MHz, 1/2 MHz steps366 data pointsInterpolation using 5 MHz HammingWindowNet frequency range: 25-200 MHz

13

' ..

E TARGET CLASSIFICATION

The basic process of target classification is illustrated in Figure

2.3. The measurement system in this case is a compact radar range. It

produces a measurement vector m, which is a set of calibrated RCS

amplitudes and phases at a number of frequencies, aspect angles and

polarizations El]. Each measurement vector is a point in M-dimensional

space (also called the observation space). The feature extractor

reduces the dimensionality of the measurement vector to produce a

feature vector n, which is a point in N-dimensional space (M > N). For

example, if the amplitudes and phase of the RCS of a vehicle were

measured at 2 polarizations, 3 aspect angles and 4 frequencies, and it

is assumed polarization and elevation are known, aspect angle is unknown

and only amplitudes are used to classify the target, then the feature

extractor reduces a 48-dimensional space to 12-dimensional space (4

frequencies x 3 aspect angles). Essentially, the feature extractor is

Lj[MEASUREMENT H FEATURE H OBJECT t -- (' ";,-,

SYSTEM EXTRACTOR CLASSIFIER DECISION

MEASUREMENT FEATUREVECTOR VECTOR

Figure 2.3 Basic process of a pattern classification system.

14

.. • ~ ~.: .

the mechanism by which a particular data file is addressed, since a

measurement vector is usually dispersed amongst several data files. The

feature vector is then passed to the object classifier which uses a

particular algorithm to make a decision as to the identity of an unknown

target (in our experiments, unknown targets are simulated by adding

errors to known target RCS amplitudes).

The ultimate goal of a classification system is to identify targets

with a minimum probability of error. Since the errors in these tests

were simulated, the classification could have been a parametric

procedure. However, in practice, the exact probability distributions

of features in the feature space are not known, thus we must resort to

non-parametric methods of classification. Two such methods which do not

require knowledge of probabilistic information are the nearest neighbor

algorithm [2] and the correlation algorithm [2], and are discussed

below.

*F. NEAREST NEIGHBOR CLASSIFICATION

The nearest neighbor (NN) algorithm uses amplitude and/or phase

returns measured at a series of frequencies f = fl, f2, ,•,, fN• (Note

that, in general, the NN algorithm is applicable to both the frequency .

and the time domain). The amplitude is defined as the square root of

the measured target cross section.

As mentioned earlier, the accurate measurement of intrinsic phase

(i.e., phase attributed to the target) is precluded by errors in the

measurement of target range. To circumvent this difficulty, the

15

..- L

. . .. . . .o .

differential quantity W is used and is defined as:

:+1 i i -i+1i+1

-~ where

6 iis the measured phase at X = Xi , and

-':-1 1, ... N - 1 (where N = Number of frequencies).

The reliability of amplitude information, A, and phase information,

1. W, is related by a variance weighting factor 3, in the NN algorithm.

Strictly speaking, a is an arbitrary choice of the experimenter, but in

the absence of any a priori knowledge as to the reliability of the A and

W features (as was the case here) a is set to 1.

Let At (fi), e (fi) be the measured target amplitude ( IRCSI) and L

phase where i = 1, 2, ...N. Let Aj (fi), ej (fi) be the catalog of

amplitude and phase values, where j is the target index, j = 1, 2, -.*me

The NN algorithm is as follows:

1. Compute the differential phases

WIF = W j X j )fi) - X+i ej (fi+i) 04

I 1, 2, ...N-1

j =1, 2, ...M t_

16

,--%

2. Calculate the sample averages of A and W in the database.

1 M N

Avg(A) zj TN I A (f1) =A

1 M- i

Avg(W) 1 M(-l N-1 W

j=1 i=1

V 3. Calculate the sample variances of A and W in the database.

1 M N 2Var (A) T I I (Aj(f 1 ) A)2

j=l I=1 ...

Var (W) M N-i - 2

M(-) X I (W3 -W)j=1 1=1

4. Select 8 the variance weighting factor, and calculate K,

LK5. Compute the distance between the estimated target feature

vector and the catalogued feature vector for each class in the database.

N f) A.( 2 N-1i )dt (At (fi A (f)) + Z (KWt " KWJ) 2 •

j = 1, 2!, , M.

6. Apply the nearest neighbor rule:

Choose smallest dt,j j = 1, 2, ... M.

If dt,m min (dt,j) classify the target (t) as m.

17

.. , v.9.-. . .~ .. *.* * ** %* N

7. Repeat experiment several times for each level of noise power.

Compile statistics as a function of noise power.

G. CORRELATION CLASSIFICATION

In the NN algorithm, phase ambiguities can be reduced by employing !71the differential quantity W. These phase ambiguities manifest

themselves as time shifts in the canonical time domain, and are

reducible by means of a correlation process. The correlation algorithm

[2] may, in fact, be implemented in either the frequency or time domain.

(Below, the latter is used.) W

1. Compute the correlation function. t _,

DIFT [X(m) Y*(m)]

2 M IX(m)I) IY(m)I "

where

X(m) = At(fk)exp((jet(fk)) t = target index

Y(m) = Ar(fk)exp(jor(fk)) r = catalog index

k = 1,2, ... ,

N = Number of frequencies

m = f1/Af, f 2 /Af, .. , f

A= f f-i i = 1, 2, ... , M -

and DIFT is the Discrete Inverse Fourier Transform of the frequency

domain data (X(m) Y*(m)). Y* is the complex conjugate of Y.

-

18 ''. .-

-•

S .-. - S -. . . . . . . . . . . .S..-S. *.~S5~S.*S** ,.-o-~ -.. S..S. *.SJS*S*.S*.5

.S-S - .5

2. Choose the time shift constant k such that p(k) is maximized.max

For a particular set of two targets this yields Pt,r, r = 1, M,

where M = number of catalog members.

max max max3. Choose the largest m,r• If Pt,q = Max (Pt,r), classify the

target (t) as q.

4. Repeat experiment several times for each level of noise power.

Compile statistics as a function of noise power.

-2j

19

CHAPTER III

EXPERIMENTAL CONSIDERATIONS

A. INTRODUCTION (

Figure 3.1 sunmmarizes the experimental classification procedure inL

the form of a flow chart. Data for M vehlc'es at a total of N

frequencies are contained in the database. The database is a directory -

of scaled data files, each corresponding to a vehicle at a particular

* aspect angle, elevation angle and polarization. Amplitudes and phases

* at the desired frequencies, aspect and polarizations are selected for

each vehicle in the database to form a catelog (this is equivalent to

* feature extraction, see Chapter 2, Section E). This selection, in a

practical situation, would be based on all of the a priori information

* pertaining to the unknown target. Zero mean Gaussian errors are then

* added to the entire catalog to produce a set of test targets.

* Classification proceeds for each noisy test target and decision

statistics are compiled. It is worth emphasizing that in the followingr

- experiments, the whole catalog of vehicle amplitude and phase returns is

corrupted by zero mean Gaussian errors and classification proceeds for

* each target in that "noisy catalog". At this point there are two

* catalogs; one an error-contaminated version of the other. The error-

* corrupted RCS values of the first vehicle are compared (by the various

* algorithms) with those of the M noise-free vehicles and a decision is

* 20

- N IS C NTA I TE CATALOG N I E C N A I E CATALOG

wk.) z 86f ki18k~fi+I) k~td DIFT A~

DCISIONAE DEAONIE C ISONAE AAO

COMPCLESSIFCTIORLATO COMPILEFSTATION

I.( E (fROR PROA )IIT R ROR PROABIIT

lot t O XP t I+-r)-4-

Figur 3. APlwCato h xeietlpoeso liia tin P

COMPILE* STT

21

made as to the identity of the test target. Since we added the noise to L

the test target, we know its identity beforehand, and can therefore

* ~determine if the classification was either correct or incorrect. We say ka

I a target has been identified correctly if we choose the right vehicle at

the right aspect and elevation angle. A target is misclassified if we

choose either the wrong vehicle, or the right vehicle at the wrong

Ielevation or aspect angle. This decision process is justified on the

* basis that we are investigating the classification properties of

parameters, such as aspect angle, polarization etc., rather than the

classification properties of individual vehicles.

The process is then repreated for subsequent members of the noisy

- catalog until all M error-contaminated test targets have been

classified. The number of misclassifications is recorded for the

particular level of injected noise power. Hence the statistics of the ~.:

-experiments apply to the collection of vehicles as a whole and

I classification properties between individual vehicles, although of

interest, are not investigated here.

The experiment is repeated a number of times in order to compile

meaningful statistics, i.e. representative curves (see below). The

entire process is then repeated for a different injected noise power so

- that curves of misclassification percentage versus post-processing SNR

can be drawn.

22

B. EXPERIMENTAL FREQUENCIES

The time domain impulse response has (by rule of thumb) a

worst-case duration of 6 transit times across a target of length L. In

order to satisfy Shannon's Sampling Theorem, we require that the

frequency sampling interval should satisfy

Af -c c/6L. (3.1)

where c is the speed of light.

The vehicle dimensions are on the order of 6 m, implying

Af 6 MHz. (3.2)

Thus, Af = 5 MHz was selected as the frequency increment. The

corresponding selection of frequencies is given in Table 3.1. .-

C. NOISE MODEL

Based on arguments discussed by Chen [2], and on the Central Limit

Theorem, we assume a Gaussian noise model. Figure 3.2 shows a noise-

free vector which is contaminated by adding in-phase and quadrature

Gaussian noise components. The resultant vector has real and imaginary

terms thus

I=A cose + n1 (3.3)

Q=A sine + nQ (3.4)

23

.4-

h IA

TABLE 3.1

RA4GE OF FREQUENCIES USED IN EXPERIMENTS

N = 2 25 MHz 4 f < 30 MHz

N = 4 25 MHz 4 f 4 40 MHz

Varying number of frequenciesN = 8 25 MHz 4 f 4 60 MHz

N = 12 25 MHz 4 f < 80 MHz

Band 1 25 MHz < f 4 60 MHz

Band 2 70 MHz <f <105 MHz Varying i ,ency sub-band inthe available 25 - 200 MHz.Band 3 115 MHz 4 f 4 150 MHz (8 frequencies).

Band 4 160 MHz 4 f < 195 MHz

22

177

24".

S---

no

A

Figure 3.2 The distribution of the noise on an I-Q plane (from [2]).

where n1 and nQ are independent Gaussian-distributed random variables

2with zero mean and variance a The power in the component n1 is equal

to that of n The noise amplitude is given as

An n1 + n Q (3.5)

and is Rayleigh distributed [5]. The noise phase is given as

en=tan Q(3.6)nn

and is uniformly distributed [5]. An expression for signal to noise

* 25

ratio is given as

S I2 + Q2 A2 ( )

VAR(n) + VAR(nQ) .).

where

2 is the noise power, and

2A2 is the average signal power estimated as

A2 1 NF NS 2 ..- NF.NNS 1 7 Aij :)""'

i=1 j=1

where

A.. is the amplitude for the ith frequency arid the jth vehicle

i is a frequency index, i = 1,2, ... , NF = number of frequencies.

and

j is a target index, j = 1,2, ... , NS = number of vehicles.r . -;,. . -

D. ESTIMATION OF THE PROBABILITY OF MISCLASSIFICATION

Chen [2] used the Maximum Likelihood EstimatE (MLE) [6] as the " * .-.

estimate of probability of error for a given target;

PE E (3.8)

where PE is the probability of error the PE is the MLE of the

probability of error. The proximity of the terms expressed in the

equation above is usually stated in terms of confidence interval. For

example, a 90% confidence interval at 30% error (a typv.al error value)

for 10 targets and 50 experiments is ., -

26

PE E 3.%(39

It must be emphasized that the confidence interval is not an expression

of probability, i.e., the above calculation does not show that 9 times

out of 10 the actual probability of error at 30% misclassification will

be within ±3.4% (or 26.6% < PE < 33.4%). In regard to this, we must

view confidence interval as an indication of the accuracy of a measured

V result, not as an absolute probability.

In classification experiments, the MLE of the probability of error

is estimated by finding the percentage of misclassification to total

number of experiments.

E. POST PROCESSED SIGNAL TO NOISE RATIO

S We define post-processing signal-to-noise ratio (post-processing

SNR) as the ratio of signal power to error variance after the received

K waveform has been processed to produce a final best estimate of the

amplitude or phase, or both, of the target radar cross section.

In future experiments, 10 dB post-processing SNR will often be used

as a reference value when comparing misclassification percentages. We.

do this mainly because the level of 10 dB starts to give us "useful"

values of misclassification percentage, (30% and lower), and therefore

represents a region of interest.

In a real situation, several measurements of a target's radar

cross-section will be made, and the 10 dB figure represents an

integrated SNR over the several measurements. To this extent, it is

27

::.9

, "- h

assumed that the radar system designer can provide us with RCS

measurements at a nominal 0 dB post-processed SNR and that integration

of measurements will yield the "useable" 10 dB or better post-processed -

SNR. .

19

28.. . . . . . . . . . . . . . . . . . . . . . . .. . . . . ..

CHAPTER IV ~

EXPER IMENTS

A. INTRODUCTION

This chapter presents most of the experimental work done in this

K study. The purpose of the experiments discussed here is to study

r classification behavior under a wide variety of the available

classification parameters. These are listed below (a list of commonly

used terms is given in Table 4.1).

1. Frequency Band (in the allotted 25-200 MHz)

2. Number of Frequencies and Algorithm

V 3. Polarization p.

4. Aspect Angle

5. Error in Aspect Angle

If each of these parameters were to be assessed in terms of the

others, thousands of curves would be needed. Clearly this is not

practical; however, an intelligent approach toward choosing the

experiments allows a thorough investigation without incurring excessive

data processing._

First, experimenting with a sub-band in the 25-200 MHz band

indicates the best region of frequencies (if any) for a particular type . *,

29

TABLE 4.1 L

DEFINITIONS

1. ASPECT ZONE: A small range of aspect angles centered on or adjacentto a particular aspect angle.

Aspect Zone Name Aspect Angles

00 Nose-On 00, 100450 ---- 400, 450, 500

90Broadside 800, 900

2. KNOWN ASPECT: If vehicle data having more than one aspect angle isused in a classification, then the aspect angle is assumedL..-known when the algorithm only allows comparisons between an'unknown' noisy target at aspect OK, with 'noise-free' catalogtargets at the same aspect angle eK. The same follows forknown elevation.

3. UNKNOWN ASPECT: If vehicle having more than one aspect angle isused in a classification, then the aspect angle is assumedunknown when the algorithm only allows comparisons between anunknown' noisy target at aspect eK, with all 'noise-recatalog targets at the same aspect angles used in thecl assi f icati on .

4. ALGORTHIM: One of the following:

1. Nearest neighbor (NN), using any of the features listedin 5.

2. Correlation Algorithm.

5. NN ALGORITHM FEATURES:

A Amplitude only

W Differential phase only

AW Amplitude and differential phase

Note that, in general , the term 'feature', in the context ofresonance region radar returns, applies to any property orquality associated with the returns. This includes, forexample, the magnitudes of the time domain impulse response. -

30

TABLE 4.1

(Conti nued)

6. PARAMETER: A variable in the measuremet or classification processsuch as frequency, aspect angle, elevation angle,polarization, feature or algorithm.

7. DATA BASE: A collection of data files containing a measurementvector for each vehicle.

8. CATALOG: A single 2-dimensional complex array containing theamplitudes and phases of all vehicles at the selected

b frequencies, and other parameters.

9. FEATURE VECTOR: A single vector for a particular target,dimensional in frequency or time, derived from the catalog byselecting a particular feature.

10. POLARIZATION: One of the following: F~

1. Vertical, V (send vertical polarization, receivevertical)

P2. Horizontal, H (send horizontal polarization, receivehorizontal)

3. Vertical divided by horizontal, V/H.

-- The termi described in 10.3 is not really a polarizationscheme, in the sense of the preceding three terms, but iscalled a "polarization" for convenience. Strictly speaking,the polarization of a wave describes the instantaneousorientation of the electric field vector; the terms listedabove refer to a particular measurement scheme or use of radarcross section returns, with respect to polarization.

31

AN-

of classification. Second, an evaluation of classification performance

* as a function of the number of frequencies (NF) provides a comparison

with work presented in [1] and also establishes the ranking of a

particular number of frequencies. Hence we can choose NF=8 for

subsequent experiments and refer to this section for results pertaining

to other values of NF. The problem of choosing the desired ~L.

classification frequencies is thus solved.

The remaining parameters are then classified using various features

and algorithms, polarizations and aspect zones. It must be remembef'ed

that certain parameters, such as V/H polarization, are not really tested

in these experiments since the difficulties which they were designed to ~

* overcome were not simulated. V/H polarization purportedly has the

property that multiplicative errors cancel out by virtue of the division

of a vertically polarized phasor by a horizontally polarized phasor. To

test the validity of this assumption, the vehicle data must be

* contaminated with multiplicative noise before classification proceeds. ~

This was not done, because a reliable multiplicative noise model was not

available at the time of experimentation.

In the following sections, each experiment is introduced, detailing

the aim of the experiments with a brief discussion of any relevent

points. The results of all experiments in the form of misclassification

percentage versus post-processing SNR curves, are contained in Appendix

A. These results are summarized by means of histograms representing

averaged misclassification percentage at 10 dB post-processing SNR, and

*are presented, along with typical misclassification curves.

32

Conclusions drawn from the histograms often compare the

classification performance of one parameter with the performance of

another by saying that misclassification percentage (at 10 dB

post-procssing SNR) is higher or lower by x%; here x is always the A

difference between the misclassification percentages, not the percentage

increase of one misclassification percentage compared with another.,',

33 i' .-

LI% m

,- , r~-- . ",

-p' .:

m",

,

experiments. The header comprises of up to 11 lines and these are

-" described as follows:

1. LINE 1. TARGET TYPE. i.e., ships, aircraft, ground

vehicles, etc. ....

2. LINE 2. POLARIZATION This can be either vertLcal('V'), horizontal ('H'), cross e:Xs) oroflasifcaio

vertical divided by horizontal V/H'). Thepolarization for each curve is printed if thisvaries from curve to curve.

3. LINE 3. A PRIORI KNOWLEDGE OF ELEVATION ANGLE. This

can be either 'known', 'unknown' or 'known/verticalunknown' if some of the curves use known

elevation angle and others use unknownelevation angle. The system is specific aboutwhich is known and which is unknown only if 2curves are present.

4. LINE 4. ELEVATION ANGLE(S). This is '27' for 270.

5. LINE 5. A PRIORI KNOWLEDGE OF ASPECT ANGLE. This iscompiled in the same way as LINE 3.

6. LINE 6. MINIMUM, MAXIMUM AND INCREMENT OF ASPECT. Theminimum and maximum aspect angles can be any ofthose discussed in Chapter 2, Section A. Ifthese parameters vary from curve to curve, theyare printed out for each curve in the same way J

as LINE 4 (or LINE 2).

7. LINE 7. NUMBER OF FREQUENCIES. This is printed foreach curve.

34

9%' "

i')' -'.- .." '..2 -.'-..-'.'-" " .-.'..'-.'- "..'. -s:. " .'..'. "-=.-"- ... -.- ", " :. ". ". ". "..'-.'.''.-".- ," .- --.." ;-""-1

TABLE 4.2

(Continued) .*V'I

8. LINE 8. NUMBER OF TARGETS. This is always a multiple of 5,the number of vehicles (No. of targets = 5 x No.aspect angles x No. elevation angles). This isprinted for each curve.

1.7"9. LINE 9. 90% CONFIDENCE INTERVAL AT 30% MISCLASSIFICATION.

This is calculated according to the discussion In I,Chapter 3, Section D, and is printed for eachcurve.

10. LINE 10. CLASSIFICATION FEATURES. These are listed as:

'A' Amplitudes only NNW' Differential phases only algorithm

'AW' Amplitudes and phases'T' orrelation algorithm. "

11. LINE 11. IDENTITY OF CURVES WITH ASPECT OR ELEVATION ANGLE -ERRORS. This is printed only if there is an errorin elevation angle or aspect angle. Note that theclassification software does not allow both types oferrors at once. For example, if 3 curves wereplotted and the last two had aspect errors, LINE 11would read

'ASPECT ERR IN CURVE 2 3'

Curves are identified as follows:

Curve I-- Curve 2

-- - -- .... Curve 3Curve 4"-" "'•................. C.urve 4

Curve 5.......... . Curve 6

3. 5,I.

35 .'..,

".. 1 =

.'...-.....-.-. "-."-..-."

B. FREQUENCY BAND

I J

1. INTRODUCTION

rhis set of experiments was designed to examine the effect of

* choosing a particular frequency band for various aspect angles and

*pol ari zati ons. In each experiment, the Nearest Neighbor algorithm with ~*I

the 'amplitudes only' (A) feature was used.

The availalbe band of 25-200 MHz was split into 4 non-overlapping I

* sub-bands, (see Table 3.1) each contianing 8 discrete frequencies -

I separated by 5 MHz.

Generally, it is expected that the lower bands of frequencies will

* provide the best classification performance. This is because the RCS at

lower frequencies tends to vary less per unit bandwidth compared with

higher frequencies. Consequently, for higher frequencies, (in the upper

resonance region) a small change in a parameter such as aspect angle, or ~ '

fl the addition of noise to the RCS amplitudes results in a greater change

in the selected features compared to the change at lower frequencies.

In this sense, the RCS frequency response is more reliable (i.e., -2*impervious to small changes in orientation, frequency, etc.) at lower

- frequencies than at higher frequencies.

2. RESULTS AND CONCLUSIONS

Figures 4.1 and 4.2 show a summary of classification results by

*means of average misclassification percentages at 10 dB post-processing

36 7

TNT~~~.'J IT7 .*1*_-Pjj W1 r

.- ,

SNR. The curves from which these histograms are extracted are contained

in Appendix A.

From Figure 4.1 it is evident that for vertical and horizontal

polarizations, there is a small but distinct preference for the lower

frequency band. However, V/H polarization does not follow this pattern,

and favors band 4.

Figure 4.2 shows that there is no distinct nor significant

preference for a particular sub-band, with respect to aspect zone. This

result is similar to the finding in the classification of ship targets[1].

Averaging across all parameters, band 2 does best with 6%

misclassification and band 3 does worst with 12% misclassification. In

general though, the distinct preference for lower frequencies

established by ship targets, is not so marked for the case of ground

vehicles. Figures 4.1 and 4.2 also show that the variation of

performance with different bands is not great, indicating that selection

of frequencies In the classification of ground vehicles is not a

critical parameter (at least with the 25-200 MHz band used in these

experiments).

37

-.. *...*. . . .-..

, -. ".. . " ". ",. ". -._ .,. '. . " ". '...... ". ".. -. ... "° -.. .... .. ' . j -, - ,*. . - '.." ,' ,'-" ," ' -. ,, -.".."-" ." '.. ,." , " .- T. ,'..'. -'--9.

~.. ,%

.*.

wIOO "-- BAND I

w BAND 2w,8o0 BAND 3 .. ,-

hi BAND 4'160 - .. ;

I40"

20

V H V/H

Figure 4.1. Classification performance of various sub-bands in theavailable 25-200 MHz band, as a function of polarization.

0 BAND I

Z BAND 2I80 - BAND 3

6 BAND 4

o-40

U)U)20 -'': " "

T 20

00 450 90-

Figure 4.2. Classification performance of various sub-bands in the tjavailable 25-200 MHz band, as a function of aspect zone.

38

-. . ''.

C. NUMBER OF FREQUENCIES AND ALGORITHM ~11

1. INTRODUCTION

The purpose of this experiment was to examine how the number of

classification frequencies affects other classification parameters, such

as polarization, aspect angle and algorithm. In subsequent experiments,

K 8 frequencies are used; the classification performance for other numbers

of frequencies can then be extrapolated from results obtained in this

experiment. Three feature (A, AW, W) from the NN algorithm, and the

correlation algorithm were evaluated at 2, 4, 6, 8 and 12 frequencies, %

using vertical polarization 270 elevation and aspect zones of 00, 450

and 900 (aspect angles assumed known).

2. RESULTS AND CONCLUSIONS .

Figures 4.3 through 4.5 all show that classification performance

improves monotonically with the number of frequencies, with 2

frequencies yielding most miscassifications and 12 frequencies yielding

the least. This result is intuitively reasonable, since a larger number

of frequencies means we have more information concerning a target, and

the more information we have, the more likely we are to classify it --

correctly.

From Figure 4.3, it is evident that vertical polarization provides

best performance, followed by horizontal, then V/H polarization.

Furthermore, these results apply for all numbers of frequencies. In

39

7 - -3;.1.:.J )F. 7.-~---.

addition to this, if the misclassification percentage at 8 frequencies k

is relatively low, the reduction in misclassification percentage, when .

going from 8 to 12 frequencies, is relatively small. (e.g., vertical

*polarization: 10% misclassification at NF=8, 8.5% at NF=12). On theA

other hand, this reduction is relatively large when the level of

misclassification at 8 frequencies is relatively high (e.g., V/H

polarization: 32% at NF=8, 25% at NF=12). Hence the advantage of using

a higher number of frequencies depends on the misclassification levels

of a particular parameter; the higher levels (20% and above) yielding

greater improvements.

Figure 4.4 shows a general preference for 00 aspect zone, which is

shared somewhat by 900 aspect zone for 8 and above frequencies. As in

the case of vertical polarization, 00 aspect zone shows little reduction

in misclassification percentage when going from 8 to 12 frequencies.

Generally speaking, for 4 or more frequencies, the correlation

* algorithm provides best performance, followed by A, then AW, then W.

* However, for 2 frequencies, the correlation algorithm, 'T', gives almost

* the worst performance. If one thinks of the correlation process as a

7'. time domain operation, then it becomes clear that the resulting

resolution will be poor because 2 frequencies is the absolute minimum

for the Inverse Fourier Transform. Figure 4.5 also shows that a higher

number of frequencies is more effective in reducing misclassifications

when the algorithm is relatively favorable (e.g., WA, 'T') compared to

an algorithm which is relatively unfavorable (e.g., 'AW', ')

40

lr

• ,hi

0100- I 2 FREQUENCIES h.-'

Z 4 FREQUENCIES80 - 8 FREQUENCIES

w 12 FREQUENCIES

0

60-

20-

20 I

V H V/H

Figure 4.3. Classification performance of various numbers offrequencies as a function of polarization.

hi_

V4 0I100- 2 FREQUENCIES

Z 4 FREQUENCIESpi S 0 - I S FREQUENCIES

h 12 FREQUENCIES-

0

IL ~40-

"i20 V/ "'"U)

I. ~Figure 4.4. Classification performance of various numbers of,-.'-.frequencies as a function of aspect zone.

.41

'- t"

U _ __ __°__o___ _,___

450 9,0

• .'. Figur. 4.4.." . " w..°. "... of. various. numbers.of

.. . . .. . ..' ".-"' . "- "-- ", "frequencies'.'.X,_',.- as._- '.' a function'. of aspect zone.T.",-. '5, ,-". ..,. , . _ " " ' -_. ' "-."._ _ ., __ _ _ ,_ -"." ". ",

bL %..1 ..

~J2 FREQUENCIEStOO ____4 FREQUENCIES

z 8 FREQUENCIESw 12 IFREQUENCIES

~60* 0

40-

U -. " ,

,"

* 4"

.~~ ........

A A W W T

Figure 4.5. Classification performance of various numbers of-frequencies as a function of algorithm.

424

424

- -

!.. ".1

Although it was stated that the correlation algorithm had best

performance (at least for 4 and above frequencies), this performance is

not substantially better than the NN algorithm using amplitudes only (1

or 2% difference). Furthermore, the correlation algorithm requires both

amplitude and phase information, the latter not always being readily

available. In view of these facts, and also of the correlation

algorithm's poor performance with only two frequencies, 'A' must be

regarded as having best all-around performance, since it provides near- ,-~

best performance for the minimum of information, consistently over a

range of NF.

The NN algorithm using amplitudes, 'A' and relative 'phases' W,

seems to always have worse classification performance than 'A' alone.

* This finding runs contrary to intuition, where one would expect the

addition of phase information to improve the chances of correct

classification. This anomaly has been the subject of further research

* and as yet, has not been resolved.

D. POLARIZATION

1. INTRODUCTION

The purpose of this experiment was to determine the relative

classification performance of each polarization; vertical, horizontal

and vertical divided by horizontal, as a function of aspect zone and

algorithm. Eight frequencies were used in this experiment.

43...

2. RESULTS AND CONCLUSION

Figure 4.7 shows that, for each algorithm, there is a distinct

preference for vertical polarization, followed by horizontal, with V/H

polarization generally doing worst. At 450 aspect zone (Figure 4.6) V/H

is particularly bad, with an average misclassification level of 47%.

Horizontal polarization is best at 450 and worst at 00. For the

*case of ship :argets [1], horizontal polarization was better at 900 than

at 00 aspect zone (which is the case here). However, for vertical

*polarization, ground-vehicles were more readily classified at 900 aspect

zone (5% misclassification on average) compared to 00 aspect zone (10%

misclassification). This is opposite to the finding for ship targets,

where vertical polarization favored 00 aspect zone.

For ship targets, it was fairly easy to attribute classification

results to the more likely scattering mechanisms associated with a ~

particular orientation and polarization. This was due mainly to the

'distinctness' of the ship structure; being long and thin, and usually

with two or more vertical structures such as masts. Ground vehicles are

* more amorphous in their shape being fairly squat and featureless,

* bar the gun turrets of tanks. Hence, for the case of ground vehicles,

IF it becomes more difficult to justify results in terms of possible

scattering mechanisms.

44

7NIM ° Tl,.,

w too VERTICAL POLARIZATION

HORIZONTAL POLARIZATION 41

Z V/H POLARIZATION

U n

00 900V. IL

Figure 4.6. Classification performance of various polarizations as a =''-:function of aspect zone.-

,.'w 1 00 VERTICAL POLARIZATION" 1 HORIZONTAL POLARIZATION,8s-

I w V/H POLARIZATION

U0

60 "'-'-"

S404"0"90

., o i :,-0 A AW w T

Figure 4.7. Classification performance of various polarizations as afunction of algorithm.

45"

HORIZONAL POLAIZATIO

~ 80 VH POLAIZATIO

-p U

% , .

-14mE. ASPECT ZONE

1. INTRODUCTION

This experiment was designed to investigate how classification is

affected by a particular aspect zone. From the above experiments, it is

evident that classification performance is dependent on the azimuithal

orientation of the ship (i.e., aspect angle) and hence, a thorough

investigation of classification performance at given aspect angles is of

interest.

Classification performance of three aspect zones; 00, 450 and 900

were examined as a function of polarization (V, H, V/H) and algorithm

(A, AW, W, T). Eight frequencies were used in each experiment.

2. RESULTS AND CONCLUSIONS

Figure 4.8 illustrates the classification performance at various

aspect zones, for different polarizations. It is evident that the

performance at any one particular aspect zone is dependent on the

polarization, and that in general, no single aspect zone has a superior

performance. Hence, while 900 aspect zone has best classification

performance using vertical polarization, it is not so good when

horizontal polarization is used (450 aspect zone being favored).

The results depicted in Figure 4.9 tend to amplify the above

findings. The fact that 450 aspect zone appears to have the poorest

performance for each algorithm is due to the particularly bad

performance of 450 aspect zone using V/H polarization (see Figure 4.8).

This tends to bias the result, somewhat, when the average values used in

these histograms are derived.

46

I%

- , Z . W71

w 100 - 0- ASPECT ZONE

450 ASPECT ZONEw 80- 900 ASPECT ZONE

w

S40-

0

Ia: V N V/H

If Figure 4.8. Classification performance of various aspect zones as a

function of polarization.

w:: 1 00 - 00 ASPECT ZONE,,..(,

" - 1Z 45" ASPECT ZONE,.,-,%

. u 90" ASPECT ZONE _.C

Z Go-"" ""

~40

0 •1

, 9-.,,..

20

Figure 4.9. Classification performance of various aspect zones as afunction of algorithm.

47 " " "

II: ASPECT ZONE.IT; .'..ASPECT ZONE

, ~~....::.]

ASPECT ZONE'i. . . , , , . . , . .. , .,•... .. ."-.. .. . . . , .h . . . . . . .. - . ... . .. . .. .. . . . . " .i.::,'

M. I,- A .. _t _ u .nrm k Kxuw , W oF X9

F. ERROR IN ASPECT ANGLE

1. INTRODUCTION

The heading of a moving vehicle, and hence its aspect can be found

quite accurately by simply plotting its position at two instances of

time. Knowing the time interval and range between the reference

positions allows the average velocity to be calculated, and this in

conjunction with the doppler shifts at the two positions allows the

aspect measurement to be refined. -

The aspect angle of a stationary target is more difficult to

estimate, and it is possible that such an estimation might contain a

large error.

The purpose of this experiment was to examine classification Lperformance as a function of various parameters when an aspect error is

introduced. For 0° aspect zone, using 00 and 100 aspects, a 1 0° error - '

was simulated by comparing 00 error-contaminated ship returns with 100 []

catalog ship returns and vice versa. A 450 aspect zone, a ± 50 error

was introduced using a catalog comprising vehicle returns at 400, 450

and 500. At 900 aspect zone a ± 100 error was used with data at 800 and

900 aspect.

2. RESULTS AND LONCLUSIONS I.L. -

In general, Figures 4.10 and 4.11 reveal that an error in aspect -.

angle results in a significant degradation in classification

performance. Comparing Figures 4.10 and 4.8, an error in aspect angle

48

S*..*. -""* ...,

- . .. o . . .° • .. - *- • -. 1 - . ~. - , -.. . I ~. . " . -° l - - o . - * * -. -. -.. - - . - . -,* ° -__-____-_°-.__________.__________-___________________-° ° %°

increases misclassification by between 10% and 20%. No single

polarization, or aspect angle seems immune to the effects of an error in

aspect angle. The correlation algorithm (T) seems to be particularly

affected by the forced error, with an average increase of about 40%

misclassification over the no-error case.

t3. '49

W 00"-0 SETZN

Z 45 ASPECT ZONE

So- 900 ASPECT ZONE

z 60 ..

40

~20

S0

C A AW W T

Figure 4.10. Classification performance of various aspect zones withforced errors in aspect angle, as a function ofpolarization.

00:W 00 ASPECT ZONE '

* 45° ASPECT ZONE0 90 ASPECT ZONE

W0-

Z0

40-

~220 .

-0

V H V/H

Figure 4.11. Classification performance of various aspect zones withforced errors in aspect angle, as a function ofalgorithm.

50

L .

CHAPTER V

CONCLUSIONS

A series of experiments investigating the classification properties 1

of a group of ground veh-icles at various azimuthal (aspect) angles,

using a variety of polarizations were performed. These experiments were

intended, to a certain extent, to represent classification based on VHF

resonance radar. A number of important findings follow from the work.

Of primary importance is the fact that it is possible to identify

K ground vehicle targets in a representatively error-prone environment, at

"useful" classification rates. For example, at 10 dB post-processing

SNR, the database of vehicles were classified at an average rate of 90%,

using 4 classification frequencies and vertical polarization.

Furthermore, we can discriminate effectively among tanks, APC's and

jeeps, even though the general shape of the vehicles is quite similar.

It is also encouraging that the NN algorithm, incorporating the

simple amplitude-only feature, performed quite well. (About 10%

misclassification for 4 frequencies, and 3% for 8 frequencies, using

vertical polarization). In view of the relative simplicity of acquiring

the amplitude-only features (compared with the acquisition of phase

data) this is a particularly useful result.

51

Generally speaking, the classification performance of the vehicle

set was found to be dependent on both aspect zone and polarization.

However, no single polarization or aspect zone proved to be

substantially superior. For example, while horizontal polarization gave

lowest misclassifications at the 450 aspect zone, this was not the case

at the 900 aspect zone, where vertical polarization had lower *.*

misclassifications. The rate of misclassifications using vertical and

horizontal polarization was found to range from 10% to 20%, depending on

* the azimuthal orientation of the target. This shows that, while some

advantage is gained from having a dual-polarization radar system, the

consequences of not having polarization agility are not drastic.

The classification performance of V/H polarization was found to be

*generally poorer than that of either vertical or horizontal%

Spolarization. The original impetus for the use of this V/H feature

stemmed from the fact that such a division of RCS amplitudes would tend ito reduce the effect of multiplicative errors. (A possible source of

multiplicative noise might be foliage.) If such errors were present in

the RCS data, then the V/H features might prove to classify better than

either the vertical or horizontal polarization features. Hence the .'*-.

relatively poor performance of V/H polarization should be viewed with

* this potential advantage in mind.

The introduction of a small, forced error in aspect angle was found

to significantly degrade classification performance, with

misclassification percentages of 30% and above being typical. However, C

additional research in this area has revealed that the majority of

52

misclassification at 10 dB post-processing SNR are the right target

type, but at the wrong aspect angle. For reasons stated earlier, a

target classified correctly by type but incorrectly by aspect angle was

considered to be a misclassification. In practice it does not matter if

the aspect angle is wrong, as long as the target type (i.e., Tank 3,

Jeep 1, etc.) is correct. Hence the consequences of inaccurately

estimating the aspect angle of a target, ( 100 errors in these

experiments), are not as severe as first suggested by the results. A

limited study showed that, if an error of ± 100 is made in the

estimation of aspect angle, then the misclassification percentage at 10

dB post-processing SNR increases by only 2 or 3 percent.

Clearly, there is a substantial potential for further work using

this database of 5 ground-based vehicles. New classification algorithms

and techniques are currently being developed. As classification

procedures gradually become more refined, we expect to see

classification rates improve on the already useful values demonstrated

in this study.

*r *.

5Not

53

it -.7.,

REFERENCES

[1] N.F. Chamberlain, "Surface Ship Classification UsingMultipolarization, Multifrequency Sky-Wave esonance Radar,"Technical Report 714190-9, Department of El .ctrical Engineering,The Ohio State University, ElectroScience Liboratory, Columbus,Ohio, 1984.

[2] J. Chen, "Automatic Target Classification Using HF MultifrequencyRadars," Ph.D. Dissertation, The Ohio State University, 1983.

[3] E.M. Kennaugh and D.L. Moffatt, "Transient and Impulse ResponseApproximations, "Proc. IEEE, Vol. 53, pp. 893-901, August 1965.

* [4] H.I. Skolnik, Introduction to Radar Systems, McGraw-Hill, New York, .-

1980.

[5] A. Papoulis, Probability, Random Variables, and StochasticProcesses, McGraw-mill, New York, 1965.

[6] P.G. Hoel, S.C. Port and C.J. Stone, Introduction to StatisticalTheory, Houghton Mifflin, Boston, 1981.

[7] D.F. Kimball, "Calibration Techniques for Broadband RadarBackscatter Measurements," M.Sc. Thesis, Department of ElectricalEngineering, ElectroScience Laboratory, The Ohio State University,1983.

A 54

* ° "

APPENDIX A

CLASSIFICATION RESULTS ME_

This appendix contains plots of misclassification percentage versus

post-processing signal-to-noise-ratio for the experiments of Chapter IV.

A guide to interpreting the headers of these curves is given in Table

4.2.

55'.

...........-

. . . . .. ... .-.-.-..... ..

So "IIq'-? a' - r S=- tra -.<6Mr s . .T r2-w 7 - . .-. Jr -& . t. - - Sk42-. Z Lft l.~lf

CLASSIFICATION OF GND VCLSPOLARIZRT11N VELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMINMAXINC ASPECT 0 10 10NO OF FREQUENCIES 8 8 8 8NO OF TARGETS 10 10 10 10907. CI (030%) .-4% 3.14% 3.4% 3.4.,CLASS. FEATURES A A A A

C; BAND 1- --- - AND 2----- ---- AND 3

.0.... .. BAND 'I

C;ii ........... N.

L±J

LLU

1L)

L)

.. ...... .-. .

-5. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN DB

Figure A.1 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200MHz band.

56

VK:- -. :.-::.-.'..:

POLAI1ZA'T ION HELEV ASSUMED KNOWN.

ASELET ASSUMED KNOWNELEVATION IDEG.) 27 ., ,,,

MIN.MAX, INC ASPECT a 10 10NO OF FREQUENCIES 8 a a aNO OF TARGETS 10 10 10 10 ll90X CI 1030X) -/- 3.47 3.0 3.0 3.YJXCLASS. FEATURES A A A A

SAND I

z .. . °..°- - - - $ D

08-

.. . .i ... .. ....... ... ..... ,.

,-,~ ~ ~ ~ ~ ~... ..... '.. .. .... . ...... ........._.... . .- -4 .. . .....

................. ....... ..... .. .....

o _"

,,' ... ---' .............. --- ..............-..............-

-. o -5. 0. 5. 10. 15. 20.£POST-PROCESSING S/N IN 08

F- Figure A.2 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200MHz band.

57

<,'> I ; -,. ,. '*

.. .. 5 .. ... .U . 5... .. . . . . . . .':.:

. I . .. * . * * . . * ... . ~ * ~ . ' .*-o.-.*..

CLASSIFICRTION OF GNO VCLS

POLARIZATION V/HELEV ASSUMED KNOWNELEVAI1ON IDEG.) 27ASPECT RSSUMED KNOWN

MIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 8 8 a 8NO OF TARGETS 10 10 10 1090% CI (030%) +/- 3.4Z 3.4% 3.Y% 3.Y% KCLASS. FEATURES A A A A

0 _______ BANO I- -- AN - - BRNO 2

--------- UAND 3o........... IRNO

I.-- -w ... ... . . . . . . .... . ....... ........

Z ~7

LJ

; .......-.. ,...

0

,.- ".f', " t

... . . . .... . .. .

C-O. -5. 0. 5. 10 15. 20.

POST-PROCESSING S/N IN DB

Figure A.3 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200

-. MHz band.

58

. . . . . . ..•4 ~ ". . . . .w . -,•... . . . .. ". "•. ". °- " =...

CLASSIFICATION OF GO VCLSPOLARIZATION VELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MRX.INC ASPECT 40 50 5NO OF FREQUENCIES a a a aNO OF TARGETS 15 15 15 i590%/ CI (630%) 4/- 2.8% 2.8% 2.8Y. 2.8V.CLASS. FEATURES A A A A ~

.....- - ------- - . .. --....r; T BAND? I- --- - UANDO2

-r CDOI

LU.

0

VZ... ---- ----)

S-- -, -----

-?. -5. 0. 5. 10. 15. 20.POST-PROCESSI NG S/N IN OB

V Figure A.4 Misclassification percentage versus post-processing SNR,comparing the performance of 4 sub bands in the 25-200MHz band.

59 .*

CLASSIFICATION OF GND VCLS

POLPRIZPTION H dELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMINMAXINC ASPECT 40 50 5NO OF FREQUENCIES 8 8 8 BNO OF TARGETS is 15 15 1590% Ci (630M) +/- 2.8% 2.87 2.8% 2.8%CLASS. FEATURES A A A A

0 _$_ AND I%

BAND? 2

--- -- ----- ---..... SANDd L

oL .... .. ,. ...

LU

C.

.. ...... ....

0

-5. .. 5. 10.. 1. 20..

ry.

Figure A.5 Misclassification percentage versus post-processing SNR, '

comparing the performance of 4 sub bands in the 25-200MHz band.

60

A1kru

CLASSIFICAT1ON OF GO VCLSPOLPRIZR7ION V/HELEY ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN MAX.INC ASPECT 40 50 5NO OF FREQUENCIES 6 8 8 8NO OF TARGETS 15 15 15 15907. CI (030/) +/- 2.87 2.8% 2.8%. 2.8%CLASS. FEATURES A A A A

j; BAND I- --- - UAND 2----- ---- AND 3

JiLUC-)

OL)

a -? -. 0. ....5 ... . ... 20.

POTPOESNZ/ NOFgr A . Micasfcto pecntg versus pos-prcesin ... S.......NR..........

................. ..............

CLASSIFICATION OF (NO VCLSPOLARIZATION VELEV ASSUMED KNOWN

ELEVATION IDEG.) 27ASPECT ASSUMED KNOWNMIN,MRX. INC ASPECT 80 90 10NO OF FREQUENCIES 8 8 8 8 1NO OF TARGETS 10 10 10 10

90. CI 130) 4/- 3.Y% 3.Y% 3.4% 3.4%CLASS. FEATURES A A A A

... ... .i .. ........... :~~. ..... .... . ......... ... .... ...CD BRNO I 1

-l AND 2

.. .... .. . ........----- AND 3

................. BAND 4

- ... .-.... . ... ...... .... ..... .......... ;. ... ..........LLJ\ -. ,.0.... .... . \ ...... .....

-,

Z

o. .. .... ..

o .I

• . . ..

% '..L. ..

.. . . ..

-TiO. -5. 0. . 10. i5. 20.PGST-PRIOCESSiNC S/N IN 05

Figure A.7 Misclassification percentage versus post-prcesslng SNR,comparing the performance of 4 sub bands in the 25-200MHz band.

62

.4

CLASSIFICA71ON OF GND VCLSPOLPRIZAT1ON HELEY ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN,MAX,]NC ASPECT 80 90 10NO OF FREQUENCIES 8 8 a 8

NO OF TARGETS 10 10 10 1090% CI (030%) +/- 3.4% 3.Y% 3.4%/ 3.4%

CLASS. FETRS*

L~~~~~BN I1o 5 . 5 5 0POST-PROCESSING S/2NO

Figure~ ~~~~~~~~----- BAND Micasfcto3ecnaeessps-rcsigSRcompaing he peformnce f 4 sb bads--nthe---20

M;z band....... AN

W6

0' .

F~~C I ..

LAh

CLASSIFICATION OF GO VCLSPOLPRIZATION V/HELEV ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 8 8 8 8NO OF TARGETS 10 10 10 1090. CI 1030X) +/- 3.4%/ 3.4/. 3.47/ 3.47.

CLASS. FEATURES A A A A ~

BAND I

- ------- BAND2

------ BAND 3

C; .............. AND 3

0

LULL..'

LL .

U

LoC " ;

POST-PROCESSING S/i'4 I,'Figure A.9 Misciassifi. ation percentage versus post-processing SNR,-

comparing t e performance of 4 sub bands in the 25-2)0MHz band.

64 ..-.

rV

CLASSIFICATION OF GNO VCLS ~POLARIZATION VELEV ASSUMED KNOWNELEVATION (DEG.) 2

MIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 4 a 12NO OF TARGETS 10 10 10 1090% CI 1030%) 4/- 3.4%/ 3.4Z 3.47 3.47%CLASS. FEATURES A A A A

L..._I2 FREQUENCIES

- 4 FREQUENCIES------ S FREQUENCIES .

......... 12 FREQuENCIE

-~~~~~~~ -- - --- ....... ....

CD(0

... .. .. ..

.. .... ... .... . .... ...... ..

........ . ..

coprn(hnefraneo auso FC;?

-._J j

L 65

-5. 0. . 0. is. 20

CLASSIFICATION OF GND VCLSPOLARIZATION HELEV ASSUMED KNOWN . .ELEVATION iDEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 10 10 10 10907 CI (30X) +/- 3. Y7 3.IL1 3.IY7 3.YXCLASS. FEATURES A A A A

..... .... ... 2 FREQUENCIES .

4I FREQUENCIES----------.. U FREQUENCIES

. ................. 12 FREQUENCIESa)

.z ........ . ........-- -.---- -... .....

LLJ ' ..- -

g i

a-. " .....~ -. .... .

wb

.. . ..... . ..,\,,..-, ._ \\ ""

CL C;,-. -

(Ir)

C

-TO. -5. 0, 5. 10. 15. 20.

POST-PROCESSING S/N IN DBFigure A.11 Misclissification percentage versus )ost-processing SNR, P.

:ompa-ing the perfcrmance of 4 value of NF.

66

. -. ~~ . :. . . : : : -.-. -

CLASSIFICAIION OF GND VCLSPOLARIZATION V/HELEV ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 LI 8 12NO OF TARGETS 10 10 10 109ox CI (03ox +/- 3.4X 3.4% 3.4X 3.4%.'CLASS. FEATURES A A A A

S . . 2 FREOUENCIES4 FREQUJENCIES

--------- FREQUENCIES. 2................ FREQUENCIES

LU

Z.a ". "-. . . ..,-. i:LaJ*

-- ", -,-,.Xi *

z.o ".. \ \0.-

U" '

-"-To. -5os 10. ,5. 20.POST-PROCESSING S/N IN OB I

I ) ~Figure A.12 Misclassification percentage versus post-processing SNR,,'"-,-..-,..'. ~~~comparing the performance of 4 values of NF..-"""-,..,

9.L

67

S.~ *.CC ~ ~ . . * * - * , . - . * . - . - . . . * . . . -. . . . .. ... . .. ....

to°% %

.C- %

• -'P,." " - " " "'," "" "_T "0.'L"

10. is.o' "20-"..' "" " " " " " "" " " ," ¢ ," ".'''. "£ " "* .

.0 %

L '- = .--.-- -- - -


POLARIZATION VELEV ASSUMED KNOWN

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMINMRX.INC ASPECT 40 50 5

NO OF FREQUENCIES 2 y 8 12

NO OF TARGETS 15 15 15 15907 CI 10307 +/- 2.87 2.87 2.8% 2.87CLASS. FEATURES A A A A

2 FREQUENCIESSI FREQUENCIES

......... .. .. - 8 FREQUENCIESD .. ........... 12 FREQUENCIES

(.'o- 'w'

LiiLo

Cc C.

. . . ... ' 'r-...'"; '

Li.J

CDz

;-:~ ~ ~~~r C;.,..\\\ ..

- ...

-- 5. 0. 5. 10. 15. 2 .

- POST-PROCESSING S/N IN DB , :-'Figure A.13 Misclassification percentage versus post-processing SNR, ,.

comparing the performance of 4 values of NF. ~

...

* - r..- r - w- r-v--r-r r. °- -r .

.7.

CLASSIFICATION OF GNO VCLS

POLARIZATION HELEV ASSUMED KNOWNELEVATION IDEG.) 27ASPECT ASSUMED KNOWNMIN,MAX, INC ASPECT 40 50 5NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 15 15 15 15907 CI (030%) -/- 2.8% 2.87 2.8% 2.87 -.

". CLASS. FEATURES A A A A

.......-...... ..

2 FREQUENCIES

o - - FREQUENCIES

----- 8 FREQUENCIES

............... 12 FREQUENC IES

z ... ,, ..

- .'

L.L

C][: x) 'N-S.'

CL A

-5. 0. . . 15. 20.

POST-PROCESSING S/N IN 0BFigure A.14 Misclassification percentage versus post-processing SNR, -.

comparing the performance of 4 values of NF.

69

- ,-.-.-.... .. ,,. .. *- ' -:. * .. , -.'-" -"..-..:-: --- ..._..._ . ... - - . . . . - . - - . . • • - - - . - -, -. -

'p CLASSIFICATION OF GOD VCLS

POLARIZATION V/H [ELEV ASSUMED KNOWNELEVATION (DEC.) 27jASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 'YO 50 5

NO OF FREQUENCIES 2 '4 B 12NO O TARETS 5 is is 1

90%. CI (030%) */- 2.8% 2.8% 2.8X 2.8%

CLASS. FEATURES A A A A '..

o ________ 2 FREQUENCIES '

14 FREQUENCIES--------- 8 FREQUENCIES

......... 12 FREQUENCIES

w I .- . ....... -...-.-.......- ..- - --........

LDo

IL

cUU-6

- r. .

;_j(A

(1Y

07

CLASSIFICATION OF GNO YCLSPOLARIZATION vELEV ASSUMED K~NOWNELEVATION (BEG.) 27ASPECT ASSUMED K(NOWNMIN.MAX,INC ASPECT 80 90 10 BWNO OF FREQUENCIES 2 4 8 12NO OF TARGETS 10 10 10 10907 CI 1@307) +/-~ 3.Y% 3.Y7% 3.Y7% 3.Y4%

CLASS. FEATURES A A A A

.. .....C ..... ..........o; 2 FREQUENCIES

- --- - 0 FREQUENCIES--- S-- FREQUENCIES ~......... 12 FREQUENCIES ~

o0 CD

........ ....\.

...... ........ ... .......

CC In -6!I

-TO -5 0 5 0. 15. 20POTPOCSIGS/ NO

Figre .16 MiclasifcaionperenAg vessp t-rc sigSR

... .. .. ~ ....

- a..A - 5. 10. i. 20. ** * ..

L . . ~ .. ~ ',. ~ LI~'~ b 5 Y7~ ~ ~... ~ *... .... .!T . , T 7 R 7 7 u "°L- - -

CLASSIFICATION OF GND VCLS .POLARIZATION H , ,ELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN

MINMAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 10 10 10 1090. CI (0307.) 4/- 3.4% 3.47. 3.47 3.4%.CLASS. FEATURES A A A A

- " " : ........... ... . .T. . . ......... ....... .. ... . ... .. ...... .. . . . . . . . . . ... . .;L , '.

ON-2 FREQUENCIES4I FREQUENCIES

--------- FREQUENCIES................. 12 FREQUENCIES

LJ

a _ ..... . . -.. .,. .. .. - . . .. ... ....... ...zLU

:(: )\," -.-<

.- %-.. .- p|

,-.'".comparng.th pefrac o\4vle f F <;'

", ._ :A

0

"-.....2'* ,.

.S..'

*r .. ... .

-S. . 5. 10. 1. 20.POST-PROCESSING S/N I N 0OB

Figure A.17 Misclassification percentage versus post-processing SNR, =


72

.-. - .-....... . ... -. ., -...%" - " ' " - ' - : - " ' " . " " " ; ' . " " . . " " . " . - . ' , " , , ' ' , . '- , - . " " " , . • , , " . " . " " • . • • " . " : " • . ., ,

S - .••°

CLASSIFICATION OF GND VCLS %.. .

POLARIZATION V/HELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN

MINAMAX.INC ASPECT 60 90 10NO OF FREQUENCIES 2 '4 8 12NO OF TARGETS 10 10 10 10 y- -

90% CI 1030%) +/- 3.0q 3.4% 3.4% 3..4%,

CLASS. FEATURES A A A A

o " ........... ..... .......... .. .......... ....... ,. . .. .. .. . ... . . .. . . . .... ............ ..... ..... ..... .o c ~ .i " : : ' :I FREQUENCIESV... 2 FREQUENCIES

------- 6 FREQUENCIESC; ..................... . . ... .. .. . .. -.. ... .. . z u ~ z : . .- : -'... 12 FREQUENCIES ,..,, .

".. -\\---i-- -,------.- IR-C;

Lj . . ....... N.

.. ... .. ... . . . . . . .... ..

L.)

0

ze

.................. ..

CC C;.............- --- - --

i i I i I '. ' . -

-o.5. 10. 15. 20.POST-PROCESSING S/N IN OB

Figure A.18 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF.

73 "'"" ""

.~ -. , -

............. ''.,

CLASSIFICATION OF GO VCLSPOLARIZATION VELEV ASSUMED KNOWN

*ELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 4 12NO OF TARGETS 10 10 10 1090%. CI (030%.) +/- 3.47V 3.4%/ 3.47*/ 3.Y7A

CLASS. FEATURES A4W A4W A4W A4W

CD______ 2 FREQUENCIESI4 FREQUENCIES

--------- U FREQUENCIES

......... 12 FREQUENCIES%

z '

L a .. .. ... .......

ccwjCL. C..;_

o *

ED'.

. -

U\ X

50 5. 0 . 10. Is. 20.

POST-PROCESSING S/N IN OBFigure A.19 Misclassification percentage versus post-processing SNR,


74

CLASSIFICATION OF GND VCLSPOLARIZAT ION HELEV ASSUMED KNOWNt.ELEVATION (DEC.) 21ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 y 8 12NO OF TARGETS 10 10 10 1090 CI (0307) 4/- 3.YX 3.YX 3.'X 3.,

CLASS. FEATURES RAW A4W A4W A4W

C 2 FREQUENCIESY FREQUENCIES

- 12 FREQUENCIESC;_ ................. 12 FREQUENCIES .- .',-

IL3

................... ............... ....... ....... ........ .... .."....... ... .. ..... .. ... ........... .... ... .....

L)

o i i ',.", i

cc

LLI

a- .. .....\.. ., .. ""-" -

% 6.* %

-TO'. -5. O.-5.\10 15. - 20 --- --

o mp-5..0 . 5.. . 10. 15..20..--

Cr.5

.... ... ..... ......N.,-

75 :: . '

f,5,

*.-.' J j ,..-. .~. .'* -- .r ... # ,,.- .. . . .- ... '-K -%, ,- -_ . ,

1w


POLARIZATION V/H

ELEV ASSUMED KNOWNELEVRIION (DEG.) 27

ASPECT ASSUMED KNOWNMIN,MAX,INC ASPECT 0 10 10

NO OF FREQUENCIES 2 y 8 12NO OF TARGETS 10 10 10 10

90% CI (030X,) +/- 3.Lj% 3.Y% 3.*%, 3.Y7

CLASS. FEATURES SAW A&W AW A4W

O ________ 2 FREQUENCIES" FREQUENCIES C:, t

---------- .I FREQUENCIES................. 12 FREQUENCIES

L ,

"JL ; ... . .... ....... ........... ... . ....

U

S....,cc

Z . ' *LL~J0 - .. ... ... .....

0

z(D

a:~~~~ .......... 3. ,..

o ! ,..-\,

- . '-c,. N" 4, :, ;

._) ,, , . .\.U

-0 5S. 10. 15. 20.

POST-PROCESSING SIN IN DB

Figure A.21 Misclassification percentage versus post-processing SNR, ".. -.comparing the performance of 4 values of NF. ,'

76"

ze,.=,

-=' ', .-

-...... ..;.,. .:,.-,:..,'; ..... ,.-. ...-.... .., .. ...,.;, -, ... .. ,:.. .-. .... .....-. : .- . .': > -.. .N. ...5

CLSIIAIN FGOVLPOLARIZAT ION

CLEASSUMETDO OFKNOWNLt:.ELEVATION (DEG.) 27

MIN.MAX. INC ASPECT '40 50 5NO OF FREQUENCIES 2 '1 a 12NO OF TARGETS 15 15 15 15 i.

90% CI (930%) +/- 2.8% 2.8% 2.8% 2.8% A

CLASS. FEATURES AAN A(N A4H A4N b

C;-~~-P - ----

2 FREQUENCIES- £1 FREQUENCIES

.- ..... S...... - - - FREQUENCIES.. .......... 12 FREQUENCIES

a: w

z .

... .......------.--

- .... ....... . ......

........ ... - %4

5 .....~ ~ s N .............. ....... .. i.-?. 20. ..... -------

77 ......

.. .. .. .. ....

-5. 0. S. 10. 15. 20

POSTPROCSSIN S/NIN1DFiueA2 icasfcto ecetg essps-rcsigSRcomparing~~~~~r th eromac f ale o F

w -3


POLARIZATION H

ELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMEL KNOWNMIN.MAX.INC ASPECT 40 50 5NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 15 15 15 15 * ,,=907 CI (0307) +/- 2.87 2.87 2.8% 2.87CLASS. FEATURES ARW AW AW ARW

2 FREQUENCIES4I FREQUENCIES

--------- 8 FREQUENCIES- ...... .. ......... 12 FREQUENC IES

I ':: " -. \- \ ' - --- - - -LLJ

0! Ao,.. ..,-:...N.4 .a...,.. ...\

o"- ' .:" \ \ U-

:..:,. \ \,' . .... , ..

0

,:; .. . .. . .. . . . .

-. r

U

_ - ... ;.. :-..ss .- .. ...- .- , .

-io, -5. 0. 5. 10. IS. 20.POST-PROCESSING S/N IN OB


78

.,o ~~*..-.,- . -. h. .. o *

-. POLARIZATION V/HELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN

MIN.MAX,INC ASPECT YO0 50 5NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 15 15 15 15

907 CI (0307) 4/- 2.87 2.87 2.87 2.87CLASS. FEATURES AAW AW AW RW

.~~~ ~ ~ ..% . ........ ....

2 FREQUENCIES

I FREQUENCIES--------- 8 FREQUENCIES

................. 12 FREQUENCIES

LU . .............. ...

- ... -. ...:..

- I .. 0"S" S p .. """"1'

LL .

L)J

...... . - - .. ... * --. ... . .

POST-PROCESSING S/N IN DB I

"C U %%'1

. ~Figure A,24 Misclassification percentage versus post-processing SNR, ... 'comparing the performance of 4 values of NF. .... -

79 ...... .... ..... ...

. ' 1

....... . .. ........... ... . ....... . ........

_ o~~~ -5 0.%1. 15"0

• v . "" ". ",.:" ; "€ .,'".".,. .- '"-"- . ," "-" R C SS N •'', .,''-. IN'''--- .DB'''-.., ,-.-." ,,- 2"Lq = ." ' _,-'',," .- " "

CLASSIFICATION OF GO VCLSPOLARIZATION VELEV ASSUMED KNOWNELEVATION (DEG.) 27

ASPECT ASSUMED KNOWNMIN.MAX,INC ASPECT 80 90 1aNO OF FREQUENCIES 2 ' 8 12 ,,NO OF TARGETS 10 10 10 10907. CI (0307.) +/- 3.Y7. 3.Y% 3.W/% 3.Y7. v- '

CLASS. FEATURES ARW R4W A&W RAW

. - . ............... ...................................... .FRE...N.......-0 _________ 2 FREQUENCIESYI FREQUENCIES

--------- FREQUENCIESo. . 12 FREQUENCIES

U j ---------... -...- .....

z .... .. .~ ...... ........... .

a• -" -'"

z

. % \

LL: -

_j-

U.)

"-. .. \ ::'::

eu0 0. 15. 2.0\

, ..... S - ... -- * " 'Z:.*"" ,, S.. ,..,* -

-?o . -5 . 0. 5 . 10. 15. 20 ."- , .


Figure A.25 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF. -

80 ..- . - -

N,.., . , ..-

p %7

CLASSIFICATION OF GND VCLSPOLARIZATION H ...

ELEY ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 y 8 12NO OF TARGETS 10 10 10 10907 CI (030.1 +/- 3.47 3.47 3.47/ 3.0CLASS. FEATURES SAW A AW RAW AW

FREQUENCIES4FREQUENCIES

------ 2 FREQUENCIES

IL .. .. I..

U~ FN

ZL --- ---

o

CC ..~ .\.. ..

.. . .. ... .. ....

C-

-5 . 5. 10. 15. 20.POST-PROCESSING S/N IN 08


81

. . . .

p .." I,' -

CLASSIFICATION OF GND VCLSPOLARIZATION V/H-ELEV ASSUMED KNOWN

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMINMAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 1 8 12NO OF TARGETS 10 10 10 10907 CI (0307) */- 3.YX 3.Y. 3. Y7 3.47CLASS. FEATURES RAW RAW RAW RAW -

";2 FREQUENCIES4 FREQUENCIES

-------- ... FREQUENCIES........ 12 FREQUENCIES

L 1 _ . -.. ... . . . . ... ....-

D

.-_ " . , . " - : ..-

• o°o

or) 4' " \..""

. . ,,4'\ . .-

a- -5. 5. O. . 0

crr

o 44

POST-PROCESSING S/N IN OB-"; Figure A.27 Misclassification percentage versus post-processing SNR,

c-ur

10. 82 20.

POST-PROCESSING . .. IN ,.

Figur A.7Mslsiiaio ecnaevru ps-rcsigSRi ~ ~ ~ ~ ~~oprn the performanc of- 4" value of NF."°".°" '

". % "- % "- ". ". ". " " % " '. % .' %

CLASSIFICATION OF GNO VCLSPOLARIZATION VELEV ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 0 10 10NO OF FREQUENCIES 2 4 a 12NO OF TARGETS 10 10 10 10907 CI (o30%.) 4/- 3.11% 3.47/ 3.Y%. 3.4%.CLASS. FEATURES W W W W

2 FREQUENCIES'I FREQUENCIES

--------- FREQUENCIES

o. ......-.. 12 FREQUENCIES

J

I.LJ

4~V.

(A L

LL- C;

0

-?. -. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN OB


83

A-A163 026 GROUND VEHICLE CLASSIFICATION USING NULTIFREGUENCY 21NULTIPOLARIZATION RESONANCE RADRCU) OHIO STATE UNIVCOLUMBUS ELECTROSCIENCE LAB N CHAMBERLAIN JUL 85

UNCLASSIFIED ESL-714196-18 NSSSI4-92-K-1637 F/G 1719 N

1-01

1-26I 11111 = 11.

NATIONAL PURMA OF STANDARDSSWfOCOPY ufSCUTC TEST CWAT

% %4

CLASSIFI]CAT'ION OF GNO VCLS

EEV ASSUMED KNOWNELEVATION (DEG. 2NO OF, EOUENCIES 2 8 1NO OF TARGETS 10 10 10 10 -

CLASS. FEATURES W W W W ! '''

0~~- -... FREQUENCIESr ! --.......... FREQ.UNCES

. . .. .. .. i .. .. [ . ... .... " ............ 12 FREQUENCIES

- --- ----- --- - - -

, : ; .. ... ..........

),. ..,C. C

POLE AST-DPROENG S/ NDELEU iONa(DEC.) 2eSPcmaSSnE tporn

LLiJ

8S 10"-0

CLASS.~ FETUE ........ .

-~..-- . . ..(2FREUENIES~; ....

LIJ* .,..- -- - - -- --- ---

_O. .5 0 . 1. 5 0

o".SN°N

Figure-J .. 29 M l io -,t ers. . p r s S,

-coprn th efomneof4vleso F

L84

CLASSIFICATION OF GO VCLSPOLARIZATION V/HELEY ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX, INC ASPECT 0 10 10NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 10 10 10 10 . .

907 CI (9307) +/- 3.47/ 3.4% 3.UVY 3.47VVCLASS. FEATURES H H H H

-; 2 FREQUENCIESII FREQUENCIES

------ FREQUENCIES *

o;......... 12 FREQUENCIES .**

z .... .-*-*7----- p -HI.. .... ...

z

NN i

CC)

..c ..... .. ....--- --- --

: ;-?. -5. 5. . 10. 15. 20.

POST-PROCESSING S/N IN 05Figure A.30 Misclassification percentage versus post-processing SNR,

* comparing the performance of 4 values of NF.

85


POLARIZATION VELEV ASSUMED KNOWN

ELEVATION IDEG.) 27ASPECT ASSUMED KNOWNMIN,AX.INC ASPECT' O SO 5

NO OF FREQUENCIES 2 8 12

NO OF TARGETS 15 15 15 1590% CI 10307) 4/- 2.8% 2.8% 2.8% 2.8%

CLASS. FEATURES 1 1 1 N""

. . . . . . . . . . . . 2 F R E Q U E N C I E S

SII FREQUENCIES- FREQUENCIES

D ..... 12 FREQUENCIES

LL J . ... . ..... .. . - .. ....... . ..' .... ... ... . ... .... ... i '

_ ,-. -

" €c; \ " .... 1

z(L.J K.

I.- \ .. ,'.:-"

a- .

•LL N *

-...

-(A -s. s. o. *s o

PN /_j

-?. -5. 0. S. 10. 15. 20.POST-PROCESSING S/N I N OB


86

* " S.O o°

CLASSIFICATION OF GND VCLSPOLARIZATION HELEY ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT YJO 50 5

NO OF FREQUENCIES 2 '4 8 12 .NO OF TARGETS 15 is is 1590Z CI (0307) */- 2.6X 2.8% 2.8% 2.8%

CLASS. FEATURES W W W W

g 2 FREQUENCIES4 FREQUENCIES

-_--------- I FREQUENCIES..

...... 12 FREQUENCIES

z

LI\

C ; .....

-------.-.---

. . . .... ............... . ..... . . . . . . . .....

CEo ...-... 0. .. 0..1...20POTPRCSSN S/V I

Figue A32 iscassiicaionperentae vrsu pot-prcesingSNRcoprnLh)efraneo auso F

08

C;2

UkA

CLASSIFICA71ON OF GND VCLSPOLARIZATION V/HELEY ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 40 50 5NO OF FREQUENCIES 2 '4 8 12NO OF TARGETS is is 15 1590%. CI (03OV) +/- 2.87 2.87 2.8X 2.87CLASS. FEATURES I W W W r

- - - 2 FREQUENCIES

4 FREQUENCIES--------- S FREQUENCIES t.......... 12 FREQUENCIES

L*LJ. ....... r........Cr CDC ---- -----

z

0-S

C;5

oz ... *S-

C;"

0

CcL)

-?. -5. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN OB .. *~v

*Figure A.33 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF.

88 -

CLASSIFICATION OF GND VCLSPOLARIZATION VELEV ASSUMED KNOWNELEVATION IDEG.) 27ASPECT ASSUMED KNOWN?IN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 41 8 12NO OF TARGETS 10 10 10 10SO0. CI (0307.) */- 3.4%i 3.4%7 3.IJ% 3.11.Z.V

CLASS. FEATURES H H W H V

....... . ....

g 2 FREQUENCIES4FREQUENCIES

---------- 0 FREQUENCIES

.......... 1 FREQUENCIES

wj ... ......... . .

C 0

CC).

C;

-5 0. 00 15 20

0TPRCSIN / IND

FiueA.4 MslasfctinprenaevesspstpoesigSR

comprin theperormnce f 4valus o NF

89b

CLASSIFICATION OF GND VCLSPOLARIZATION M

ELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10

NO OF FREQUENCIES 2 y 8 12 . .,-NO OF TARGETS 10 10 10 10907. CI (030Y) */- 3.47 3.47 3.4. 3.47CLASS. FEATURES W W 1 W

C 2 FREQUENCIES- 'I FREQUENCIES -.

6 FREQUENCIES................. 12 FREQUENCIES ,

CCoL ) ... ..

0-

-C;

a: n

E Z) I-. *-

C;

- -5. 0. . 20.

POST-PROCESS ING S/N I N DOBFigure A.35 Misclassification percentage versus post-processing SNR,


90

CLASSIFICATION OF GND VCLSPOLARIZATION V/HELEV ASSUMED KNOWNVELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 so 10NO OF FREQUENCIES 2 y 8 12NO OF TARGETS 10 10 10 1090%. CI (0307) 4/- 3.4% 3.4% 3.4%i 3.47.CLASS. FEATURES W W W W

____ 2 FRNEQUENCIES

0 C

CC.0

.. '

CLn

-. LL I II

-TO. -5 0. 5 10. 15. 20.POST-PROCESSING S/N IN OB

Figure A.36 Misclassification percentage versus post-processing StIR,.1 comparing the performance of 4 values of NF.

91

+6X

CLASSIFICAION OF GND VCLSPOLARIZATION V i,*

ELEV ASSUMED KNOWN L .

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWIN

MIN.MRX.INC ASPECT 0 10 10NO OF FREOUENCIES 2 4 8 12NO OF TARGETS 10 10 10 1090%. CI (o30%.J +/- 3.14% 3.4J% 3.4%/ 3.147

CLASS. FEATURES T T T T

2 FREQUENCIES'4 FREQUENCIES

--------- FREQUENCIES- ................ 12 FREQUENCIES

LAJ,, ; . +-. I - ; . II.... -... ..... ...... ... .. . +. . . . .... . . + , - ....

,.- . \. .\ UZ(D

-,+"~ . " . \ . -::

, - . .. " \ ....rU

u--

-toa. -5. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN 0B

Figure A.37 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF. .....

...... ..... ...... ..... ...... ..... .....92 ,+-..-.".-." •

• .- . %" -*-

" -V. Sb .2b . S I = ' . .2 k1 V- q= %.V),l q. ' . ,- ( . L.C A<'P ..W --..7......W.. J. -- c rz .r.rtrn m a - --.. . . a . --=-i.t.T flr

Vk

CLASSIFICATION OF GNO VCLSPOLARIZATION HELEV ASSUMED KNOWN

ELEVATION IDEG.) 27ASPECT ASSUMED KNOWN

MIN,MAXINC ASPECT 0 10 10NO OF FREQUENCIES 2 '4 8 12NO OF TARGETS 10 10 10 10 . .

907 , CI ( 0 ) 4 )/- 3.47 , 3 ,7, 3.Y7 , 3.YC,

, -CLASS. FEATURES T T T T

6 . 2 F R E Q U E N C I E S4 FREQUENCIES

- - 18 FREQUENCIES

D *

u." .- ---... --.-.

,...

z

0- C;A

Z -. .,

U ".. \ S

-) C_ -

. \

), o_ ). - ...... ....

. . . ....

-"-- . : ,,-,

01. .~ ,, \.,

.-

\-.

-70. -5 O 5. 10. 15. 20.POST-PROCESSING S/N IN OB

Figure A.38 Misclassification percentage versus post-processing SNR,comparing the performance of 4 values of NF. .-.

S3 ; ' ''

% I;.r. . ..-...,., . I ,%., .. .., , ,., ...-........

..... .,..., .,...-.,.... ..... ... . . . ."-., ... ,... . ..........

.. .,-.,.,: , ...

'.'.


POLARIZATION V/HELEV ASSUMED KNOWN ,. -ELEVATION IDEG.) 27ASPECT ASSUMED KNOWNMIN,MAX.INC ASPECT 0 10 10 PNO OF FREQUENCIES 2 y 8 12NO OF TARGETS 10 10 10 10

907 CI 10307)a4/ 3.4%/ 3.4% 3.4%/ 3.4%XCLASS. FEATURES T T T T

2 FREOUENCIES4 FREOUENCIES.. ... .. ... . - .. FREQUENCIES

.. ................ 12 FREQUENCIES

Ii Iz

L.. , \\. -.LLi

.... ,, \\: :--

0 - ,. , . ..\ .. .. -. ...... ., ? .' -'

0

a:~, o: , ' .. .... ':;

0 -0. is. 20.-

T~~~~~7 7,". ,, \ ,

94 .

ii'-o" -5. o. 5. 10. 15. 20. <---. POST-PROCESSING SIN IN OB

. i Figure A.39 Misclassification percentage versus post-processing SNR, -...- 'comparing the performance of 4 values of NF. ""

i:". ~94 "-??:

. . .'af.

.........-*.**.*..,-** .-- ,..,.,.... .. . .. .

A X. tt. C t.-! 5 - .-. Sa.- P A S !.S- k!i

CLASSIFICA7ION OF GO VCLS%POLARIZATION V

ELEV ASSUMED KNOWN

ELEVATION [DEG.) 27 :z I

ASPECT ASSUMED KNOWN

MIN.MAX, INC ASPECT 40 50 5 __

NO OF FREQUENCIES 2 4 8 12

NO OF TARGETS 15 i5 15 1s907 CI 1@307) /- 2.67. 2.87 2.8 2.8X

CLASS. FEATURES T T T

o ________ 2 FREQUENCIES'I FREQUENCIES

--------- S FREQUENCIES

... .......... 12 FREQUENCIES

Li -.- /. - -4 - ''""

Z -LLJ 0

a:

-

Ca c;-A I :: :. , ", I \ I :: :-

-'O. -5. 0. 5. 10. 15. 20.POST-PROCESSING 5/N IN DB


I -:

', : .::..95

"" ":"::"

E.** :* i.. .. ...............................................

. . . . . . . . . . . ......._:- . ..:. . . ..::. .:. . . . . . ..:. .......-... p:. .:... .........-... :: ,.::;,:....::.:.;:;: .:;-.:..,:- - .-. ..:.,,,


POLARIZATION H.ELEV ASSUMED KNOWN LELEVATION IDEC.) 27ASPECT ASSUMED KNOWN

MIN.MAX.INC ASPECT '40 50 5

NO OF FREQUENCIES 2 '4 a 12 ~NO OF TARGETS 15 15 15 1590/. CI (e30%) 4/- 2.8% 2.8% 2.87. 2.87 .. .CLASS. FEATURES T T T T

'I FREQUENCIES-- - 8 FREQUENCIES............... 1 FREQUENCIES

-- . . . .. ... ... ... ...... .. .. ... . ......... . . 2 F E U N I S, ' .-w. - -.. -''

,,..-\ •\'zLaJ

z to,.- ,, \:::.::

... \ \ ,,-

U

LL. " ",C\\ . ..L')

_-~

-,u. -5. o. . 1o. 15. 20.POST-PROCESSING SIN IN DB6,

Figure A.41 Misclassification percentage versus post-processing SNR,-.....comparing the performance of 4 values of NF.

EQ ...

"96 0 '°..',.'

. . . . . .. . . . . .01 . .

.. ... . . . . . . . . ..* .** . *-**.* * : .

:,.- 9. . -. ,-..- -,... -. - - -....- , . , . . ., -.-.. .. . .. ,..... . .* .. -,,'..' ..- ' ,..-,-. . .,.- -< ,-,-.-.

CLASSIFICATION OF GNO VCLSPOLARIZATION V/HELEV ASSUMED KNOWN "ELEVATION (DEC.) 27ASPECT ASSUMED KNOWNMIN,NAX,INC ASPECT 40 O 5NO OF FREQUENCIES 2 Li a 12NO OF TARGETS 1s is 15 is90% CI (0307) +/- 2.87 2.87 2.87 2.8%CLASS. FEATURES T T T T

". ... ... .... .. . ............. ..... . . . . . . . . ... .. . . . .. . . . . . .... W 6

C; I 2 FREQUENCIES- FREQUENCIES:

---------- 8 FREQUENCIES. ................ 12 FREQUENCIES*

::- ...... C'" "".w" . ..... .. . .........

r .. '.C. : - . ' i

_~ ... . . ... ... . ...... ...., ... .... .. ...... ........

a-= .. ,a

C D" ",Za

_ .... . ..-. " ".. ..

-?o -..-.. 5 10 15

PO5T-PRiOCE5SING S/N IN DB P-"

,. ~Figure A.42 Misclassification percentage versus post-processing SNR, '.,.'" ~~comparing the performance of 4 Values of NF. """

_97

V -

. . . . . -". .

. ....... . . .......

"_2 --2 _2---5 .2 2 -2 -,.2:.2 - O. -5'22 22. .-0-..2-.- : . .. '. 1"0".'£ 15:2: ;2 . 20 ._. 2 2 22 '" ': 2 2 :

CLASSIFICATION OF GNO VCLSPOLARIZATION V

ELEV ASSUMED KNOWNELEVATION (DEG.) 27

ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 4 a 12

NO OF TARGETS 10 10 10 10907 CI (e30.) 4/- 3.YV 3.47/. 3.YV. 3.Y% r. ,CLASS. FEATURES T T T T 1%

; 2 FREOUENCIESII FREOUENCIES

. - FREQUENCIES,. .............. .... REQUENCIES

-" ..

Cc)

z

L

9-",, •

•J* . ,.4\..... .

zo

-s. o. 5. 10. 15. 2o.--LPOST-PROCESSING S/N IN 0B .. ,-

Figure A.43 Misclassification percentage versus post-processing SNR, ,comparing the performance of 4 values of NF.

98

.4..- °U*.4 ".

*Ib% , ''. * * * w .--o".. " ;r. . r~lrg- °*. . -1 =.-i I .. " z rr. *° r;--, -.. r.'- cs .": J .., . ,: .. ,. .- 1., ,%,. rl[ _.rl.:

CLASSIFICATION OF GNO VCLSPOLARIZATION MELEV ASSUMED KNOWNELEVATON IDEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 2 4 6 12NO OF TARGETS 10 10 10 1090%. C I 30.) 4/- 3.47. 3.Y% 3.0 3..%CLASS. FEATURES T 7 T T

C; 2 FREQUENCIESI V FREQUENCIES

--------- S FREQUENCIES~2 . .......... 2 FREQUENCIES -..... ..... .... ...... c ..

f:K z

(A

-.. 4--,i

I*

CC V

-?o -5. . 5. 1. 15\0C C .. ...... --

_ . -5 0. '. 10 5 0

POST-PROCESSING S/N IN DBI.

t.[..iueA4 icasfcto percentage versus post-processing SNR, ...comparing the performance of 4 values of NF. .."

ILL"-

CLASSIFICATION OF GND VCLSPOLARIZATION V/HELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN,MAX,INC ASPECT 80 90 10NO OF FREQUENCIES 2 4 8 12NO OF TARGETS 10 10 10 10

90% CI 1030%) */- 3.YJ% 3.YJ% 3.Yd% 3.4rACLASS. FEATURES T T T T

_ L42 FREQUENCIES

SI FREQUENCIES.............. ......... .... ...... FREQUENCIES I-. '

0 ................. 12 FREQUENCIES

-. . ..... ....... ........ ..

z ' . ... ....... ........ ..... ... ..... .. ..... ' : ':

Up.,.r'. ',

ILJ

0zw

(•).', \.. .-

• , .. %.

C-

... .* . .. , . .\ * .9= . .

__=

-,o. -5. 0. 5. 10. is. 20.POST-PROCESSING S/N IN DB


100 ..

CLASSIFICATION OF GNO VCLSpPOLARIZATION v M V/H

NIN.MAX. INC ASPECT 0 10 10NO OF FREQUENCIES 8NO OF TARGETS 10 10 1090%. CI 1630.) +/- 3.4%7 3.4Z% 3.4%7

CLASS. FEATURES A A A -

I _________ VER7ICAL POLARIZATION- -- - HORIZONTAL POLARIZATION

------_- ---- V/H POLARIZATION

C;_S

LA~cA

La.I

C;- -

U

y1o. -5. 0. 5. 10. 15. 20.POST-PRIOCESSING S/N IN 05

Figure A.46 Misclassification percentage versus post-processing SNR, -.

comparing the performance of various polarizations.

101 ':.

CLASSIFICATION OF GO VCLSPOLARIZATION V H V/HELEV ASSUMED KNOWNELEVATION IDEG.) 2?ASPECT ASSUMED KNOWNMIN,MAX. INC ASPECT 40 50 5NO OF FREQUENCIES aNO OF TARGETS is 15 15SO0. CI (030%) +/- 2.8% 2..8X. 2.8%

CLASS. FEATURES A A A-----------------------------------.---.--------- 1--7

C2 VERTICAL POLARIZATION- HORIZONTAL POLARIZATION

--------- V/N POLARIZATION

-L - ---- .... ......

LUCC .

aj(0Z

CC . ...

copain th efrac'fvrosplrztos

P _q

CLASSIFICATION OF GNO VCLSPOLARIZATION V H V/H

ELEV ASSUMED KNOWNELEVATION IOEG.3 27 -,ASPECT ASSUMED KNOWN

NIN.MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 10 1090. CI 1e30Y. 4/- 3.YX 3.Y. 3.Y7. ;CLASS. FEATURES A A A

o | VERTICAL POLARIZATION ". =".'

HORIZONTAL POLARIZATION--------.... V/ POLARIZATION C

......~ ~~~ ~ ~~~~ ........... ........ .... ............ ........CC.3i i %

-' /II \ , iI S'-% 'Lig

4- - -- -] ii h _

\.I *i " . .-. ;

L ,

Cr... + -

.. .. .....LA_ C;

,% %

.. .... -.. .. . .. . - .- -.. - --- .-..-

-- ?. -. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN 05

Figure A.48 Misclassification percentage versus post-processing SNR,comparing the performance of various polarizations.

103

b,

SCLASSIFICATION OF GNO VCLSSELEV RSSUMED KNOWN

i ELEVATI]ON IDEG. ) 27

IASPECT ASSUNEO KNOWN

NIN.MAXINC ASPECT 0 10 10.NO OF FREQUENCIES 8

NO OF TARGETS 10 10 10

907 CI (430%) +/- 3.4% 3.47. 3.Y7.

CLASS. FEATURES A4W A4W A4H

VERTICAL POLARIZATIONSi . . HORIZONTAL POLARIZATION

-------.- V/H POLARIZATION

--

.-

- z /, I IU0 6-. zc I**.. . N . . . .

"- -, \x .,. ...o -

LiJ

r - .\. . .. .. .-

., 0- ", \ .... -

". I :/I % I I I

-. . -/ 0. 5. \0 15 0

-POT-PROCESSING S- N -N ..Figure A.4c M ificatio p v p '."

cpnhercfrslzo

_ * -T.-. . 5 IO. 15. 20~ :.:. K K K §{ K2'-

' P ST PRO ES ING SI I DB'-- "-''--. Figure A.49 His~c; asfcto ecnaevru ps-rcsigSR -'"""

.c -.-._-j' -

.. ,. " o " " ; .r . o "P - " " , . P ' " " -". " : . . " '' ' "' ' '' ' ' - ' ' , "U" j"" ".'

: %'', ~ ~ ~ ~ 3 M..,., % .. •.. .... % ... ,.. '.... .,,.% ..... %,..%.%...-.. .. ,. _ .,_¢z_,_ ;_.; ,.crrL¢ ,.-,.__,.,: .#_; .,. ,.-. . ._'.a.r L ., _ . . ._"... . _. ... - ,, ' C,, .,.

TO -5 1 5 1 1. 1 s 20

- -,. - -",.- -L A

CLASSIFICATION OF GND VCLSPOLARIZATION V H V/H

ELEV ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWNNINMRXINC ASPECT 4O 50 5

NO OF FREQUENCIES eNO OF TARGETS 15 15 1590%. CI (0307.) */- 2.87. 2.87. 2.87.CLASS. FEATURES A4W A4W RAW

C;- VERTICAL POLARIZATION- - -.... HORIZONTAL POLARIZATION

- --------- V/H POLARIZATION

! . I ' "

L.)

LIJi \ ,r,, . ,

.. ... ... T .. ....... . .--.*--'-- I 'S - i. 1

0 . '*... I'---.-"-

-105-.---.----.,

I .' . '°

1 1 1 I*, I" °

-?o. -5. 0. 5. .0. 15. 20

0105

..................................Z.. .

* . .. . . . . . . . . . . . . . . . * ** ~ ~ ~ -..

L). .*-.- .- ~sp.p-

CLASSIFICATION OF GND VCLSPOLARIZATION V MI V/MELEV ASSUMED KNOWNELEVATION IDEG.) 27ASPECT ASSUMED KNOWNMIN.MAX.INC ASPECT 80 90 10NO OIF FREQUENCIES 8NO OF TARGETS 10 10 1090OX CI (6301 -+/- 3*I4Y/ 3.47 3.4%7 -

CLASS,~ FEATURES 4W R4W R(W

aVERTICAL POLARIZATION- -- - HORIZONTAL POLARIZATION

- --------- V/H POLARIZATION

0

LU - . . ....-

o; N

_j 0

En~S%C;

-?. -. 0. 5. 10. 15. 20.POST-PROCESSING S/N I N OB *


106

CLASSIFICATION OF ONO VCLS

POLARIZATION V H VIMELEV ASSUMED KNOWNELEVATION IDEG.) 27ASPECT ASSUMED KNOWN .e'eMINMAX. INC ASPECT 0 10 10NO OF FREQUENCIES 8NO OF TARGETS 10 10 10907 CI 14307) */- 3.Y% 3.0rI 3.Y7CLASS. FEATURES W W N

a VERTICAL POLARIZATION- I HORIZONTAL POLARIZATION

-. ----- V/H POLARIZATION

iii

IDI

I.- C; T

ccI If

U -.-.- .I---~

oC;

(nK

I - - -

Li

- 0TPOESN 4.... IND

Fiur A.5 Micasfcto ecnaeIessps-rcsigScomparin th pefrac o vaiosplrzt n.

-. ... -.. 2.107-. .. .

CLASSIFICATION OF GOD VCLSPOLARIZATION V H V/HELEY ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWNMIN.MAX,INC ASPECT YO0 So 5NO OF FREQUENCIES 8NO OF TARGETS 15 15 15

90%CI160/) /~2.8% 2.8% 2.87CLASS. FEATURES W W W

C__ VETCLPLRZTO

HORIZONTAL POLARIZATION--------- V/H POLARIZATION

LiiI

S..

C;,

IN,

U,

U

U

-?. 5. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN OB

Figure A.53 Misclassification percentage versus post-processing SNR,comparing the performance of various polarizations. *

t3

108


POLAR1]ATION V H V/HELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWNMIN,MAX.INC ASPECT 80 90 10NO OF FREQUENCIES 8

NO OF TARGETS 10 10 1090X CI 1030%) */- 3.J% 3.4 3.U7

CLASS. FEATURES W W.

o jVERTICAL POLARIZATION- - HORIZONTAL POLARIZATION- ------- V/H POLARIZATION

,--:®-o <::- I - . ! I 1 ..

z - ",hu. 1.

• o - ,,. .

" \. - i, j .'-]

.~~0.......... .......... ... ... ... .,

-"z oiI , I\" ..

( . - . . . . . . . ... . . ... .. .. .I'

LL

-- __ _ -- i---~ r

(INS.

U0N"

• i" igreA,4 islasi ic inpretg ves s .s -rcesn.NR.-i .

CC%

L - .

-

• *.-I% %

POTPOCSIGSE NO

109

C.5

-5. 0..5.:.0--.,20

, ir

CLASSIFICATION OF GO VCLSPOLARIZATION V H V/H

ELEV ASSUMED K(NOWN ~ELEVATION (DEC.] 27 ~ASPECT ASSUMED KNOWNMIN.MAX. INC ASPECT 0 10 10NO OF FREQUENCIES 8NO OF TARGETS 10 10 10

907~ CI (0307.) +/- 3.4% 3.4X~ 3.'4%

CLRSS. FEATURES I T T

o j VERTICAL POLARIZATION- - - H ORIZON'TAL POLARIZATION

------ V/H POLARIZA TION

Lii

zrO

Z~0j

U(0 A

- .... V..\

x~ CD

-5. 0. 5. 10. is. 20.

POST-PROCESSING S/N IN OB

Figure A.55 Misclassification percentage versus post-processing SNR,comparing the performance of various polarizations.% 6

110

CLASSIFICATION OF GND VCLSPOLARIZATION V H V/H

" ELEV ASSUMED KNOWN

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWNNNRX, ]NC ASPECT 40 50 5 m .

NO OF FREQUENCIES 8NO OF TARGETS 15 15 15

90% CI f@30%) +/- 2.87 2.8 2.8,CLASS. FEATURES T T T

a I VERTICAL POLARIZATION- - - NORIZONTAL POLARIZATION

--------- V/N POLARIZATION

. . .. --__. _ __ _ _ .. ___ II*. .

-

LLJ~ Q

I " .. ...... i .... ,, ..... .. ......... U ......... ...........

-" .o - . . . ,. 2 .-

I. 1-D-

copr 9 h efrmneo arosplar atn. .'-rC- .1 1 .1% -

C; ----- i1

rn-i- ------- _---------..- ---

-TO. -5. .5. . 15. 20.POST-PROCESSING S/N I N DB


CLASSIFICATION OF GND VCLS" POLARIZATION V H V/H

ELEV ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWN

MIN.MAX,INC ASPECT 80 90 10 %,NO OF FREQUENCIES 8NO OF TARGETS 10 10 1090% CI (630%) */- 3.4% 3.4% 3.,%


C; VERTICAL POLARIZATION- - - - ORIZONTAL POLARIZATION

. V/H POLARIZATION

L), - 4.. . !;: .--....

a.,L.)

S.......... .

LL. -V o .-

Li I' I -o°-

CI

r-7

I p t t

I \ N.'.t

-" I , I 9.

______- __ .- _________ .

I ! \ A-"--

-?o. -5. 0. 5. 10. 5. 20.POST -PROCESSING S/N IN DB"."0

Figure A.57 Misclassification percentage versus post-processing SNR, """'comparing the performance of various polarizations. '..-

*f'.--

°- .- ,

* . . . . . . .. ..o. . -. .- , - - ~ - . ° - ... . . . . . . . . . . . . . . ° .•. -°. o . •-. - .-

PR ". -"A-.%

CLASSIFICATION OF GNO VCLS ,,S.

POLARIZATION VELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWNMINNAX.INC ASPECT 0 10 10. 40 50 5, 60 90 10

NO OF FREQUENCIES 8NO OF TARGETS 10 15 1o907. CI 19307l 4/- 3.4%. 2.8Z. 3.4% '"

CLASS. FEATURES A A A

- - -S oEG. ASPECT ZONE,---------- 90 DEG. ASPECT ZONE.

CIC

z . It- ..r

. -; . .... . ... ... . . . ... .... .. ....... . .. ... . ...... ... . .

, * ii --

I-a - I I*,'.-'..\.

_-- _ -. o s. Io -A. 'o -.- ,-S.

,*5'... ..:

LnJ

............

C;

.,

- o. -5. 0. 5. 10. .s 20.POST-PROCESSING S/N IN 5B

Figure A.58 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones.

* 113

CLRSSIFICRTION OF GND VCLS

POLARIZATION H - :

ELEV ASSUMED KNOWN

ELEVATION (DEG.) 27

ASPECT ASSUMED KNOWN /KNOWN

MINMAX. INC ASPECT 0 10 10, 'J0 50 5. 8o go o0NO OF FREQUENCIES 8NO OF TARGETS 10 15 10

907. C (0307) #/- 3.J7 2.87 3.47 . .


.. 71

* ".. ....... .......... .......... ........ .. ........ ....... ....- " -.... .. ..... ... .--- I

. __-_T__ 0 DEC. ASPECT ZONE.

- i -- 4S DEG. ASPECT ZONES------ 90 OEG. ASPECT ZONE

.. . . . .. .. .

I- I _ '_

I s ,%'

• ,-- -~ ,\ . . . . . . -. .-'-- - -°

-- P TPRCSSN S/N IN D

* ' ,' * 5'

(') L - A

I V

0. .. ...0..

-... ,.

ooto

. . . . . -.. ... . S. .... .5 .- *.-0 ... .

POST-PROCESSING.....I....

... copain the* pefrac of vaiu .*-- , . ~ .a.sec zoes

CLASSIFICATION OF GN0 VCLS

POLRRZR'ITON V/H

ELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWNMIN,MAX,INC ASPECT 0 10 10. YO 50 5, 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 15 1090. C] 1030.) */- 3.YZ. 2.8Z 3.,.CLASS. FEATURES A A A

a 0 DEC. ASPECT ZONE-- - - - - MS DEG. ASPECT ZONE,----- 90 DEG. ASPECT ZONE'

,.;'.22,,2".I- I

a.-.. - -1

'" .... ...

L)

I I "

; --

--- ---

"'.-?,.. -5. 0. 5. 10. 15. 20. "'"'- .+..

;" ~POST-PROCESSING S/N IN DB 1 "

:-. Figure A60 Misclassification percentage versus post-processing SNR,,'-" ~~comparing the performance of various aspect zones. -. '..

V)~

115 . ....CC

• -" : ::::_:::

.U

-TO --. 0. . 10. 15 ini2.. .0.-- -

CLASSIFICATION OF GNO VCLSPOLARIZATION V

ELEY ASSUMED KNOWN ., .!

ELEVATION [DEG.) 27ASPECT ASSUMED KNOWN / KNOWN

MIN,MRX. INC ASPECT 0 10 10. YO0 50 5, 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 is 10 ~'907. CI (630.) #/- 3.Y% 2.8. 3.Y%7,

CLASS. FEATURES A4W A4W A4H

. .. . .... 0 DEG. ASPECT ZONE l-

45 DEG. ASPECT ZONE L :-- - 0 DEC. ASPECT ZONE

---- -- ---- --- - ~

-U C; --- - -- - -.. .. . ... ........ . . .. .... . . .. . . . .

,, , i Izi

01DII 0 " X I

"To. .. . S. .0:1,.20

er \

F" I \I...

POTPOESN S/ INO

- . .. ..... . ... ... ...... . .. . .. .. . . . . -N- ...-

zoi I •,

__ - -- ,,... *

t , I I I ,S- o. -5. o. 5. 10. 15. 20.,.



116

":. .::-::%° =~..*"

CLASSIFICATION OF GND VCLSPOLARIZATION HELEY ASSUMED KNOWNELEVATION (DEG.) 2?ASPECT ASSUMED KNOWN /KNOWNMIN.MAXINC ASPECT 0 10 10. .0 50 5, 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 15 1090% CI (030Z) 4/- 3.0J 2.6x 3.11%CLASS. FEATURES RAW A4M AW

______ 0 DEG. ASPECT lONE1--- -.... VS DEG. ASPECT ZONE:

- I;-L.__- I ." .p L.

- - ------ - -"Z

z - -- -- **\ ,, 1 !

m I ,, ".-. I I

- a i K-,! , :.,:,:.:.

-?o. -s. 0. 5. 10o. 15s. 20.:::::'POST-PROCESSING S/N IN 08 .

- Figure A.62 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones.

_-- :.... X.... -.... ..

cc" " '

40. .:-... ., . . . .%.. . . . . ... . . . . . - . , , ,, . , . , . . . . . . . .. .. :.

hi

CLASSIFICATION OF GNO VCLSPOLARIZATION V/H

ELEV ASSUMED KNOWN

ELEVATION (DEG.) 27 b

ASPECT ASSUMED KNOWN / KNOWN

MINMRX.INC ASPECT 0 10 10. YO 50 5. 80 90 10NO OF FREOUENCIES 8NO OF TARGETS 10 15 1090% CI 1030%) 4/- 3.47 2.87 3.0CLASS. FEATURES A4N R4W R4W

aTo I 0 DEG. ASPECT ZONE

.5 DEC. ASPECT ZONE----------- 90 DEG. ASPECT ZONE,

II i m - "i i -- -- -- - --.-- -

--- .. - -- - ," "--

- - \ -- " "-- .-1Lw C

c

- . .. . I S',!'" • % -

-TO. -5 0 5. 10 15 20.-" ..- -P.ST-PRCESSIN S, ,N OB ,. --- ,1

Figure A63 Misclassif i pe v s pco aspect zones.C;18""0

-o. -5 . o. . 10. is. 20." - ".. P OST-PROCESSING SIN IN DB . :..

." Figure A.63 Misclassification percentage versus post-processing "" '.:R,[ ~comparing the performance of various aspect zones. :-,

.: 118 .

le

CLASSIFICA71ON OF GN0 VCLSPOLARIZAT1ON V .-.

r., ELEY ASSUMED KNOWNELEVATION (BEG.) 27bASPECT ASSUMED KNOWN /KNOWNHINIMAX. INC ASPECT 0 10 10. 410 50 5. 80 90 10NO OF FREQUENCIES aNO OF 7ARGETS 10 15 1090% Cl (6307.) */- 3.41% 2.8Z. 3.14%x

CLASS. FEATURES W W W

T; 00DEG. ASPECT ZONE ..

- -6 -D-VSBG. ASPECT ZONES -- .90m DEG. ASPECT ZONE

L c -- --I

CL

Z

LU

X: acyn N

C;_~

119 * \\Z

CLASSIFICATION OF GNO VCLSPOLARIZATION HIELEV ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWN /KNOWNMIN.NAX.INC ASPECT 0 10 10, 410 50 5. 80 90 10NO OF FREQUENCIES aNO OF TARGETS 10 15 1090% CI 1@30%) 4/- 3.4%/ 2.8% 3.4%/CLASS. FEATURES W W W

C ______ 0 DEG. ASPECT ZONE-- -- - 15 DEG. ASPECT ZONE--------- 90 DEG. ASPECT ZONE

.. .. ... .. .. . .

0

0x

(jY

-to -0 . 0N s 0



120

CLASSIFICATION OF GNO VCLSPOLARIZATION V/HELEV ASSUMED KNOWN

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWN

MINNAX. INC ASPECT 0 10 10, YO 50 5, 80 90 10

NO OF FREQUENCIES 8NO OF TARGETS 10 15 10

90% CI 1030%) */- 3.4% 2.8% 3.4%CLASS. FEATURES W W W

C; 0 DEG. ASPECT ZONE- D. oEG. ASPECT ZONE

------------------------------- --------- 90 DEG. ASPECT ZONEC

..

------------------------

I-- i - \ •-S

2 S.'

a: ....

L o '"

-5 ... 0 0 1. 20.POST-PROCESSING S/N IN DB


:'" 121

.- *

[20.

, . .. . , ._ , . ".- .'." - POST. .• - PR. . CESSING... S/N ,. N. -. O-..-,-B . . . -. . . .'. -. -. -.

-.- ... ,

I: CLASSIFICATION OF GND VCLS

POLRRIZRTION V

ELEV RSSUMED KNOWN

ELEVATION (DEG.) 27RSPECT ASSUMED KNOWN /KNOWN

MIN.MAX,INC ASPECT 0 10 10, 'JO 50 5. 80 90 10

NO OF FREQUENCIES 8

NO OF TARGETS 10 Is 1090% Cl (30X) +/- 3.4% 2.BX 3.W/-


o _________ 0 DEG. ASPECT ZONE-- i M5 DEG. ASPECT ZONE - -

'"- .... .... 90 DEG. ASPECT ZONE

I- € . - _w.7

Lii

m-o ,,\

- . .

Cr'

cJ*)

(A A

"II

0

- . -5. 0. 5. 10. 15. 0.-" POST-PROCESSING SIN IN 08(

II

-5.: 0 ' 5.' 10 IS. "-0.


122

.. . . . . . . . ... .. . .. . ..: '*.** . .. .. . .. :: : *:-

CLASSIFICATION OF GNO VCLSPOLARIZATION HELEV ASSUMED KNOWN

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWNMIN,MRX.INC ASPECT 0 10 10. 40 50 5, 80 90 10NO OF FREQUENCIES 8

NO OF TARGETS 10 15 10 wt90% CI 10307.) +/- 3.Y% 2.8. 3.Y.CLASS. FEATURES T T T

• ....... ... ......... ; .... .... .~ ~ ~.... ... ......... . ..... . .. ... . . - '

0 0 DEC. ASPECT ZONE.15 DEG. ASPECT ZONE

---------- 90 DEG. ASPECT ZONE

C;!

L I . .. .'-.*..'..

°C ,LLJ

CL . .... .. .. = = . .. ,

..

0

I L . .. . .. . ... . ... .. .\ .. ....L) \ '.

- . I -

U

X:0

-j0.' -5. 0. 5. 10. 15. 20.



123

......... ........

s-.. "

CLASSIFICATION OF GND VCLSPOLARIZATION V/H

ELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWN

MIN.MAX.INC ASPECT 0 10 10, 4O 50 5. 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 i5 10

90Y CI (0307) 4/- 3.4X 2.87 3.r/.


o ____ _ 0 DEC. ASPECT ZONE

.5 DEG. ASPECT ZONE-- o DEC. ASPECT ZONE

J . .. . - - - ----

_ ...,

POS-P-CES -N -/ I- Bw--

LU

o.- .- oN

*oo--.-

a

U..0V)S

-.. .5% , -

00Cu

-TO. -5. 0. 5.. -is 20.

POST-PROCESSING S/N I N D6


124

CLASIFCATONOF G:O:VLSPOLARIZATION V

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWN /KNOWN 5HIN.MAX,INC ASPECT 0 10 10, 40 50 5, 80 90 10NO OF FREQUENCIES 8Na OF TARGETS 10 1s 10 .-.

90% CI 1030%) 4/- 3.YX 2.8% 3.14% *.,*


ASP ERR IN CURVE 1 2 3

C;~~ DEC. AfSPE-CT -ZONECD1 E.ASETZN

4S DEC. ASPECT ZONE

LJ

.. . . ....

L)S

-?L -. 0. 5... 10 . 2 .POTPOCSIGS/;ND

.i ..... . ...

.r .n

* 1!CLASSIFICATION OF GNO VCLS

POLARIZATION H-

ELEV ASSUMED KNOWN -

ELEVATION (DEG.) 27

ASPECT ASSUMED KNOWN / KNOWN

MINMAX.INC ASPECT 0 10 10. YO 50 5, 80 90 10

NO OF FREQUENCIES 8 - "

NO OF TARGETS 10 Is 10

907 CI (@307) +/- 3.4J% 2.87 3.YX -. ,.

CLASS. FEATURES A A R


____; _ 0 DEC. ASPECT ZONEV5 DEG. ASPECT ZONE

------- g9 DEG. ASPECT ZONE0

LU -.. .. - .. . . ......... ....

C; j_ C; -. -----UIJ -

z .. . .......

- - ! 4' .C;

rC nUa:'7 C; ....

U

C"

, .. ,

N7o o

-5. 0. 5. 10. 15. 20.POST-PROCESSING S/N IN DB

Figure A.71 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle.

126

." • • "

-. 2 2.~2;2. 2~AA..2 A AA .* A 2 ~ '-.---~ -~ d~ .'.

CLASSIFICATION OF GO VCLSPOLARIZATION V/HELEV ASSUMED KNOWNLELEVATION (DEC.) 27ASPECT ASSUMED KNOWN /KNOWNMIN.MAX.INC ASPECT 0 10 10. 4O 50 5. 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 is 1090X CI (0307) #/- 3.47. 2.8. 3.7JXCLASS. FEATURES A A A


"- I a DEC. ASPECT ZONE- ' -- 45 DEC. ASPECT ZONE---------- 90 DEG. ASPECT ZONE

* - --- ..... ...... ---- .

POSP CSSN S/ N B

LD

.. .

a-C - -S . - .

0

-?o. 0..5. 10..15. .20

POST-PROCESSING S/N IN 06Figure A.72 Misclassification percentage versus post-processing SNR,

comparing the performance of various aspect zones, witha forced error in aspect angle.

127

CLASSIFICATION OF GOD VCLSPOLARIZATION V

-. ELEV ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWN /KNOWNMIN,MAX. INC ASPECT 0 10 10. 'YO 50 5. 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 is 10907. C1 (8307) +/- 3.4Zi 2.67 3.'Y7CLASS. FEATURES A4W RAW A4W


a; 0 DEG._____ ASPECT ZONE t XMIS DE.ASPECT ZONE

--------- 90 DEG. ASPECT ZONE

wj ............ - ----- - - -. -

U.'

LLJL0.a

U

0. . 1. 15. 20



128 *-

CLASSIFICATION OF GO VCLSPOLRRIZRT1ON IIELEY ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWN /KNOWNMIN.MRX.INC ASPECT 0 10 10. YO0 50 5, 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 i5 10

CLASS. FEATURES A4W A4W ACM

C; 0S DE. ASPECT ZONE


;-

C; ----

L)

... ... ..

C ;. .. ........... . ... ...... ...

Cc II

-5U. 5 0 5 0

129%

. .. .... . . .. .... .... ........ .----- -- ........ .. . .

CLASSIFICATION OF GND VCLSPOLARIZATION V/H

ELEV ASSUMED KNOWN iELEVATION (DEG.) 27

ASt ECT ASSUMED KNOWN / KNOWNMIN.MRX,INC ASPECT 0 10 10. 40 50 5. 80 90 10NO OF FREQUENCIES 8, .NO OF TARGETS 10 15 10907. CI (0307) 4/- 3.4. 2.8% 3.4/-CLASS. FEATURES A&W A4W A4 .


0 DEG. ASPECT ZONE ",- - .. . . . . NS DEC. ASPECT ZONE

--------- 90 DEC. ASPECT ZONE

LDD

S . .. ....\ .:.......

% %%

.. , ~ ~ .... . ... .... .: .... :

-To: -5. 0. s. 10. 15. 20. p -POST-PROCESSING S/N IN SB

Figure A.75 Misclassification percentage versus post-processing SNR," :".'.-. comparing the performance of various aspect zones, wit~hk ~~a forced error in aspect angle. ; :.

"130

''""

- S"S*5

,

CLASSIFICATION OF GO VCLSPOLARIZATION VELEV ASSUMED KNOWNELEVATION (BEG.) 27ASPECT ASSUMED KNOWN /KNOWNMIN.MAXINC ASPECT 0 10 10. '40 50 5. 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 15 1090% CI (030X) */- 3.11% 2.87 3.4%/CLASS. FEATURES W W W

ASP ERR IN CURVE I 2 3

C; 0 DEG. ASPECT ZONE- --- - S DEG. ASPECT ZONE--------- 1g DEC. ASPECT ZONE

C;*%

LJ

C o

LLCL5

LI)

X0

-?. 5 0 . 10. i5. 20.POST-PROCESSING S/N I N DB

Figure A.76 Misclassification percentage versus post-processing SNR, .:.:


131

-,V L7. ""'j 7 .

=.,k.


POLRRZRTION H

ELEV ASSUMED KNOWN "*

ELEVATION (DEG.) 27ASPECT ASSUMED KNOWN / KNOWN

MIN.MAX,INC ASPECT 0 10 10. YO 50 5, 80 90 10NO OF FREQUENCIES 8

NO OF TARGETS 10 15 10

90X CI 18307) +/- 3.Y% 2.8% 3.Y7,

CLASS. FEATURES N W N

ASP ERR IN CURVE 1 2 3 . *

g 0 DEG. ASPECT ZONE-5 EG. ASPECT ZONE

- 9...... 0 DEG. ASPECT ZONE

a ® C'-

I~ ..... .. .... ... ... ... .. .

0

iii..-- .. *.-,.. ..

LU -

(CrsUN

• "- ...'.- -

.. .. .... .................... ... ....... ... ... ... ........... ... . . .... ..... ...... . . -. "

-5 0 s. 10. 1 . 20.-

POST-PROCESSING S/N IN DBE

Figure A.77 Misclassification percentage versus post-processing SNR, ;,.-

comparing the performance of various aspect zones, with".".'a forced error in aspect angle. ;/-

-132"

%A-

, ,., ,,,. -... . ., ... .-... . . . .... .,.. .. .. . . .. . . . .. . . .. . .- -,. . ,.. .. ., ... .. .,,;j .e

CLSSFIATONOF GNO VCLS/KNN

EL0VACIO (DE.) 27 .% 2.% 34

ASPECT ASUE KNEWN ASPCTWONMIN.MAX.INC ~M CE.ASPECT 0Z010ONE0 5 0 01

4010 DEG. ASPECT ZONE

LI

L&J0 -........

C;

0CI ...........-----0i

* 0. . 10 is 20. I

POST-PROCESSING S/N IN DBFigure A.78 Misclassification percentage versus post-processing SNR,


133 :~a. .~%

CLASSIFICATION OF GO VCLSPOLARIZATION VELEY ASSUMED KNOWNELEVATION (DEC.) 27ASPECT ASSUMED KNOWN /KNOWN?IIN.MAX. INC ASPECT 0 10 10. 40 50 5. 80 90 10NO OF FREQUENCIES 8NO OF TARGETS 10 1s 1090% CI (0307) 4/- 3.4% 2.8% 3.4%CLASS. FEATURES T T T


0 DEG. ASPECT ZONE45 EG.ASPCTZONE

........ 90 EG.ASPECT ZONE

~~. ....... --...-

LACC .---D-

(Li

.... .... ..

- -W

ZL

N

.... ..... ...........

-To. -S. 0. 5. 10. 15. 20.POST-PROCESSING SAN IN 0i

Figure A.79 Misclassification percentage versus post-processing SNR,comparing the performance of various aspect zones, witha forced error in aspect angle. -

134

?S.

CLASSIFICATION OF GND VCLSPOLARIZATION HELEV ASSUMED KNOWNELEVATION (DEG.) 27ASPECT ASSUMED KNOWN /KNOWNNIN.MAX.INC ASPECT 0 10 10. 40 50 5, 60 90 10NO OF FREQUENCIES 8NO OF TARGETS ]a is 10

ASP ERR IN CURVE 1.4 2.S 3.i SET ZN

C; ~~0 DEG.ASETZN15 DG SETZN

S-90 DEG. ASPECT ZONE

Cr

LU C

J

CcN.

.. . . . . . . .. ..... .. ...... ..... .- ~

Lr)

C1

0''. s . S. 10. 15. 20.POST-PROCESSING S/N IN DB


135

' 26

CLASSIFICATION OF GO VCLSPOLARIZA71ON V/H

ASETASMD KNOWN KNOWN 5. 8 o i

NO OF TARGETS 10 is 109O0% CI 1030%) 4/ 3.Y7 2.% .4


ASP ERR INCURVE 1 2 3

O ________ 0 DEG. ASPECT ZONE--- - -- - - -U5 DEC. ASPECT ZONE


0L

h 'f

C D2

cc '

LL)

0

Z..

cc II

ac foce eror.nasec.nge...,.

_ o

-5 0 S 10, s1360

OSTP OCSSIN S/NIN OFi gre .81 Miscassfictionperentge vrsu pot-prcesingSNR

copain te erorane fvaios..pctzoes .wt

APPENDIX B

AMPLITUDE AND PHASE RETURNS

V This appendix contains plots of the amplitudes and phases of

processed radar returns for one vehicle at 3 aspect angles (00, 450 and

9Q0), at 270 elevation angle, using vertical and horizontal

polarizations.

137

p~ • i-

04

0

CIZ

a:

-- O0 68.88 11L. 160.63 206.507 FREQUENCT [MHZ)

LI) e.

UL)- CD.v

cn.

o --

, )

93.00 68.88 114.75 160.63 206.50

FREQUENCY (MHZ) -. ~ -, ",',,,".

Figure B.1 RCS magnitude and phase response for Vehicle A, at 0*aspect zone using vertical polarization.

138 r%

.,

t33

C )

C)-

LU

CrIL

0

" : ...' .' #

a)J

3.00 68.88 14716.3206.50

V FREQUENCY (MHZ)

U.-' %..

U-)

, .% .

O

LU

.., ~ ~ I ' -'

SI ]I I- - -

C

93.00 68.68 114.75 160.63 206.50 i2"I FREQUENCY IMHZ)

Figure B.2 RCS magnitude and phase response for Vehicle A, at 450aspect zone using vertical polarization.

139

"," ." -" ." -" "" "" "" "," "' "" "" -....-" ," "".""." ""." ,"" ",. ". , " .". ,-'.-'. " -. , -' -" ', " -. -.,-' -' .' ,' ', .,.' .".v - " "- ,':,', * "

.,%.

LU"

0

LUJ

LrL

3.O0 68.88 114.75 160.63 206.50FREQUENCY (MHZ)

L .-(\j

(I-)

LUJ

9300 68.88 11.75 160.63 206.50FREQUENCY (MHZ) .. :-

Figure B.3 RCS magnitude and phase response for Vehicle A, at 900

aspect zone using vertical polarization. ,..

140

Z77

C-)

. ._ ,. .- .- --..-...-- -. - .03

SI' o *

Li-

;" 3006888. 4. 7 10. 065

LOL

IFREQUENCY MHZ)..

U .-- -%,

- ao -

I- nLUJ

U)

3. 00 68.88 1 1q4.75 160.63 206.50

FREQUENCY (MHZ]Figure BA4 RCS magnitude and phase response for Vehicle A, at 00

aspect zone using horizontal polarization. .

141

66P.

F.7 -C w-7 AM: a~. . ~ - C - "-

-. 5

CD

0)

S 0j

.CD

0L

CC0 :0

C)

p3.00 68.88 114.75 160.63 206.50FREQUENCY (MHZ)

LtLJ

C; 0

c-fl

CD

CD5

93.00 68.88 114.75 160.63 206.50

FREQUENCY (Mt-Z)

Figure B.5 RCS magnitude and phase response for Vehicle A, at 450 .aspect zone using horizontal polarization.

142

LUI

LUJ

U.)C;

3.068.88 1114.7S 160.63 206.50FREQUENCY (MHZ)

iC

LU

0 o

CO

3. 00 68.88 114.7S 160.63 206.50

Figure B.6 RCS magnitude and phase response for Vehicle A, at 900aspect zone using horizontal polarization.

r ~ ~143 ,.

I.

'34

4

'I'

~'J.

FILMEDP. a

a- *,~,

I

.4DTICa-

a'I.

~ ~.: ~ %'a'%~ %''~' % ' 'a '~I

Documents

CLASSIFICATION USING NULTIFREGUENCY V/2 … · ad-r162 026 ground vehicle classification using nultifreguency v/2 nultipolarization resonance radar(u) ohio strte uuiy columbus electroscience