Modelling and simulation of non-Rayleigh radar

Prof. E. Conte, DrEng

Prof. M. Longo, DrEng
M. Lops, DrEng

Indexing terms. Radar, Clutter, Modelling, Simulation

Regarding the amplitude distribution of the clutter, it

Abstract: As the undesired echo from a radar is most often assumed to be a Rayleigh distribution,
environment (clutter) is often not Rayleigh- implying that the quadrature components of the received
distributed, problems arise in specifying the clutter echo are jointly Gaussian processes: the usual intepreta-
as a non-Gaussian random process and in tion is that the clutter arises from the superposition of
computer-generating realisations of clutter for returns from a large number of equivalent elementary
simulation purposes. The paper shows that model- scatterers, independent of one another, whence the
ling the clutter as a compound Gaussian or a resulting process is Gaussian as a consequence of the
spherically invariant random process allows a central limit theorem. However, experimental data indi-
complete specification of the clutter suitable for cate that large deviations from Rayleigh statistics are
use in radar design. It also lends itself readily to observed in situations such as low grazing angles and/or
computer simulation procedures. Specific pro- high-resolution radars. In such cases the empirical dis-
cedures for generating clutter realisations with tributions exhibit both higher tails and larger standard
assigned correlation properties and with K or deviation-to-mean ratio than predicted by the Rayleigh
Weibull amplitude distributions are suggested, distribution (spiky clutter). Thus a single amplitude prob-
and are validated by statistical tests. Then the ability density function (APDF) cannot accommodate all
issue of resolving between the Weibull and K dis- of the different scenarios. Rather, we must consider a
tributions is considered, and empirical operative family of amplitude distributions, and we require that it
characteristics of the associated hypothesis test are include the Rayleigh one as a member.
computed. This requirement excludes from further consideration
the lognormal family which, although fitting data
obtained in some experiments, does not include the Ray-
leigh distribution. Instead, the Weibull and the K dis-
1 Introduction tributions reduce to the Rayleigh distribution for special
We consider the mathematical modelling of the undesired values of their parameters; moreover, both them have
return from a radar environment (clutter). The value of been validated by experimental data [l-81.
an accurate model is twofold: at the analysis stage it This paper introduces a coherent model for the clutter,
i.e. one in which the clutter in each range bin is specified
allows the performance of radar processors to be predict-
ed and at the design stage it allows optimum, or at least as a complex stochastic process, implying the joint sta-
sub-optimum, receivers to be synthesised. It is also desir- tistical characterisation of arbitrary order of its quadra-
able, for cases where computer simulation is required, ture components.
This model has to fulfil the cited constraints in terms
that clutter patterns conforming to the given model
be easily generated by algorithmic procedures. of amplitude distribution and correlation functions.
Clutter modelling involves in the first place the specifi- We adopt a compound model for clutter, namely we
cation of a random process, that is the assignment of a assume that the received clutter echo is the product of a
rapidly varying complex Gaussian process and an inde-
family of distributions of finite dimensional random
vectors that is internally consistent, in the sense that joint pendent, slowly varying, modulating process; with this
distributions of given order imply those of lower order. approach, the correlation properties of the overall
Such higher order joint distributions are of interest in the process are dictated by those of the Gaussian component.
design of optimum radar processors, e.g. Neyman- The compound model has been the subject of a sub-
Pearson detectors. stantial bulk of literature. In Reference 9, based on theo-
Further, in order that the model be realistic, some retical arguments, the PDF of the scattered power is
external consistency conditions must be satisfied to modelled as a gamma variate with random parameters,
ensure that the expectations computed on the given but no attention is paid to the correlation properties of
model comply with those obtained from experimental the received clutter echo.
data. Specifically, the expectations we have in mind are With reference to K-distributed clutter, the compound
the amplitude distribution and the correlation functions. model, in the form of the product of a Rayleigh rapidly
varying process and a chi-distributed, slowly varying one,
P a p a 7829F (E15).first received 14th March and in revised form 15th
has been considered in References 3 and 5-7. Since the
September 1990 first experimental results, presented in Reference 3, the
The authors are with the Dipartimento di lngegneria Elettronica, Uni- validity of this model for sea clutter has been supported
versita di Napoli, Via Claudio 21, 80125 Napoli, Italy by a complete set of evidence, as summarised in Refer-
I E E PROCEEDINGS-F, Vol. 138. N o . 2, A P R I L 1991 121
ence 7, accounting for the effect of various geometrical We also recall that specifying a process means assigning
and electromagnetic factors on the clutter statistics, such its joint PDFs of any order; therefore, a rule should be
as the shape parameter of the K distribution and the provided to derive such PDFs based on the correlation
correlation properties of both components. properties and on the APDF.
The correlation properties of the slowly varying com- For sake of example, we focus on two families of
ponent are studied in Reference IO. Following that work, ADPFs: the Weibull, namely
the extension to the bidimensional case is considered in
References 11 and 12, and simulation algorithms for such fA(u) = acuC-' exp (-au') U 20 (5)
slowly varying component are presented in Reference 13.
In this paper, a more general compound model is con- where c is a shape parameter and a is related to the
sidered, in the sense that it allows in principle for a wider average power U' of the quadrature components by
class of APDFs, although the worked examples refer to
the Weibull and K distributions, since these are the
APDFs for clutter returns which have received the most
attention. We pay little attention to the correlation
properties of the modulating component, since it is the and the K-APDF, namely
overall correlation of the compound process which deter-
mines the system performance in the applications we (7)
have in mind, namely predicting and analysing the per-
formance of radar processors subject to non-Gaussian
clutter. In fact, in Reference 6 it has been shown that the where v is the shape parameter, r(' ) the Eulerian func-
tion and b is related to U* by
spatial correlation of the clutter, which is reflected in the
correlation of the slowly varying component, can always
2v
be taken into account, for performance prediction pur- bZ =-
poses, by suitably modifying the shape parameter of the U2

received clutter.
For other applications, e.g. modelling the observed The Rayleigh APDF is the member of the Weibull family
spatial properties of coherently illuminated surfaces [l 11, corresponding to the special case c = 2, as well as the
it is the slowly varying component that contains informa- member of the K-family corresponding to the limiting
tion about such properties. Thus, a different approach is case v = CO.
required, in which the slowly varying component In these special cases the specification of the process is
becomes the primary focus of the modelling and pro- a problem which can be solved by classical methods
cessing, whereas the rapidly varying component - called since, as the quadrature components are jointly Gaussian
speckle in this content ~ is usually averaged out. In this processes, the joint PDFs of any order depend only upon
sense, our work complements that presented in Refer- the correlation function. One further advantage of Gauss-
ences 10 and 13. ian processes, related to simulation tasks, is their closure
property under linear transformations. This property
2 Clutter modelling makes it possible to generate first a white Gaussian
process and then to induce the desired correlation
2.1 Statement of the problem
properties by linear filtering. We shall see shortly that
We refer to the usual bandpass representation of the
clutter: these two features, namely the closure property and the
specification of the process based solely upon the covari-
x,,(t) = Re [x(t)ej"'] (1) ance matrix, are not unique to Gaussian processes but
where x ( f )is the complex envelope and can be written as are shared by a broader class of processes called spher-
+
x(t) = x,(t) j x s ( t ) = A(t) exp [jO(t)] (2)
ically invariant processes.
As the amplitude is no longer Rayleigh-distributed, the
in which x,(t) and x,(t) are the inphase and quadrature task of modelling and simulating radar clutter becomes
components, A(t) is the envelope and Q(t)is the phase more involved, since, in this case, assigning the corre-
process. We consider the problem of specifying the lation properties does not generally lead to the complete
complex envelope so that its first and second order sta- statistical characterisation of the process.
tistics fulfil given constraints of consistency with physical
requirements and of compliance with experimental
results. With regard to the complex envelope of the 2.2 Generalised Wiener model
clutter we impose two requirements: The Wiener model [17] for real processes with assigned
(a) the APDF, namely the first-order PDF of A(t), first-order PDF and correlation function can be gener-
must belong to a prescribed family of distributions. This alised to account for complex processes. The resulting
family should include the Rayleigh PDF as a member scheme is outlined in Fig. 1, in which, as in the following
(b) no constraint must be imposed on the correlation
function of x(t),i.e. the quantity
rJt, T ) = E[x(t)x*(t - r ) ] (3)
can be arbitrarily assigned. Note that this complex corre-
lation function amounts to the set of the following four
real correlation functions: Fig. 1 Model for a complex non-Gaussian correlated sequence: the
Wiener approach

figures, it is understood that the flow of complex quan-

(4) tities is indicated by thick lines. The desired process x(k)
is generated by transforming the envelope of a complex
correlated Gaussian sequence via a zero memory non-
122 I E E PROCEEDINGS-F, Vol. 138, No. 2, A P R I L 19YI

I
linearity (ZMNL), which leaves the phase process where uz(k) is the mean square value of the clutter quad-
unchanged. The ZMNL converts the Rayleigh input rature components and s(k) has been assumed, without
APDF into the desired non-Rayleigh output APDF. The loss of generality, to have unit root mean square value.
linear time-invariant filter H ( z ) is included to induce the Thus, the amplitude distribution is uniquely determined
desired correlation properties in the final output x(n). T o by the marginal PDF f(s) of the modulating process s(k).
this end, the relationship between the autocorrelation Moreover, as s(k) and y(k) are independent, the complex
function of y ( k ) and that of x(k) across the nonlinearity, correlation function of x(k) is
say
rx(n, m) = r3(4r,(n, m) (1 1)
r,(m) = dr,(m)I (9) where rdm) is the autocorrelation of s(k) and r,(n, m) is
should be determined, so that H ( z ) can be calculated via the (complex) autocorrelation of y(k). Since r,(m) r 1 for
spectral factorisation based on the autocovariance r,(m) any m where i-,(n, m) is not vanishingly small, the overall
resulting from eqn. 9. To be meaningful and realisable, correlation properties of x(k) are practically those of the
this scheme requires that r,(m) be a covariance function, underlying Gaussian sequence.
and in particular that it be non-negative definite. By Note that not all amplitude distributions admit the
referring to results of Reference 17 concerning the effects exogenous representation. In fact, a necessary and sufi-
of nonlinear processing on the spectral (or correlation) cient condition for a process x(k) with APDFf,(u; k ) to
properties of real (wide sense) stationary random pro- be admissible as an exogenous one is that there exists a
cesses, we observe that the above requirement is not PDF f(s), s > 0, which satisfies the type-I integral equa-
always satisfied. For example, it cannot be satisfied - tion of Fredholm (eqn. 10). It turns out that both the
unless the ZMNL is a polynomial - if the spectrum of Weibull and the K distributions are admissible [19, 201.
the output amplitude process A,(k) is to have some finite The appropriate f ( s ) for the K distribution is a gener-
zeros, and in particular if it is strictly bandlimited. As a alised chi-PDF [ S , 19, 211, namely
consequence, the autocovarance of A,(k), and hence the
complex autocovariance r,(m), cannot be arbitrarily
assigned for any shape of the ZMNL, that is to say for
any desired output PDF. The f(s) for the Weibull distribution can be expressed in
This approach was first proposed in Reference 14 for
terms of Meijer’s G functions, and hence reduces to some
lognormal clutter and then applied in References 15 and
known function for special values of their parameters.
16 to Weibull clutter. It is unmanageable, though, if the
However, in the general case an algorithm for its numeri-
output envelope is to have the K distribution: in fact, an
cal computation is still lacking, so that such an expres-
explicit expression for the ZMNL does not exist and,
sion is not very useful for practical purposes.
hence, no relationship between the input and the output
Besides the cited mathematical properties, exogenous
correlation functions can be established. Therefore, we
models are appealing in that they can be physically inter-
have to resort to a different approach, as outlined in the
preted in the light of the composite surface scattering
next subsection.
theory [22]. According to this theory, the returns from
any illuminated patch result from two independent con-
2.3 Exogenous model
tributions. The first component, with a Rayleigh-
With this model, introduced in Reference 18, the complex
distributed envelope, accounts for the reflection from a
envelope x(k) is considered to be the product s(k)j(k) (Fig.
very large number of independent, identically distributed
2), where y(k) is a zero-mean complex Gaussian sequence,
elementary scatterers and exhibits a decorrelation period
tied to the internal motion of the scatterers within the
resolution cell as well as to the radar wavelength. A
decorrelation time of the order of milliseconds is typical
of sea clutter [3]. The second component accounts for
the gross reflectivity characteristics of the illuminated
Fig. 2 Exogenous model for a complex non-Gaussian correlated patch and exhibits a decorrelation period which is associ-
sequence
ated with bunching of scatterers. For sea clutter, such a
component decorrelates on a time scale of several
and s(k) is a real, non-negative, stationary sequence, inde-
seconds [3, 71. This compound model is validated by the
pendent of y(k), which decorrelates on a time scale much
results reported in References 5 and 6, referring to sea
longer than y(k). The complex Gaussian process y(k) of
clutter with K-distributed amplitude. The data reported
Fig. 2 may be modelled as a complex white Gaussian
therein indicate that any variation in radar resolution or
process w(k) filtered through a linear (possibly time-
grazing angle results in a corresponding variation of the
varying) system, whose impulse response is determined
by the correlation properties of y(k). In other words, the shape parameter v of the generalised chi distribution of
the modulating process s(k). In Reference 6 it is also
clutter is modelled as a correlated Gaussian process
shown that the returns from groups of L correlated, iden-
modulated by the exogenous sequence s(k).
tically distributed resolution cells can be related to a
The main feature of this model is that it allows inde-
pendent control of the APDF, which is dictated by the single return from a single, L times larger, resolution cell
first-order PDF f(s) of the modulating sequence s(k), and whose radar cross-section has a chi distribution whose
parameter v depends upon the spatial correlation of the
of the correlation properties, which are essentially those
clutter.
of the Gaussian process. In fact, by the total probability
law, the envelope A(k) = s(k)1 y ( k ) 1 of the resulting
sequence has the marginal PDF 2.4 SlRP model
The exogenous model allows the amplitude PDF and the
correlation function to be controlled in accordance with
eqns. 10 and 11, respectively, but otherwise does not
I E E P R O C E E D I N G S - F , Vol. 138, No. 2, APRIL IY91 123
specify the higher order statistics. To achieve the Nth A further relevant feature of SIRPs concerns their
order characterisation of the process ( N finite) as reprentation in hyperspherical co-ordinates. Let x =
required for the analysis of signal processors operating llxl, ..., X 2 N / I T be a 2N-dimensional zero-mean white
with blocks of size N, we resort to the SIRP model. A SIRV. Its hyperspherical co-ordinates (R, 4’, . . ., d Z N -
spherically invariant random process (SIRP) [23, 241 can are defined as
be thought of as a ‘degenerate’ exogenous process in the
limiting case of constant s(k), in the sense that it is the
product of a Gaussian process y(k) and a modulating
x1 = R n
2N- 1

j= 1
sin 4j

random variate s, rather than a modulating random 2N-k

x k = R cos r $ 2 N + 1 - k fl sin qbj 2<k < 2N - 1

Tp
process s(k).
J=I
Two equivalent SIRP schemes are depicted in Fig. 3.
Fig. 3a stems directly from the general block diagram of x Z N= R COS qbl
hkm) h(k. m) O<R<CE 0<4,N-,<2n
O<qhJ<n j # 2 N - l (17)
The angular co-ordinates (I$’, ..., 4ZN-l) are indepen-
o b
dent of each other as well as of the radius R , and their
Fig. 3 SIRP modelfor a complex non-Gaussian correlated sequence
marginal PDFs do not depend on the marginal PDF of
n Premadulation hltering
the modulating variable s. On the contrary, the PDF of
b Postmodulation filtering R, is specified by the APDF of the complex SIRP, as it
h(k, m)is the impulse response ofthe (possibly time varying) linear filter can be written as [25]
an exogenous process (Fig. 2) with the sequence s(k) UZN- 1
replaced by the random variate s. The alternative imple- fdU)= 2N-”)h 2 N ( u 2 ) (18)
mentation of Fig. 36 is derived from the scaling property
of linear systems and is a direct proof of the closure of We refer to Appendix 8 for further details on the cited
SIRPs with respect to linear transforms. properties.
Most relevant features of exogenous processes apply Unlike a general exogenous process, a SIRP is non-
to SIRPs as well: in particular, SIRP‘s accommodate ergodic [19]. However, this is not a drawback as long as
both stationary and nonstationary situations, and allow we deal with short sequences. In fact, we can still assume
the APDF and the correlation properties to be indepen- that the clutter process is an exogenous one, in the
dently controlled. The applicability of the SIRP model general sense specified above. Nevertheless, if it is
for approximating the clutter process in sufkiently short observed and processed on time intervals much shorter
time intervals is in keeping with the results in Reference than the average decorrelation time of the modulating
6. sequence s(k) (as is actually the case for most practical
The multivariate PDF of the complex SIRP x(k) can situations), then it can be considered as a SIRP to a
be evaluated in a straightforward manner. From the reasonable degree of approximation (s approximately
scheme of Fig. 3a it is seen that, if we denote by constant).
x = IlX<,, ..,, x,,, x,,, ..., X,Il* (13) 3 Short-time simulation of non-Rayleigh clutter
a 2N-dimensional vector whose entries are N samples We distinguish the task of generating clutter samples by
from the inphase and quadrature components, then the ‘blocks’, whose duration is short compared to the decor-
PDF of x is relation time of the modulating process s(/~), from that of
f,(x) = (2x)-NIMI-’/Z

x exp
l+m
[-(I - F ) ~ M‘(x
-
s - ~ ~

- p)/(2sz)]f(s) ds
generating infinitely long streams of clutter samples, or,
practically, streams much longer than the said decorrela-
tion time. We call the first task ‘short-time simulation’
and the second task ‘continuous simulation’.
= ( 2 ~ )1 MI
- ~- ” Z h 2 N [ (-~ p)*M- ‘(x -p)] (14) In a short-time simulation by blocks of length N , we
have to generate a complex N-dimensional vector, whose
where p and M are the average and the covariance components are samples from the complex envelope of
matrices of x, respectively. Due to the structure of x (eqn. the clutter process, or, equivalently, a 2N-dimensional
13), its covariance matrix takes on the structure real vector (eqn. 13), whose components are samples from
the inphase and quadrature components of the clutter.
Two different approaches can be envisaged, depending
on whether the auxiliary variable s, whose P D F is the
where M,, and M,, are the covariance matrices of the solution of eqn. 10, can be generated or not.
inphase and quadrature components, respectively, and
M, and M, are their crosscovariance matrices. 3.1 Generation using the auxiliary variable: K-APDF
Since f,(x) depends on x through the quadratic form If an efficient algorithm producing the auxiliary variable
(x - p)*M-’(x - p), the vector is referred to as a spher- s exists, then a generation procedure based on either of
ically invariant random vector or SIRV. The function the models of Fig. 3 suggests itself, and can be imple-
h2N(. ) is non-negative and is related to the APDF of the mented by either of the schemes of Fig. 4. Here the trans-
SIRP by formation matrix G is determined by the correlation
matrix M by spectral factorisation. More precisely,
letting E be the matrix of the normalised eigenvectors of
M and D be the diagonal matrix of its eigenvalues, the
where U’ is the common mean power of the quadrature transformation matrix is simply [26]
components of clutter [19]. C = ED’/2 (19)
124 IEE PROCEEDINGS-F, Vol. 138, No. 2, A P R I L 1991
Short-time simulation of processes with K-distributed puted by subdividing this population into ten groups of
amplitude can be performed by this method since, in this size 1OOO; for all values of the shape parameter, the null
case, the auxiliary variable s is related to a gamma hypothesis that the generated data were from the desired
variate with parameters (v, v) by theoretical distribution could not be rejected at a signifi-
cance level of 5%.
y = s2

S
0

Fig. 4 S I R P modelfiir a complex non-Gaussian correlated sequence

n Premodulation filtering
b Postmodulation filtering

and since efficient algorithms exist to produce a gamma

variate. We cite the following from Reference 27:
(a) If v is an integer, then y is the sum of v exponential
variates with parameter v.
(b) If v is a semi-integer, then y is a chi-square variate
and hence can be easily generated as the square root of
the sum of 2 v squared independent standard Gaussian Fig. 5 CDFs for K-APDF with varying shape parameter v (SIRP
variates. model)
(c) If v is neither an integer nor semi-integer, it can be ~ ~ _ _
empincal
written as v = v 1 + v,, with v I an integer or semi-integer ~ theoretical
and 0 < v 2 < 0.5.Then y is the sum y1 + y 2 where y l is
gamma with parameter (vl. v) and hence can be generated
according to (a) or (b) above and y , is gamma with With regard to the fulfilment of correlation requisites,
parameters ( v 2 , v) and hence can be generated as the we refer to Fig. 6, where the estimated entries of the
product covariance matrix M , for v = 0.1, namely the values
i,,(m), are shown and compared to the desired auto-
Y z = vB/v
where q is an exponential variate with unit parameter
and B is a beta variate with parameters (v, , 1 - v,). The
beta variate B is in turn generated by the following stan-
dard rejection method [27]:
(i) generate two uniform variates with unit param-
eter, Ul and U ,
(ii) evaluate Y, = and Y2 --
(iii) if Yl + Y2 > 1 then go to step (i), otherwise
deliver

Using the above procedure, we generated segments of

length N = 8 from a zero-mean process with K-APDF, 0
0 1 2 3 4 5 6 7
assuming the following correlation specifications: 7

Fig. 6 Estimated correlation values (discrete) and design correlation

IIM,, 1 ;. j = lIMS8Ili,
j =
function (continuous) of correlated process with K-APDF ( v = 0.1, S I R P
(20) model)
= l l ~ c s l l i=
llMcslli,j .j 0
where p is the one-lag correlation coefficient. Owing to
these specifications the complex envelope x(t) is obvi- correlation function, shown as a continuous line plot.
ously wide-sense stationary (WSS) and its power spectral Besides this visual appreciation, the goodness-of-fit of the
density (PSD) is Gaussian-shaped. Moreover, the radio- estimated correlation to the theoretical one was tested via
frequency process x&t) of eqn. 1 is itself WSS [ 2 8 ] . Fig. Fisher-z transformation of the calculated coefficients.
5 compares the empirical cumulative distribution func- Results indicate that the null hypothesis that the covari-
tions (CDFs) obtained from M = 10000 independently ance matrix of the generated data is M,, as given by
generated complex vectors, each of length N = 8, to cor- eqns. 20 can never be rejected at a significance level of
responding theoretical CDFs, for several values of the 5%. Similar results were obtained with reference to M,,.
shape parameter v, the second parameter b being deter-
/ I i , = 1.
mined from eqn. 8 so that ( r 2 = 1 and hence llkfcc 3.2 Generation by co-ordinate transformation:the
The goodness-of-fit was assessed through double Weibull APDF
Kolmogorov-Smirnov tests, using a population of size Lacking an algorithm for the generation of the modulat-
loo00 made of one complex component from each gener- ing variate, we exploit the statistical properties of the
ated complex vector. The discrepancy indices were com- hyperspherical co-ordinates of any SIRV cited in section
IEE PROCEEDINCS-F, Vol. 138, N o . 2, A P R I L 1991 125
2 and suggest the generation scheme of Fig. 7. Starting where
with a white Gaussian random vector w,the rectangular-
to-spherical co-ordinates transformation produces a set
of 2N - 1 angular co-ordinates bl, . . ., $2N- which are

4I
rectangular-to-
spherical
co-ordinates
transformation
1 L
b r k
:%z::r!&r
co-ordinates
transformation
I
mapping G
Finally, using eqn. 18, we obtain the PDF of R :

fdu) =
l
2N")k1= ,
N vkc~'ea"'
(26)

Fig. 7 Generation scheme using spherical co-ordinates For purposes of simulation, one is interested in the C D F
FR(r).Integrating eqn. 26 yields the following expression:
suitable for any SIRV, independently of the APDF. To
produce the desired APDF, the radial co-ordinate R,
separately generated according to this APDF, merges
with the previous set of angular co-ordinates. Then a with
spherical-to-rectangular co-ordinate transformation
yields the Cartesian representation of the SIRV x,
having the desired APDF but uncorrelated components.
Finally, the desired covariance matrix is induced by the
linear mapping x = Gx,, with G as given in eqn. 19, and
a non-zero mean may be achieved by adding a constant
at the output. This procedure was implemented for generating segments
The only nontrivial step in this procedure is determin- of length N = 8 from a sequence with Weibull APDF,
ing the P D F of the radius R to match the APDF specifi- using the same correlation specifications as in the pre-
cation. This step involves the concatenation of eqns. 16 vious K-APDF example. Again, the empirical CDFs were
and 18. The following development refers to the Weibull
APDF. 99.99
Assuming, without loss of generality, that the clutter
quadrature components have unit mean square value,
and using the Weibull APDF (eqn. 5) with parameters (a,
c). eqn. 16 gives

50

dNe - o y r l l
= (-2)N - (21)
dYN
Using the rule for the Nth derivative of a composite func-
tion [29], we obtain
J
0.1 02 0.L 0.6 0.8 1 2 4 6 8 10
R
Fig. 8 Empirical and theoretical CDFs for a Weibull ADPF with
varying shape parameter c (SIRP model)

Since

we can write
Fig. 9 Estimated correlation values (discrete) and design correlation
function (continuous) of correlated process with Weibull APDF
~ (24) c= 15. SIRP model
k= 1

126 I E E PROCEEDINGS-F, Vol. 138, No. 2, APRIL 1991

computed from M = loo00 realisations, using several F,( ' ) of s(k) as
values of the shape parameter c and selecting the second
parameter a so that U' = 1 from eqn. 6. These empirical
CDFs were compared to the corresponding theoretical
ones both graphically (Fig. 8) and using the double
The autocorrelation of s(k) depends on both h,(m) and
Kolmogorov-Smirnov test, with the same data partition
go( ). Typically, the nonlinear processing enlarges the
of 10 groups of lo00 samples; again, the null hypothesis bandwidth, which does not meet the requirement that
of Weibull APDF could not be rejected at a significance r,(m) should remain close to one. Therefore h,(m) should
level of 5%.
be designed such that the bandwidth prior to the nonlin-
The fulfilment of the correlation specifications was ear processing is sufficiently narrow for the cited require-
assessed by the same statistical procedures as outlined for ment not to be endangered. Thus the requirements on
the previous K-APDF example. Fig. 9 compares the
h,(m) are essentially in terms of bandwidth, so that the
computed FCc(m)against the desired rcc(T),for c = 1.5. The
classical design procedures for lowpass digital filters are
Fisher-z test yielded acceptance of the null hypothesis of adequate, with bandwidth specifications dictated by the
compliance at a significance level of 1%. criterion that the correlation of s(k) should be close to
unity in the range of interest for rx(m).In the case of non-
4 Continuous simulation stationary x(k), a trial-and-error procedure might be
necessary.
In this case the time fluctuations of the modulating The procedure of Fig. 10 was applied for the gener-
process are no longer negligible with respect to the dura- ation of a WSS complex sequence with K-APDF and the
tion of the sequence to be generated, so that the SIRP following correlation specifications:
model does not apply and a general exogenous model is
needed. According to the scheme of Fig. 2, we are to gen- rcc(m)= rss(m)= (0.9)'"
erate a highly correlated sequence s(k), namely one whose
normalised correlation function is close to one as speci- r&) = r S M = 0 (32)
fied with reference to eqn. 1 1 . Since the actual shape of In this case the filter h,(m) is a simple autoregressive
the autocorrelation is immaterial, one possible generation system of the first order:
scheme is that of Fig. 10.
h,(m) = (1 - 0.81)1'2(0.9)"u(m)
where u(m) is the (discrete) unit step function and the
h l (k. m ) ZMNL is implicity defined by

where y(q. p ) is the incomplete Eulerian function. In Fig.

11 plots of the function g,(z) are shown for increasing
Fig. 10 Generation scheme Jar a correlated exogenous sequence
h,(k. m) is the impulse response 01 a possibly time-varying filter; h,(m) is the
impulse response a l a linear time-invanant filter

Here, a white complex Gaussian sequence w , ( k ) is gen-

erated and is fed to a linear filter whose time-lag impulse
response h,(k, m) is specified by the desired overall corre-
lation functions. If the process to be simulated is WSS,
then this filter is linear time invariant and its energy
transfer function I H,(B)IZ is proportional to the output
PSD.
Independently, a white real Gaussian sequence w,(k) is
generated and filtered by a linear time-invariant system
with impulse response h,(m), and is passed to a ZMNL,
gl(z) say, designed so that the output sequence has the
desired marginal PDFf(s). It is advantageous to imple-
ment gl(z) in the form Fig. 11 Behaviour oJgo(z)for various shape parameters
Sl(4 = I Yo(4 I Only positive values are represented, since qo( - 2 ) = -q&)

where go(z) is an odd function of z, producing the sym-

metric PDF
values of v. As v tends to infinity, s(k) converges to a con-
stant, which implies that Igo(z)l = 1, Vz. Note that the
implemented scheme does not allow the shape of the
autocorrelation of the modulating sequence to be easily
so that gl(z) is given by predicted, but this is not a true drawback in our
approach, in the sense specified in the introduction.
However, if one is interested in controlling the autocor-
Here Q(x) is the area under the trailing tail of the stan- relation of s(k), several simulation techniques can be
dard Gaussian variate and Fs0( . ) is related to the C D F adopted. In particular, one can generate the process s'(k)
IEE PROCEEDINGS-F, Vol. 138, No. 2, A P R I L 1991 127

r
by resorting to either of the two simulation algorithms values i,(m) are compared with the design correlation
introduced in Reference 13, which allow correlated obser- function r c c ( T ) . After the Fisher test the null hypothesis
vations to be generated from a gamma-distributed that the generated sequence possesses the design corre-
process. The first algorithm achieves the desired PDF, lation could not be rejected at a significance level of 5%.
but is applicable only for obtaining an exponential corre-
lation; the second algorithm on the other hand, allows 5 Resolvability of the Weibull and the K
wider control of the correlation function, but achieves the distributions
wanted PDF only approximately. Even so, however, an
exact control of the final correlation of the modulating Several authors have referred to either Weibull or K dis-
process requires a knowledge of an analytical relation- tributions after studying their goodness-of-fit to the avail-
ship between the correlation functions of s2(k)and s(k). able experimental data [l, 61. The question has then been
For several values of v, a sequence of loo00 samples raised as to whether the two distributions are essen-
was generated and empirical CDFs were computed and tially equivalent, whenever their parameters are con-
compared with theoretical CDFs. The results are present- nected by an appropriate relationship [30, 311.
ed in Fig. 12. A possible relationship results from the method of
moments (MM) by forcing the first two moments of the
two distributions to be equal, so that their parameters
are implicitly defined by the nonlinear equations system

where a, c, b, v have been introduced in eqns. 5 and 7.

Using this criterion, the closeness of the two distributions
has been studied in Reference 31 and it has been shown
that the theoretical CDFs exhibit similar behaviour.
An alternative relationship is based on a criterion of
minimum distance (MD), in the Kolmogorov-Smirnov
0.0001 0.001 0.01 01 1 10
X sense, between the Weibull and K distributions. If we
Fig. 12 Empirical and theoretical C D F s for K - A P D F with varying denote by F,(x; c) the Weibull CDF with parameter c
shape parameter v (exogenous model) (the parameter a being such that U’ = 1 in eqn. 6) and by
~~~~
empirical theoretical F,(x; v ) the K-CDF with parameter v (the parameter b
being such that u2 = 1 in eqn. 8), then their
Since independent realisations were not available, an Kolmogorov-Smirnov distance is defined as
indirect test of goodness-of-fit was carried out exploiting
the necessary and sufficient condition that the modulat- d(c, v) = max I F,(x; c) - F,(x; v) I x > 0 (34)
ing variate s has the generalised-chi distribution of eqn. The relationship between the parameters c and v under
12 in order that the exogenous sequence has the the MM and MD criteria are plotted in Fig. 14; note that
K-APDF. Therefore, we tested the hypothesis that the
ZMNLs corresponding to the implemented v’s convert a
Gaussian variate into a generalised chi variate. After
Kolmogorov-Smirnov tests, this null hypothesis could
not be rejected at the significance level of 5%.
The fulfilment of the correlation requirements was
tested on a generated K-distributed sequence with
parameter v = 0.5. In Fig. 13 the estimated correlation

0.01 0.1 1.0 10

v
Fig. 14 Relationship between the shape parameters of K and Weibull
APDFs under method ofthe moments ( M M )and minimum distance ( M D )
criteria

U the two criteria are equivalent except for very low values
7 of v (or e), that is for very large deviations from the Ray-
Fig. 13 Estimated correlation values (discrete) and design correlation leigh CDF.
Junction (continuous) of correlated procesb with K - A P D F The problem of discriminating between these two
Y = 0.5. exogenous model populations can be handled as a hypothesis testing
128 I E E PROCEEDINGS-F. Vol. 138, N n . 2 , A P R I L l Y Y l
problem. Thus, letting 3 , be a sample of size M drawn with moderate sample sizes, such as M = 100 or 200. On
from a Weibull population with CDF F d x ; c) and X2be the other hand, for v = 2 (hence c = 1.53 according to the
a sample of the same size from a K-population with CDF MM criterion) the two distributions remain hardly
F , ( x ; v), we consider the decision problem resolvable, even if a sample size of M = loo0 is used.
H , : X I and Xzare drawn from the same
population 6 Conclusions
(35)
H I : TI and Xzare not drawn from the same This paper has addressed the problems of clutter model-
population ling and simulation in situations where the Gaussian
assumption no longer applies. The exogenous model can
As the test statistic we adopt the usual Kolmogorov- be adopted whenever the clutter amplitude is a
Smirnov measure of discrepancy [32], that is compound-Rayleigh process, i.e. when it fulfils an admis-
D , = max I F , ( x ) - F,(x) I x > 0 (36) sibility condition (eqn. 10). From a physical viewpoint,
this model corresponds to the composite surface scat-
wherein F , ( . ) and F2( .) are the empirical CDFs of X I tering theory, which regards the clutter as the product of
and Xz, respectively. two independent processes: the spiky component, which
The empirical operating characteristics of this test controls the gross reflectivity characteristics of the illumi-
were evaluated for several values of the sample size M nated patch and exhibits high average decorrelation time,
and for different values of the distributions parameters. and the speckle component, a Gaussian process which
The results are shown in Figs. 15 and 16, and demon- exhibits a much shorter decorrelation time.
The main advantage of exogenous models is that they
allow the marginal PDFs of the clutter quadrature com-
ponents and the covariance functions of the clutter
complex envelope to be independently controlled. The
former can be accommodated by a suitable choice of the
marginal P D F of the modulating process, while the latter
can be controlled by the correlation properties of the
underlying Gaussian process. Moreover, if the clutter
observation time is small enough, when compared
against the average decorrelation time of the modulating
process, then the spiky component can be thought of as a
random variate, i.e. the overall process degenerates into a
SIRP. In such a situation, the multivariate PDFs of any
order can be written by a simple recursive formula once
the covariance functions and the envelope PDF are
known.
With regard to the simulation of non-Rayleigh clutter,
Fig. 15 Operaling characferistics ofthe tesf resolving between K - and some procedures are presented and implemented for gen-
Weibull APDF for v = 0.1, c = 0.53 and varying sample sizes erating clutter patterns with assigned envelope PDF and
correlation properties. These techniques have been
applied both to the case of Weibull-distributed clutter
and to the case of K-distributed clutter. The results have
been validated by statistical as well as graphical tests.
Applications of proposed modelling and simulation
methods include the analysis and design of all radar pro-
cessors involving a single observation and no adaptive
thresholding technique. With some limitations, they also
include cases where multiple observations are processed,
either for adaptation purposes (CFAR) or for integration.
Besides the trivial, though unrealistic case of independent
00.2
.4[ ,,,/ observations, the case of a highly correlated modulating
component of the clutter is fully amenable to the
methods discussed here.
On the other hand, if the modulating component of
the clutter is only partially correlated over the relevant
observations (due either to spatial nonstationarity over
Fig. 16 Operafiny characferistics ofthe fesf resolving between K - and
the reference window in CFAR processing or to the dwell
Weibull APDF for Y = 2, c = 1.53 and varying sample sizes time exceeding the scale of stationarity of the gross reflec-
tivity characteristics of the illuminated cell in the case of
pulse integration) then the exogenous model is still
strate the power of the test (that is, 1 - /I, if /l
is the prob- applicable as long as the interest is in reproducing the
ability of a type-2 error) against the probability of a correct autocovariance function and the correct APDF of
type-1 error (that is a, or significance level). the compound process, but it does not allow complete
With increasing sample size, the resolvability of the specification of the same compound process, i.e. know-
two distributions improves. However, the shape param- ledge of PDFs of any order.
eter is also influential. In fact, for v = 0.1 (hence c = 0.53 A rather different approach is needed in applications,
under the MM criterion), that is for high deviations from such as remote sensing, where the interest is not in the
the Rayleigh CDF, the two distributions are resolvable overall compound process, but only in its modulating
IEE PROCEEDINGS-F, Vol. 138, No. 2, A P R I L 1991 129
component, since it contains information about the 22 WRIGHT, J.W.: ‘A new model for sea clutter’, IEEE Trans., March
surface characteristics one is willing to sense. 1968, AP-16, pp. 217-223
23 BREHM, H.: ‘Description and generation of spherically invariant
An incidental result, which is made possible by the speech-model signals’, Signal Process., 1987, 12, (2). pp. 119-141
models proposed here, is the assessment of the extent to 24 YAO, K.: ‘A representation theorem and its application to
which the Weibull and K distributions are actually dis- spherically-invariant random processes’, IEEE Trans., September
tinguishable from one another in statistical experiments. 1973, IT-19, (51, pp. W 6 0 8
25 GOLDMAN, J.: ‘Detection In the presence of spherically symmetric
Our analysis indicates that the two distributions can be random vectors’, IEEE Trans., January 1976, IT-22, pp. 52-59
distinguished quite easily for very spiky clutter, in the 26 FUKUNAGA, K.: ‘Introduction to statistical pattern recognition’
sense that the Kolmogorov-Smirnov test has an accept- (Wiley, New York, 1972)
able resolving power for moderately low sample sizes. 27 RUBINSTEIN. R.: ‘Simulation and the Monte-Carlo method’
Conversely, if the clutter is not very spiky, the two dis- (Wiley, New York, 1981)
28 PAPOULIS, A.: ’Probability, random variables and stochastic pro-
tributions are practically indistinguishable, even if the cesses’ (McGraw Hill. 1965)
sample size of the test is increased considerably. 29 GRADSHTEYN, IS., and RYZHIK, I.M.: ’Table of integrals, series
and products’ (Academic Press, London, 1983, 4th edn.)
30 DADDIO, E., GALATI, G., GIANNATEMPO, S., and PROTO-
7 References PAPA, A.: ‘Modelli non Gaussiani per il clutter superficiale’, Riuista
Tecnica Selenia, 1987, 10, (3), pp. 23-38
1 SEKINE, S., OHATANI, S., MUSHA, T., IRABU, T., KIUCHI, E.,
31 WARD, K.D.: ‘A radar sea clutter model and its application to per-
and TOMITA, Y.: ‘MTI processing and Weihull distributed clutter’,
formance assessment’. Proc. Radar-82 Conf., IEE ConJ Puhl. 216,
IEEE Trans., November 1982, AESIE, (4), pp. 729-730
London, October 1982
2 SEKINE, S., MUSHA, T., TOMITA, Y., HAGISAWA, T., IRABU, 32 MOOD, A.M., GRAYBILL, F.A.. and BOES, D.C.: ’Introduction to
T., and KIUCHI, E.: ‘On Weibull distributed weather clutter’, IEEE the theory of statistics’ (McGraw-Hill. 1974)
Trans., 1979, AESIS, pp. 824-828
33 POLLARD, J.K.: ‘A handbook of numerical and statistical tech-
3 WARD, K.D.: ‘Compound representation of high resolution sea
niques’ (Cambridge University Press, 1979)
clutter’, Elect. Lett., 6th August 1981, 17, (16), pp. 561-563
4 WATTS, S., and WARD, K.D.: ’Radar sea clutter’, Microwave J . ,
June 1985,22, pp. 109-121
5 WATTS, S.: ‘Radar detection prediction in sea clutter using the
compound K-distribution model’, IEE Proc. F, Commun., Radar & 8 Appendix
Signal Process., 1985, 132, (7), pp. 613-620
6 WATTS, S., and WARD, K.D.: ‘Spatial correlation in K-distributed Let x be a 2N-dimensional white SIRV. By eqn. 14, its
sed clutter’, IEE Proc. F, Commun., Radar & Signal Process., 1987, multivariate P D F can be written as
134, (6), pp. 526531
7 WARD, K.D., BAKER, C.J., and WATTS, S.: ’Maritime sur-
veillance radar. Part 1: Radar scattering from the ocean surface’, (37)
IEE Proc. F, Radar & Signal Process., 1990, 137, (2). pp. 51-62
8 WATTS, S., BAKER, C.J., and WARD, K.D.: ‘Maritime sur-
veillance radar. Part 2: Detection performance prediction in sea where h Z N ( . )is given in eqn. 16. The Jacobian of the
clutter’, IEE Proc. F, Radar & Signal Process., 1990, 137, (2), pp. hyperspherical transformation eqn. 17 is
63-72
9 LEWINSKI, D.J.: ‘Nonstationary probabilistic target and clutter
scattering models’, IEEE Trans., 1983, AP-31, (3). pp. 4 9 W 9 8
10 OLIVER, C.J.: ‘Correlated K-distributed clutter models’, Opt. Acto,
1985, 32, (12), pp. 15151547
I1 OLIVER, C.J.: ‘The representation of correlated clutter textures in Therefore, the multivariate P D F of the hyperspherical
coherent images’, Inverse Prob., 1988, 4, pp. 843-866 co-ordinates of the vector x can be written as
12 OLIVER, C.J.: ‘The representation of radar sea clutter’, IEE Proc.
F, Commun., Radar & Signal Process., 1988, 135, (6), pp. 497-500 fR,+i....,+2N-i(u’ ai, ‘ “ 1 a2N-1)
13 OLIVER, C.J., and TOUGH, R.J.A.: ‘On the simulation of corre- 2N-2
lated K-distributed random clutter’, Opt. Acta, 1986, 33, (3), pp.
223-250
14 FARINA, A., RUSSO, A., and STUDER, F.A.: ‘Coherent radar
detection in lognormal clutter’, IEE Proc. F, Commun., Radar & which proves that the hyperspherical co-ordinates are
Signal Process., 1986, 133, (I), pp. 39-54 independent of each other.
15 FARINA, A., RUSSO, A., SCANNAPIECO, F., and BARB-
AROSSA, S.: ‘Theory of radar detection in coherent Weibull Moreover, by writing the above expressions in a fac-
clutter’, IEE Proc. F, Commun., Radar & Signal Process., 1987. 134, torised form and introducing suitable normalisation
(2), pp. 174-190 factors, we find the following expressions for the marginal
16 LI, G., and YU, K.-B.: ‘Modelling and simulation of coherent PDFs:
Weibull clutter’, IEE Proc. F, Radar & Signal Process., 1989, 136,
(1)- PP- 2-12 .2N - I
17 WISE, G.L., TRAGANITIS, A.P., and THOMAS, J.B.: ‘The effect
of a memoryless nonlinearity on the svectrum of a random orocess’.
IEEE Trans., 1977, IT-23, (0, pp. 84-89
18 CONTE, E., GALATI, G., and LONGO, M.: ‘Exogenous modelling
of non-Gaussian clutter’. J. Inst. Electron. & Radio Ena.. 1987. 57.
(4), pp. 191-197
19 CONTE, E., and LONGO, M.: ‘Characterisation of radar clutter as
a spherically invariant random process’, IEE Proc. F, Commun.,
=(2N -k + 1)
Radar & Signal Process., 1987,134, (2), pp. 191-197
20 CONTE, E., LOPS, M., and ULLO, S.: ‘A new model for coherent
Weibull clutter’. Proc. Int. Conf. on Radar, Paris, 24th-28th April
1989, pp. 482487
21 JAKEMAN, E., and PUSEY, P.N.: ‘A model for non-Rayleigh sea
echo’, IEEE Trans., November 1976, AP-24, pp. 80&814 O<l<n kf2N-1 (40)

130 IEE PROCEEDINGS-F, Vol. 13X, N o . 2, A P R I L 1991