IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL.

19, 2022 8015805

A Statistical Model Based on Modified

Generalized-K Distribution for Sea Clutter
Penghui Huang , Member, IEEE, Zihao Zou , Xiang-Gen Xia , Fellow, IEEE,
Xingzhao Liu , Member, IEEE, and Guisheng Liao, Senior Member, IEEE

Abstract— Sea clutter magnitude distribution exhibits generalized Pareto (GP) distribution [5], generalized K (GK)
important guiding significance for the design of marine target distribution [6], [7], G distribution [8], KummerU distribu-
detection algorithms and the selection of the constant false alarm tion [9], and so on; and 3) the models adding the additional
rate (CFAR) detection threshold. In this letter, to deal with the
limitation of the traditional generalized-K (GK) distribution scattering component to match the sea spikes, such as KA
in highly heterogeneous clutter scene, a modified GK (MGK) distribution [10], KK distribution [11], and so on. Compared
distribution is proposed for sea clutter magnitude. By combining with the two-parameter models, GK distribution can be applied
the traditional GK and generalized Pareto distributions, the value to precisely fit the sea clutter magnitude distribution in a wider
range of power parameter and the applicable range of the dis- range based on three-parameter model, which is composed of
tribution model are extended. The applicability of the proposed
distribution model is verified by real-measured sea clutter data. generalized Gamma distribution and complex Gaussian distri-
bution, naturally adapting to the compound motion properties
Index Terms— Generalized-K (GK) distribution, parameter of sea clutter with the surface roughness modulated by gravity
estimation, sea clutter magnitude distribution.
and capillary weaves.
I. I NTRODUCTION However, the power parameter of the traditional GK dis-
tribution may not be precisely obtained in the case of highly

M AGNITUDE distribution is an important measure to

reveal the sea clutter properties, which plays an
important role in the marine target detection fields in last
heterogeneous scene since it is only limited to a positive num-
ber. To address this issue, a modified GK (MGK) distribution
is proposed to extend the application range by inheriting the
decades [1]. When radar resolution is relatively low, a large merits of the traditional GK and GP distributions. Therefore,
number of statistically independent scatterers will appear in compared with GP distribution, MGK distribution can more
a resolution unit, and then from the central limit theorem, precisely describe sea clutter due to more degrees of freedom,
the magnitude follows the Rayleigh distribution. However, and has a better fitting performance for relatively homoge-
with the increase of radar resolution, the sea clutter will exhibit neous clutter. Different from the traditional GK distribution,
more complex characteristics. Due to the influence of the sea MGK distribution has a wider scope of applications. In addi-
spikes, the tail of the sea clutter magnitude distribution is tion, compared with G distribution and KummerU distribution,
enhanced, causing the magnitude distribution strongly devi- MGK distribution exhibits a lower computational load in the
ating from Rayleigh distribution [1]. parameter estimation procedure, which will be analyzed in
To deal with this issue, many more flexible sea clutter Section III. Both the ground-based and airborne real-measured
distribution models were proposed in recent years, which can sea clutter data are utilized to illustrate the effectiveness and
be mainly categorized into three kinds: 1) multiparameter applicability of the proposed distribution model.
non-Gaussian models, such as log-normal distribution [2], The remainder of this letter is organized as follows.
Weibull distribution [3], and so on; 2) compound models Section II introduces the proposed clutter distribution
based on the two-scale model dividing sea clutter into texture model. Section III gives the distribution parameter estima-
component and speckle component, such as K distribution [4], tion result of the proposed distribution model. Section IV
presents some real-measured radar data processing results
Manuscript received February 2, 2021; revised April 7, 2021 and June 25,
2021; accepted June 28, 2021. Date of publication July 12, 2021; date and gives discussions and comparisons to validate the pro-
of current version January 6, 2022. This work was supported in part by posed distribution model. Finally, some conclusion is drawn
the National Key Research and Development Program of China under in Section V.
Grant 2017YFB0502700 and in part by the National Natural Science Foun-
dation Program of China under Grant 61801289. (Corresponding author: II. D ISTRIBUTION M ODEL
Zihao Zou.)
Penghui Huang, Zihao Zou, and Xingzhao Liu are with the School In this section, MGK distribution is proposed for sea clutter
of Electronic Information and Electrical Engineering, Shanghai Jiao Tong magnitude. First, some conventional distribution models are
University, Shanghai 200240, China (e-mail: [email protected];
[email protected]; [email protected]).
briefly introduced.
Xiang-Gen Xia is with the Department of Electrical and Computer
Engineering, University of Delaware, Newark, DE 19716 USA (e-mail: A. Conventional Distribution Models
[email protected]). The commonly applied non-compound distribution mod-
Guisheng Liao is with the National Laboratory of Radar Signal Processing,
Xidian University, Xi’an 710071, China (e-mail: [email protected]). els for the clutter magnitude include: Rayleigh, log-normal,
Digital Object Identifier 10.1109/LGRS.2021.3093975 Weibull, and so on. The interested readers can refer to [2]
1558-0571 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Peloponnisos. Downloaded on April 08,2024 at 13:15:19 UTC from IEEE Xplore. Restrictions apply.
8015805 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 19, 2022

and [3]. In the meantime, sea clutter motion characteristics Therefore, the pdf of |T | can be expressed as
can be considered into the magnitude distribution applications. ⎧ ν(a−1)  ν
According to the two-scale model of sea clutter, sea clutter can ⎪
⎪ t t ν−1
⎨ ba (a) exp − b · νt ,
⎪ ν>0
be divided into texture component and speckle component,
pT (t) =  ν (5)
where speckle component is modulated by texture compo- ⎪
⎪ t ν(a−1) t
⎪
⎩− a exp − · νt ν−1 , ν < 0.
nent [12]. Based on this theory, many compound-Gaussian b (a) b
models are proposed, including K, GP, GK, G, KummerU,
and so on. The probability density function (pdf) of GK From (5), it can be seen that T follows the modified
distribution is given as below. generalized Gamma distribution [15]. Then the pdf of MGK
GK distribution [6], [7] distribution can be obtained by combining the pdf of |T | and
 +∞ νμ−3   ν 
νt x t x2 Rayleigh distribution, i.e.,
p(x) = exp − − dt (1)  +∞  ν  
σ νμ (μ) σ 2t 2 |ν|t νa−1 t x x2
0
p(x) = exp − · 2 exp − 2 dt. (6)
where x denotes the clutter magnitude, (·) denotes the 0 ba (a) b t 2t
Gamma function, σ denotes the scale parameter, μ denotes Let μ = a, σ = b 1/ν , then the pdf can be rewritten as
the shape parameter, and ν denotes the power parameter,  +∞   ν 
|ν|t νμ−3 x t x2
ν > 0 [7], [13]. When ν = 2, GK distribution is equivalent to p(x) = exp − − 2 dt (7)
K distribution, indicating that K distribution is a special case 0 σ νμ (μ) σ 2t
of GK distribution [13]. where μ and σ denote the shape parameter and the scale
Compared with K distribution, it is known that GP distribu- parameter of MGK distribution, respectively. From (7), it can
tion is more suitable for highly heterogeneous clutter scene [8]. be seen that when ν > 0, the pdf of MGK distribution is
The conventional K and GK distributions designed against the same as that of GK distribution in (1). In particular, when
highly heterogeneous sea clutter may encounter the fitting ν = 2, MGK distribution is equivalent to K distribution. When
performance degradation, limiting their scopes of applications. ν = −2, MGK distribution is equivalent to GP distribution.
Furthermore, for relatively homogeneous sea clutter scene, Therefore, the proposed distribution model has a wider appli-
K and GK distributions will exhibit a better fitting performance cation range.
than GP distribution. Therefore, to obtain a distribution with The second characteristic function of the second kind of
a wider application range, an MGK distribution is proposed MGK distribution can be expressed as
 +∞ 
in the following.
φ(s) = ln s−1
p(x)x d x
B. Proposed Distribution Model  0     
1 s+1 s − 1 √ s−1
To deal with the limitation of the traditional GK distribution = ln   μ+ 2σ . (8)
in the complex sea clutter scene, MGK distribution is proposed (μ) 2 ν
in this section, which is composed of the Gamma distribution Then the second-kind log-cumulant functions can be written
to the power of 1/ν and complex Gaussian distribution, where as
√ 
the power parameter ν can be any value except 0. When d[φ(s)] 1 1
c1 = = (0, 1) + (0, μ) + ln 2σ (9)
ν > 0, the proposed distribution model is equivalent to the ds s=1 2 ν
traditional GK distribution model. When ν < 0, the proposed d 2 [φ(s)] 1 1
distribution model is composed of the reciprocal of generalized c2 = = (1, 1) + 2 (1, μ) (10)
ds 2 s=1 4 ν
Gamma distribution [14] and complex Gaussian distribution.
In particular, when ν = −2, the proposed distribution model d 3 [φ(s)] 1 1
c3 = = (2, 1) + 3 (2, μ) (11)
is equivalent to GP distribution model. ds 3 s=1 8 ν
Suppose that Z is a random variable with the magnitude where (m, x) = ((d m+1 ln[(x)])/d x m+1 ) denotes the
conforming to Gamma distribution. The pdf of |Z | [14] is polygamma function.
z a−1  z
p Z (z) = a exp − (2) The log-cumulant diagram of MGK distribution is shown
b (a) b in Fig. 1. From the figure, it can be seen that K and GP
where z denotes the magnitude, a and b denote the shape para- distributions exhibit a line in the log-cumulant space, which
meter and the scale parameter of Gamma distribution, respec- may not suitably work in a broader region. On the contrary,
tively. Therefore, the distribution of |Z | is given by MGK distribution covers the region ①–④ in the log-cumulant
 z a−1 function
 y
y space, indicating its wider application range. As for the other
FZ (z) = exp − d y. (3)
0 b (a)
a b three-parameter models, GK distribution covers the region ①
Let T = Z 1/ν , then Z = T ν , and the distribution function and ②, G and KummerU distributions cover the region ② and
of |T | can be expressed as ③, indicating that MGK distribution covers the widest region
⎧  tν  y in the log-cumulant space among these distributions.
⎪ y a−1
⎪
⎨ exp − d y, ν>0 In the following, the log-cumulant statistics of the
ba (a) b
FT (t) = 0 +∞ a−1  y (4) real-measured airborne radar data covering homogeneous and
⎪
⎪ y
⎩ exp − d y, ν < 0 heterogeneous clutter are shown in Fig. 2. It can be observed
tν ba (a) b from the figures that for relatively homogeneous clutter,
where t denotes the magnitude. GK distribution is suitable to fit the data. However, for
Authorized licensed use limited to: University of Peloponnisos. Downloaded on April 08,2024 at 13:15:19 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: STATISTICAL MODEL BASED ON MGK DISTRIBUTION 8015805

Therefore, the estimation of μ can be obtained as

 3  2
(1, μ) (2, μ)
μ̂ = arg minμ − .
c2 − (1, 1)/4 c3 − (2, 1)/8
(14)
Then the power parameter ν and the scale parameter σ can
be estimated through (9) and (10) as
⎧
⎪ (1, μ̂)
⎪
⎨ , c3 < (2, 1)/8
c2 − (1, 1)/4
ν̂ =  (15)
⎪
⎪ (1, μ̂)
⎩− , c3 > (2, 1)/8
c2 − (1, 1)/4
Fig. 1. Log-cumulant diagram of MGK distribution.  
1 1 1
σ̂ = √ exp c1 − (0, 1) − (0, μ̂) . (16)
2 2 ν̂
Based on (14)–(16), the estimated parameters of the pro-
posed MGK distribution can be obtained. Compared with the
traditional 2-D parameter search, the parameter estimation
procedure in (14)–(16) only requires a 1-D search, effectively
reducing the computational complexity.

Fig. 2. Log-cumulant statistics of the real-measured airborne radar data. IV. R EAL -M EASURED DATA P ROCESSING AND A NALYSIS
Log-cumulant statistics of the real-measured (a) homogeneous clutter and
(b) heterogeneous clutter. In this section, several sets of real-measured sea clutter data
are used to test the fitting performance of the proposed distri-
relatively heterogeneous clutter, the log-cumulant statistics bution model and other typical distribution models including
tend to be in the region with ν < 0, where GK distribution Rayleigh, log-normal, Weibull, K, GP, GK, G, and KummerU
may not suitably work, indicating the advantage of MGK distributions.
distribution. The goodness-of-fit tests are required to evaluate the fitting
accuracy [17], and χ 2 test is applied in this section. The
III. D ISTRIBUTION PARAMETER E STIMATION expression of χ 2 test can be written as
The traditional parameter estimation method [6] for GK K
( f i − N p i )2
χ2 = (17)
distribution requires a 2-D search, which leads to a heavy N pi
i=1
computational load. In the following, we apply the logarithmic
where K denotes the number of regions, f i denotes the number
cumulants method [16] to deal with the parameter estimation
of samples in the i th region, and pi denotes the estimated
of MGK distribution.
probability of the i th region.
The second-kind log-cumulant functions of MGK distribu-
tion are shown in (9)–(11). For the observed data, the estima-
tion of ck can be written as [16] A. Ground-Based Real-Measured Data Analysis
⎧
⎪ 1 
N In this section, the effectiveness of the proposed MGK dis-
⎪
⎪ = ln(x i )
⎪
⎨ ĉ 1 tribution model is validated through a set of ground-based real-
N i=1
(12) measured radar data, which was recorded by a Ku-band radar
⎪ 1 
N
⎪
⎪ with the marine target scanning mode. For this ground-based
⎪
⎩ ĉk = N [ln(x i ) − ĉ1 ] , k ≥ 2
k

i=1
radar with VV polarization, the echoes of the near-range sea
scene are received, where the sea sate is about 1–2 level,
where N denotes the number of samples, and x i denotes the
covering a sea area of China.
i th sample.
The estimated pdf curves of Rayleigh, log-normal, Weibull,
Due to the fact that (2, x) < 0 for x > 0, it can
K, GP, G, KummerU, GK, and MGK distributions are shown
be clearly observed from (11) that when c3 > (2, 1)/8,
in Fig. 3. For this sea clutter data, due to the fact that the
the power parameter ν can be calculated as a negative number.
estimated power parameter of MGK distribution is a positive
In this case, the clutter magnitude distribution is closer to GP
number, MGK distribution in this case is equivalent to GK
distribution. When c3 < (2, 1)/8, the power parameter ν can
distribution. In addition, all K, G, KummerU, and MGK
be calculated as a positive number. In this case, the clutter
distributions have almost the same fitting accuracy in the tail
magnitude distribution is closer to K distribution.
of the pdf curve. On the contrary, the pdf curves of the other
From (10) and (11), a function of the shape parameter μ
distribution models derivate from that of the measured data
can be obtained as
 3  2 to a certain extent. The parameter estimation results of these
(1, μ) (2, μ) distribution models are listed in Table I, and the results of
= . (13)
c2 − (1, 1)/4 c3 − (2, 1)/8 χ 2 test are given in Table II, where the number of regions

Authorized licensed use limited to: University of Peloponnisos. Downloaded on April 08,2024 at 13:15:19 UTC from IEEE Xplore. Restrictions apply.
8015805 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 19, 2022

Fig. 3. Estimated pdf curves of the real-measured ground-based sea clutter

data. (a) Estimated pdf curves. (b) Estimated pdf curves in logarithmic plot. Fig. 4. Range-Doppler spectrum of the airborne sea clutter data.
TABLE I
PARAMETER E STIMATION R ESULTS

Fig. 5. Estimated pdf curves of the airborne sea clutter data. (a) Estimated
pdf curves. (b) Estimated pdf curves in logarithmic plot.
TABLE III
PARAMETER E STIMATION R ESULTS

TABLE II
G OODNESS - OF -F IT T EST R ESULTS

is set as 30. Assume that the thresholds are 42.6, 41.3, and
40.1 for one-parameter, two-parameter, and three-parameter
TABLE IV
models with the confidence of 95%; therefore, all K, G,
G OODNESS - OF -F IT T EST R ESULTS
KummerU, GK, and MGK distributions are suitable to fit the
magnitude distribution of this ground-based radar data with
relatively low sea state.

B. Airborne Real-Measured Data Analysis

In this section, the effectiveness of the proposed MGK dis-
tribution model is validated by a set of airborne real-measured
sea clutter data, covering an area of sea surface with about
four-level sea state. Fig. 4 shows the range-Doppler spectrum
of this airborne radar data, from which it can be observed that
the clutter spectrum in an airborne radar system is significantly tables, it is clearly seen that GP, G, KummerU, and MGK
broadened due to the platform motion. distributions exhibit better fitting performance for this rela-
For this airborne real-measured clutter data, the estimated tively heterogeneous clutter. Considering that the estimated
pdf curves are shown in Fig. 5, and the corresponding para- power parameter of MGK distribution is a negative number,
meter estimation results and χ 2 test results are, respectively, the applications of K distribution and GK distribution are
listed in Tables III and IV, where the searching ceiling of limited in this relatively heterogeneous clutter scene. From
shape parameter of GK distribution is 50. From the figures and the χ 2 test results, it can be observed that G, KummerU,

Authorized licensed use limited to: University of Peloponnisos. Downloaded on April 08,2024 at 13:15:19 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: STATISTICAL MODEL BASED ON MGK DISTRIBUTION 8015805

is required. On the contrary, the parameter estimations of G

and KummerU distributions are time-consuming because of
the computational load of 2-D search. Therefore, MGK distri-
bution may be more suitable for the complex sea clutter with
the considerations of fitting performance and computational
complexity.

V. C ONCLUSION
In this letter, an MGK distribution for complex sea clutter
scene is proposed, which effectively expands the application
Fig. 6. χ 2 test results of the real-measured airborne radar clutter data. range for the sea clutter magnitude distribution. The effective-
TABLE V ness of MGK distribution model is verified through both the
P ROBABILITIES OF S UITABLY F ITTING real-measured ground-based and the airborne sea clutter data.
The processing results show that MGK distribution model
has a better fitting accuracy for the sea clutter scene and
a time-efficient parameter estimation solution. The proposed
distribution model will provide valuable guidance for the
algorithm design in marine target detection.

R EFERENCES
[1] H. Ding et al., “Overview and prospects of research on sea clutter
property cognition,” J. Radars, vol. 5, no. 5, pp. 499–516, Oct. 2016.
TABLE VI
[2] G. V. Trunk and S. F. George, “Detection of targets in non-Gaussian
RUNNING T IME OF PARAMETER E STIMATION sea clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-6, no. 5,
pp. 620–628, Sep. 1970.
[3] M. Sekine, T. Musha, Y. Tomita, T. Hagisawa, T. Irabu, and E. Kiuchi,
“Weibull-distributed sea clutter,” IEE Proc. F Commun., Radar Signal
Process., vol. 130, no. 5, p. 476, 1983.
[4] E. Jakeman and P. Pusey, “A model for non-Rayleigh sea echo,” IEEE
Trans. Antennas Propag., vol. AP-24, no. 6, pp. 806–814, Nov. 1976.
[5] A. Balleri, A. Nehorai, and J. Wang, “Maximum likelihood estimation
for compound-Gaussian clutter with inverse Gamma texture,” IEEE
Trans. Aerosp. Electron. Syst., vol. 43, no. 2, pp. 775–780, Apr. 2007.
[6] L. Verrazzani, M. Diani, M. V. Greco, and F. Gini, “Performance analysis
of two adaptive radar detectors against non-Gaussian real sea clutter
data,” IEEE Trans. Aerosp. Electron. Syst., vol. 36, no. 4, pp. 1429–1439,
Oct. 2000.
and MGK distributions have a good fitting accuracy for this [7] X. Pan, G. Liao, Z. Yang, and H. Dang, “Ship detection using online
radar data. update of clutter map based on fuzzy statistics and spatial property,” IET
Radar, Sonar Navigat., vol. 12, no. 5, pp. 499–505, May 2018.
[8] A. C. Frery, H.-J. Muller, C. C. F. Yanasse, and S. J. S. Sant’Anna,
C. Goodness-of-Fit Test and Running Time Statistics “A model for extremely heterogeneous clutter,” IEEE Trans. Geosci.
Remote Sens., vol. 35, no. 3, pp. 648–659, May 1997.
In this section, 322 sets of airborne real-measured clutter [9] L. Bombrun, G. Vasile, M. Gay, and F. Totir, “Hierarchical segmentation
data covering homogeneous and heterogeneous clutter are of polarimetric SAR images using heterogeneous clutter models,” IEEE
Trans. Geosci. Remote Sens., vol. 49, no. 2, pp. 726–737, Feb. 2011.
applied to compare the χ 2 test results and running time of [10] Y. Dong, “Distribution of X-band high resolution and high grazing
MGK distribution and other typical distributions. The χ 2 test angle sea clutter,” DSTO, Edinburgh, U.K., Tech. Rep. DSTO-RR-0316,
results of each dataset are shown in Fig. 6. At the confidence Jul. 2006.
[11] L. Rosenberg, D. J. Crisp, and N. J. Stacy, “Analysis of the
of 95%, the probabilities of suitably fitting of these distribution KK-distribution with medium grazing angle sea-clutter,” IET Radar,
models are given in Table V. From the Table V, it can be Sonar Navigat., vol. 4, no. 2, pp. 209–222, 2010.
observed that the probabilities of suitably fitting of Rayleigh, [12] K. Ward, R. Tough, and S. Watts, Sea Clutter: Scattering, the
K Distribution and Radar Performance. London, U.K.: IET, 2013.
log-normal, Weibull, K, and GP distributions are lower than [13] V. Anastassopoulos and G. A. Lampropoulos, “A generalized compound
those of G, KummerU, and MGK distributions, indicating the model for radar clutter,” in Proc. IEEE Nat. Radar Conf., Atlanta, GA,
flexibility of three-parameter model. In addition, the fitting USA, Mar. 1994, pp. 41–45.
[14] E. W. Stacy, “A generalization of the gamma distribution,” Ann. Math.
performance of MGK distribution is better than that of GK Statist., vol. 33, no. 3, pp. 1187–1192, Sep. 1962.
distribution due to the extension of power parameter. [15] E. W. Stacy and G. A. Mihram, “Parameter estimation for a general-
The running time of the parameter estimation for 50 datasets ized gamma distribution,” Technometrics, vol. 7, no. 3, pp. 349–358,
Aug. 1965.
with 2400 points is shown in Table VI, where the running [16] V. A. Krylov et al., “On the method of logarithmic cumulants for
environment is the same for all the experiments. From the parametric probability density function estimation,” IEEE Trans. Image
Table VI, it can be seen that, by applying the logarithmic Process., vol. 22, no. 10, pp. 3791–3806, Oct. 2013.
[17] X. L. Pan et al., “Experimental analysis of sea clutter using airborne
cumulants method, the parameter estimation of MGK distri- circular scanning SAR in medium grazing angle,” J. Syst. Eng. Electron.,
bution is time-efficient due to the fact that only 1-D search vol. 30, no. 1, pp. 68–77, Feb. 2019.

Authorized licensed use limited to: University of Peloponnisos. Downloaded on April 08,2024 at 13:15:19 UTC from IEEE Xplore. Restrictions apply.