September 4, 2006 20:51 00321

Fractals, Vol. 14, No. 3 (2006) 187204c World Scientic Publishing Company

TARGET DETECTION WITHIN SEA CLUTTER:A COMPARATIVE STUDY BY FRACTAL

SCALING ANALYSES

JING HU,, JIANBO GAO,, FRED L. POSNER, YI ZHENG and WEN-WEN TUNGDepartment of Electrical and Computer Engineering, University of Florida

Gainesville, FL 32611, USANaval Research Laboratory, Radar Division, Code 5313, 4555 Overlook Avenue SW

Washington, DC 20375, USADepartment of Earth and Atmospheric Sciences, Purdue University

West Lafayette, IN 47907, USAjhu@ecel.ufl.edugao@ece.ufl.edu

Received October 4, 2005Accepted March 3, 2006

AbstractSea clutter refers to the radar returns from a patch of ocean surface. Accurate modeling of seaclutter and robust detection of low observable targets within sea clutter are important problemsin remote sensing and radar signal processing applications. Due to lack of fundamental under-standing of the nature of sea clutter, however, no simple and eective methods for detectingtargets within sea clutter have been proposed. To help solve this important problem, we applythree types of fractal scaling analyses, uctuation analysis (FA), detrended uctuation analysis(DFA), and the wavelet-based fractal scaling analysis to study sea clutter. Our analyses showthat sea clutter data exhibit fractal behaviors in the time scale range of about 0.01 seconds to afew seconds. The physical signicance of these time scales is discussed. We emphasize that timescales characterizing fractal scaling break are among the most important features for detectingpatterns using fractal theory. By systematically studying 392 sea clutter time series measuredunder various sea and weather conditions, we nd very eective methods for detecting targets

Corresponding author.

187

September 4, 2006 20:51 00321

188 J. Hu et al.

within sea clutter. Based on the data available to us, the accuracy of these methods is close to100%.

Keywords : Fractal; Pattern Recognition; Sea Clutter; Target Detection.

1. INTRODUCTION

Sea clutter is the backscattered returns from a patchof the sea surface illuminated by a radar pulse.Accurate modeling of sea clutter and robust detec-tion of low observable targets within sea clutter areimportant problems in remote sensing and radarsignal processing applications, for a number of rea-sons: (i) identifying objects within sea clutter suchas submarine periscopes, low-ying aircrafts, andmissiles can greatly improve coastal and nationalsecurity; (ii) identifying small marine vessels, navi-gation buoys, small pieces of ice, patches of spilledoil, etc. can signicantly reduce the threat to thesafety of ship navigation; and (iii) monitoring andpolicing of illegal shing is an important activityin environmental monitoring. Since sea clutter isa type of electromagnetic wave, sea clutter studymay also help understand fading and non-Gaussiannoise in wireless communications, so that wirelesscommunication channel characterization and signaldetection can be greatly improved.

Due to massive reection of radar pulses fromwavy or even turbulent ocean surfaces, sea clut-ter is often highly non-Gaussian,18 even spiky,9

especially in heavy sea conditions. Hence, sea clut-ter modeling is a very dicult problem, and alot of eort has been made to study sea clutter.Traditionally, sea clutter is often studied in termsof certain simple statistical features, such as themarginal probability density function (pdf). Thenon-Gaussian feature of sea clutter has motivatedresearchers to employ Weibull,1 log-normal,24

K,57,10 and compound-Gaussian8 distributions tomodel sea clutter. However, such simple phe-nomenological modeling of sea clutter only oerslimited analytical or physical understanding.

To gain deeper understanding of the nature of seaclutter, the concept of fractal has been employed forthe modeling of the roughness of sea surface andinvestigation of scattering from rough surface.1113

Possible chaotic behavior of sea clutter has alsobeen studied.1421

Since the ultimate goal of sea clutter study isto improve detection of targets embedded within

clutters, a lot of eort has been made to design inno-vative methods for target detection within sea clut-ter. Notable methods include time-frequency analy-sis techniques,22 wavelet-based approaches,23 neuralnetwork-based approaches,2426 and wavelet-neuralnet combined approaches,27 as well as utilizing theconcept of fractal dimension28 and fractal error,29,30

and multifractal analysis.31,32 Note that most ofthe above works were based on the analysis ofradar images. To improve detection accuracy, someresearchers resort to higher resolution more power-ful millimeter wave radars.33 The status of the eldclearly indicates that one needs to adopt a system-atic approach, work on a large number of datasetsmeasured under various sea and weather conditions,and design a few readily computable parametersthat can accurately and easily detect targets withinsea clutter.

In this paper, we employ methods from randomfractal theory to study sea clutter. Specically, weapply three types of fractal scaling analyses, uc-tuation analysis (FA), detrended uctuation anal-ysis (DFA), and the wavelet-based fractal scalinganalysis to analyze three types of data, the mea-sured sea clutter amplitude data u(n), the datav(n) obtained by integrating u(n), and the dataw(n) obtained by dierencing u(n). These analysesshow that sea clutter data exhibit fractal behaviorsin the time scale range of about 0.01 seconds toa few seconds. By systematically studying 392 seaclutter time series measured under various sea andweather conditions, we nd very eective methodsfor detecting targets within sea clutter by applyingFA to u(n), DFA to both u(n) and v(n), and thewavelet-based approach to all three types of data.Based on the available data, the accuracy of thesedetectors is found to be close to 100%.

The remainder of the paper is organized as fol-lows. In Sec. 2, we briey describe the sea clutterdata. In Sec. 3, we introduce the three types offractal scaling analyses, FA, DFA and the wavelet-based fractal scaling analysis. In Sec. 4, we applythe three types of fractal scaling analyses to analyzethe three types of data mentioned above, and make

September 4, 2006 20:51 00321

Target Detection Within Sea Clutter 189

careful comparisons among the three types of anal-yses. Finally, some concluding remarks are made inSec. 5.

2. SEA CLUTTER DATA

Fourteen sea clutter datasets were obtained froma website maintained by Professor Simon Haykin:http://soma.ece.mcmaster.ca/ipix/dartmouth/data-sets.html. The measurement was made using theMcMaster IPIX radar at the east coast of Canada,from a clitop near Dartmouth, Nova Scotia. Theoperating (or carrier) frequency of the radar was9.39GHz (and hence a wavelength of about 3 cm).The grazing angle varied from less than 10 to a fewdegrees. The wave height in the ocean varied from0.8m to 3.8m (with peak height up to 5.5m). Thewind conditions varied from still to 60 km/hr (withgusts up to 90 km/hr). Data of two polarizations,HH (horizontal transmission, horizontal reception)and VV (vertical transmission, vertical reception),were analyzed here. Each dataset contains 14 spa-tial range bins of HH as well as 14 range bins of VVdatasets. Therefore, there are a total of 392 sea clut-ter time series. A few of the range bins hit a target,which was made of a spherical block of styrofoamof diameter 1m, wrapped with wire mesh. This is avery small target, more dicult to detect than, say,a ship. Usually, the range bin where the target isstrongest is labeled as primary target bin, and theneighboring range bins where the target may alsobe visible labeled as secondary target bins. How-ever, due to the drift of the target, it is possible thatthe target in a primary range bin may not be thestrongest and some secondary target bins may nothit the target at all. Each range bin data contains217 complex numbers, with a sampling frequency of1000Hz. We analyze the amplitude data. Figure 1shows two examples of the sea clutter amplitudedata without and with target. Note that similarsignals have been observed in many dierent elds.Therefore, the analysis used in this paper may alsobe applicable to those elds.

3. FRACTAL SCALINGANALYSIS

In this section, we describe one of the prototyp-ical models for random fractals the fractionalBrownian motion (fBm) model, then present thethree types of fractal scaling analyses.

0 20 40 60 80 100 120 1400

5

10

15

20

25

Ampl

itude

Time (sec)(a)

0 20 40 60 80 100 120 1400

5

10

15

20

25

Ampl

itude

Time (sec)(b)

Fig. 1 Examples of the sea clutter amplitude data (a) with-out and (b) with target.

3.1. Fractional BrownianMotion (fBm)

FBm BH(t) is a Gaussian process with mean 0,stationary increments, variance

E[(BH(t))2] = t2H (1)

and covariance:

E[BH(s)BH(t)] =12{s2H + t2H |s t|2H} (2)

where H is the Hurst parameter. The incrementprocess of fBm, Xi = BH(i + 1) BH(i), i 1,is called fractional Gaussian noise (fGn) process.It is a zero mean, stationary Gaussian time series.Noting that

E(XiXi+k) = E{[BH(i + 1)BH(i)] [BH(i + 1 + k)BH(i + k)]},

by Eq. (2), one can easily obtain the autocovariancefunction (k) for the fGn process:

(k) =12{(k+1)2H2k2H+|k1|2H}, k 0 (3)

September 4, 2006 20:51 00321

190 J. Hu et al.

hence,

(k) k2H2 as k , (4)when H = 1/2, the process is called memoryless orshort range dependent, the most well-known exam-ple being white Gaussian noise and its integrationbeing the standard Brownian motion process. When0 < H < 1/2, we have negatively correlated incre-ments; a jump up is more likely followed by a jumpdown and vice versa. When referring to fBm, thisis called anti-persistence by Mandelbrot.34 For1/2 < H < 1, we have positively correlated incre-ments. This means that a jump tends to be followedby another jump in the same direction. In fBm, thisis called persistence.34 Such processes have longmemory properties.

By the Wiener-Khinchin theorem, one nds thatthe power spectral density (PSD) for the fGn pro-cess follows a power-law,

EX(f) 1/f2H1. (5)Furthermore, the PSD for the fBm BH(t) time seriesis of the form

EBH(t)(f) 1/f2H+1. (6)The processes under study are thus often called1/f noise. Such type of noise is very ubiquitous.For classic examples, we refer to Refs. 35 to 37. Morerecently, it has been found that network trac,3840

DNA sequence,4143 human cognition,44 ambigu-ous visual perception,45 coordination,46 posture,47

dynamic images,48,49 the distribution of primenumbers,50 etc. all belong to such type of stochasticprocesses. It is further observed that principle com-ponent analysis of such processes leads to power-lawdecaying eigenvalue spectrum.51

For a general random-walk-type 1/f pro-cess (i.e. not necessarily Gaussian process) with1 < < 3, it can be proven that52,53

= 2H + 1. (7)

We shall see later that Eq. (7) may hold for a widerrange of or H, depending on which method isemployed to analyze a dataset. However, when Hdoes not belong to the unit interval, then one cannotsay the process to have persistent or anti-persistentcorrelations.

Note that while in principle Eq. (5) or (6) maybe used to estimate the Hurst parameter, when thepower-law-like PSD is only valid for a limited fre-quency range, it may be dicult to determine asuitable region to dene the power-law scaling by

this approach. This point will be made more con-crete later.

3.2. Fluctuation Analysis (FA)

FA characterizes the second order statistic thecorrelation, in a time series. It works as follows.We consider a covariance stationary stochastic pro-cess X = {X(i), i = 1, 2, . . .}. A stochastic pro-cess is covariance stationary if it has constant mean = E[X(i)], nite variance 2 = E[(X(i) )2],and covariance E[(X(i) )(X(i + k) )] thatdepends only on k. We rst subtract the mean from the time series. Denote the new time series asx = {x(i), i = 1, 2, . . .}, where

x(i) = X(i) .Then we form the partial summation of x to con-struct a new time series y = {y(n), n = 1, 2, . . .},where

y(n) =n

i=1

x(i). (8)

Often, y is called a random walk process of x,while x an increment process. One then examineswhether the following scaling laws hold or not,

F (m) =|y(n + m) y(n)|2 mH , (9)

where the average is taken over all possible pairsof (y(n + m), y(n)). The parameter H is called theHurst parameter. When the scaling law describedby Eq. (9) holds, the process under investigation issaid to be a fractal process.

3.3. Detrended Fluctuation Analysis(DFA)

When a measured dataset contains some trends(say, linear), it is advantageous to employ DFA54,55

instead of FA. Since a lot of ocean waves of dierentwavelength contribute to the complexity of sea clut-ter, it is desirable to examine whether DFA may giveadditional information on the nature of sea clutter.When applying DFA, one works on a random-walk-type process, as described by Eq. (8). It involvesthe following steps. First one divides the time seriesinto N/m non-overlapping segments (where thenotation x denotes the largest integer that is notgreater than x), each containing m points; then onecalculates the local trend in each segment to be theordinate of a linear least-squares t for the randomwalk in that segment, and computes the detrendedwalk, denoted by ym(n), as the dierence between

September 4, 2006 20:51 00321

Target Detection Within Sea Clutter 191

the original walk y(n) and the local trend; nally,one examines if the following scaling behavior (i.e.fractal property) holds or not:

Fd(m) =

mn=1

|ym(n)|2 mH (10)

where the angle brackets denote ensemble average ofall the segments. For ideal fractal processes, FA andDFA yield equivalent results.52 In practice, whenthe data under study contains certain trends or isnon-stationary, DFA often works more reliably.

3.4. Wavelet-Based Fractal ScalingAnalysis

The wavelet-based fractal scaling analysis is basedon the coecients of a discrete wavelet decomposi-tion. It involves a scaling function 0 and a motherwavelet 0. The scaling function satises

0(n)dn = 1.

The wavelet 0 must have zero average and decayquickly at both ends.56 The scaled and shifted ver-sions of 0 and 0 are given by

j,k(n) = 2j/20(2jn k),j,k(n) = 2j/20(2jn k),

j, k Z,

where j and k are the scaling (dilation) and theshifting (translation) index, respectively. Dierentvalue of j corresponds to analyzing a dierentresolution level of the signal. One popular tech-nique used to perform the discrete wavelet trans-form (DWT) is the multiresolution analysis (MRA).The procedure of performing MRA is detailed asfollows:56

(1) At the j = 1th resolution, for each k =0, 1, 2, . . . , compute the approximation coef-cient ax(j, k) and the detailed coecientdx(j, k) according to the following formulae:

ax(j, k) =n

x(n)j,k(n)

=n

x(n)2j/20(2jn k)

dx(j, k) =n

x(n)j,k(n)

=n

x(n)2j/20(2jn k).

(2) The signal approximation SAj and the signaldetail SDj at the jth resolution level are com-puted as

SAj =k

ax(j, k)j,k(n)

SD j =k

dx(j, k)j,k(n).

(3) Repeat steps (1) and (2) for the (j +1)th reso-lution level, using the signal approximation SAjobtained in step (2) as the input signal.

Let the maximum scale resolution level chosen foranalysis be J . The signal can be reconstructed usingthe following equation:56

x(n) = SAJ +J

j=1

SDj =k

ax(J, k)J,k(n)

+J

j=1

k

dx(j, k)j,k(n). (11)

The rst term represents the approximation atlevel J , and the second term represents thedetails at resolution level J and lower. MRA buildsa pyramidal structure that requires an iterativeapplication of the scaling and the wavelet functions,respectively. This is schematically shown in Fig. 2.

To make the above procedure more concrete, letus take the Haar wavelet as an example. The scalingfunction and the mother wavelet of the Haar waveletare dened as

0(n) ={

1, 0 n < 1,0, elsewhere.

0(n) =

1, 0 n < 1/2,1, 1/2 n < 1,0, elsewhere.

Signal

SA1 SD1

SA2 SD2

SD3SA3SA: Signal ApproximationSD: Signal Detail

Fig. 2 Pyramidal structure of the output of wavelet mul-tiresolution analysis.

September 4, 2006 20:51 00321

192 J. Hu et al.

0

1

1

0.5 1 n

0(n) 0(n)

0

1

0.5 1 n

Fig. 3 The scaling function 0(n) and the mother wavelet 0(n) of the Haar wavelet.

They are shown in Fig. 3. We consider the signalx(n) consisting of noisy blocks, as shown in Fig. 4a.The signal approximations and details at resolutionlevels 1 through 3 are shown in Figs. 4b, 4d and 4f(left column) and Figs. 4c, 4e and 4g (right column),respectively. We have

x(n) = SA1 + SD1 = SA2 + SD2 + SD1= SA3 + SD3 + SD2 + SD1.

Let

(j) =1nj

njk=1

|dx(j, k)|2,

where nj is the number of coecients at level j,then the Hurst parameter is given by

log2 (j) = (2H 1)j + c0, (12)where c0 is some constant. When log2 (j) versusthe scale j curve is approximately linear for certainrange of j, the process x(t) is said to be fractal, withslope being 2H 1. In particular, a at horizontalline corresponds to H = 1/2.

Recalling that when applying FA or DFA, oneworks on a random-walk-type process. When oneemploys the wavelet-based fractal scaling analysis,one works on the original time series. When thisis the case, the Hurst parameters estimated by thethree methods would be consistent. However, if onealso applies the wavelet-based fractal scaling anal-ysis to the random walk process, then the esti-mated scaling exponent would be H+1, where H isobtained by either FA or DFA (when they are equiv-alent). This point will be made clearer in Sec. 4.

4. FRACTAL SCALINGANALYSES OF SEA CLUTTER

Given a measured sea clutter data, one does notknow a priori whether the original data should be

treated as a random-walk-type process or as anincrement process. For this reason, we analyze threetypes of data, the original data, the integrated data,and the dierenced data. Let us denote the originalsea clutter amplitude data by u(n), n = 1, 2, . . ..The integrated data v(n) is obtained by rst remov-ing the mean value u and then forming the partialsummation,

v(n) =n

i=1

[u(i) u], n = 1, 2, . . . . (13)

The dierenced data w(n) is dened as

w(n) = u(n + 1) u(n), n = 1, 2, . . . . (14)We rst discuss FA of these three types of data.

4.1. Fluctuation Analysis of SeaClutter

For notational clarity, we re-denote F (m) in Eq. (9)by F(u)(m), F(v)(m), or F(w)(m), depending onwhether FA is applied to the u(n), v(n), or w(n)time series.

4.1.1. Analysis of original sea clutter data

Let us apply FA to the original sea clutter ampli-tude data u(n) rst. We can directly apply Eq. (9)by replacing y(n) by the sea clutter data u(n). Rep-resentative results of log2 F(u)(m) versus log2 m forthe 14 range bins of one measurement are shownin Fig. 5a, where the curves denoted by open cir-cles are for data with the target, while the curvesdenoted by asterisks are for data without the tar-get. We observe the curves are fairly linear in therange of m = 24 to about m = 212. They corre-spond to the time scale range of about 0.01 seconds

September 4, 2006 20:51 00321

Target Detection Within Sea Clutter 193

200 400 600 800 1000

0

10

20

Sign

al x

(n)(a)

200 400 600 800 1000

0

10

20

Sign

al A

ppro

x. S

A 1

200 400 600 800 10005

0

5

Sign

al D

etai

l SD 1

(b) (c)

200 400 600 800 10005

05

1015

Sign

al A

ppro

x. S

A 2

200 400 600 800 10005

0

5Si

gnal

Det

ail S

D 2

(d) (e)

0 200 400 600 800 100010

0

10

20

Index n

Sign

al A

ppro

x. S

A 3

200 400 600 800 10005

0

5

Index n

Sign

al D

etai

l SD 3

(f) (g)

Fig. 4 (a) The input signal x(n), (b, d, f) and (c, e, g) are the signal approximations and the signal details at resolutionlevels 1 through 3, respectively. x(n) = SA1 + SD1 = SA2 + SD2 + SD1 = SA3 + SD3 + SD2 + SD1.

to 4 seconds, since the sampling frequency of thesea clutter data is 1000Hz. Thus in this time scalerange the sea clutter data can be classied as frac-tal. The H parameter of each curve is estimatedby tting a straight line to the log2 F(u)(m) versuslog2 m curve in the range of m = 24 to m = 212. Theestimated parameter is explicitly shown in Fig. 5b.We notice that the H parameters of the curves forthe data with the target are much larger than thosefor the data without the target. It turns out this isa generic feature for all the measurements.

What is the physical signicance of the two timescales, one about 0.01 seconds and the other arounda few seconds, identied in FA? We observe thatfor time scale up to 0.01 seconds, the amplitudewaveform of sea clutter is fairly smooth, as can

be evidently seen from Fig. 6. The time scale ofa few seconds may correspond to how fast the wavepattern on the sea surface changes. These timescales may vary slightly with sea and weather condi-tions. Interestingly, these two time scales have beenexplicitly accounted for by the compound-Gaussianmodel,57 where the time scale of a few seconds isconsidered as the decorrelation time of the texture.

FA suggests that the sea clutter data is a type of1/f noise for the time scale range of around 0.01seconds to a few seconds. It is interesting to esti-mate the PSD of sea clutter data to check whether itindeed decays as a power-law in the frequency rangecorresponding to the time scale range identied, andif yes, to check whether the relation of = 2H + 1holds.

September 4, 2006 20:51 00321

194 J. Hu et al.

0 2 4 6 8 10 12 14 163

2

1

0

1

log 2

F (u)

(m)

0 2 4 6 8 10 12 140

0.1

0.2

0.3

0.4

H

Primary Target Bin

Bins With Targets

(a) (b)

0 2 4 6 8 10 12 14 1602468

101214

log 2

F (v)

(m)

0 2 4 6 8 10 12 140.93

0.94

0.95

0.96

0.97

0.98

H

Primary Target Bin

Bins With Targets

(c) (d)

0 2 4 6 8 10 12 14 163

2.5

2

1.5

1

0.5

log2 m

log 2

F (w)

(m

)

0 2 4 6 8 10 12 140.2

0.1

0

0.1

0.2

Bin Number

H

(e) (f)

Fig. 5 The left column (a, c, e) shows the results by FA (log2 F (m) vs. log2 m) for the 14 range bins for the original seaclutter data u(n), the integrated data v(n), and the dierenced data w(n), respectively. The corresponding H values are shownin (b, d, f). Open circles denote bins with target, while * denote bins without target. This rule applies to all other gures.

(a) (b)

Fig. 6 Two short segments of the sea clutter amplitude data.

We have systematically estimated the PSD fromall the sea clutter data. Two representative PSDcurves are shown in Fig. 7 in log-log scale. Thedashed straight lines in Fig. 7 are in the frequency

range of 1 to 100Hz, which corresponds to the timescale range of around 0.01 seconds to a few sec-onds. Those two straight lines are obtained by theleast-squares t to the PSD curves in that frequency

September 4, 2006 20:51 00321

Target Detection Within Sea Clutter 195

102 101 100 101 102 103105

100

105

1010E y

(f)

f

slope = 1.14

(a)

102 101 100 101 102 103105

100

105

1010

f

E y(f)

slope = 1.72

(b)

Fig. 7 (a) and (b) show the PSD of the two sea clutterdata shown in Figs. 1a and 1b, respectively.

range. The slopes are 1.14 and 1.72, which areequivalent to H = 0.07 and 0.36. The Hurst param-eters for the same datasets estimated by FA are0.07 and 0.37. Hence, FA and spectral analysis arevery consistent. It is worth noting, though, that itis not easy to identify the scaling range by spectralanalysis alone.

Next, let us examine if a robust detector fordetecting targets within sea clutter can be devel-oped based on H estimated in the time scalerange of 0.01 seconds to a few seconds identied.We have systematically studied 392 time series ofthe sea clutter data measured under various seaand weather conditions. To better appreciate thedetection performance, we have rst only focused onprimary target bins, but omitted those secondarytarget bins, since sometimes it is hard to deter-mine whether a secondary target bin really hitsa target or not. After omitting those secondarytarget range bins, the histograms (equivalent topdfs) for the H parameter under the two hypothe-ses (the bins without targets and those with pri-mary targets) for HH and VV datasets are shownin Figs. 8a and 8b, respectively. We observe that

the histograms completely separate for the HHdatasets. This means the detection accuracy canbe 100%. The accuracy for the VV datasets is alsovery good, except for two measurements. Interest-ingly, those measurements correspond to the twoHH measurements with the smallest H values. Wesuspect there might be some kind of experimentalerror in those two measurements.

Before proceeding, we make a comment. If onetries to estimate H from other intervals of time byusing maximum likelihood estimation, then H failsto detect targets within sea clutter. This makes itclear that characterization of fractal scaling breakis among the most important when detecting pat-terns using fractal theory. This feature is particu-larly important in practice, since experimental datais nite, and therefore, may not conrm to idealmathematical denition of fractal processes withlong range correlations.

4.1.2. Analysis of the integrated data

Now we apply FA to the data v(n) obtained by inte-grating u(n). A typical result of log2 F(v)(m) versuslog2 m is shown in Fig. 5c and the H values (esti-mated by tting a straight line to log2 F(v)(m) ver-sus log2 m in the range of m = 24 to m = 212) forthe 14 range bins shown in Fig. 5d. From Fig. 5c,one would conclude that the data have excellentfractal scaling behavior. However, this is an illu-sion due to the large y-axis range in the gure.This point will be further discussed in Sec. 5. Whilethe variation of H versus the range-bin numberstill indicates which bins hit the target, overall,the H values are very close to 1. Because of this,FA-based fractal scaling analysis becomes ineec-tive for the purpose of distinguishing sea clutterdata with and without targets. This can be readilyseen from Figs. 8c and 8d, where we observe thatthe histogram for the H parameters for the datawithout targets signicantly overlaps with that forthe data with primary targets for both HH and VVdatasets.

Let us explain why FA may fail for detecting tar-gets within sea clutter. This lies in the observationthat the largest Hurst parameter given by FA is 1.To explain this idea, let us assume y(n) n, > 1.Then |y(n + m) y(n)|2 = [(n + m) n]2 isdominated by the terms with large n. When this isthe case, (n+m) = [n(1+m/n)] n[1+m/n].One then sees that |y(n + m) y(n)|2 m2, i.e.H = 1. We call this the saturation phenomenon

September 4, 2006 20:51 00321

196 J. Hu et al.

0 0.1 0.2 0.3 0.4 0.50

10

20

30

40

His

togr

am

0 0.050.10.150.20.250.30.350

10

20

30

40

50

His

togr

am

(a) (b)

0.7 0.8 0.9 10

10

20

30

40

His

togr

am

0.8 0.85 0.9 0.95 10

5

10

15

20

His

togr

am

(c) (d)

0.01 0.005 0 0.005 0.010

10

20

30

H

His

togr

am

0.01 0.005 0 0.005 0.010

10

20

30

H

His

togr

am

(e) (f)

Fig. 8 The left column (a, c, e) shows the results by FA (histograms of the bins without targets and those with primarytargets of HH datasets) for u(n), v(n) and w(n), respectively. The histograms of VV datasets are shown in (b, d, f). Openboxes denote the range bins without targets, while solid black boxes denote the bins with primary targets. This rule appliesto all other gures.

associated with FA. An important implication ofthis discussion is that whenever one observes aHurst parameter very close to 1, one has to bealerted that it may be advantageous to re-do theanalysis by treating the original time series as arandom walk process instead of an incrementprocess. In other words, apply FA on the originaldata instead of the integrated data. It should benoted that one can similarly prove that H estimatedby FA cannot be negative.

4.1.3. Analysis of the dierenced data

We now apply FA to the data w(n) obtained bydierencing u(n). A representative result of

log2 F(w)(m) versus log2 m for one single measure-ment is shown in Fig. 5e and the variation of Hversus the range-bin number shown in Fig. 5f. Weobserve that all the curves are almost at in therange of about m = 24 to m = 216, thus theH values for the 14 range bins are all very closeto 0. Since the dierence between the H parame-ters for the range bins with and without the targetis very small, FA-based fractal scaling analysis againbecomes ineective for the purpose of detecting tar-gets within sea clutter data. This can be readilyappreciated from Figs. 8e and 8f, where we observethat for both HH and VV datasets, the histogramsof the H parameters for the sea clutter data withand without targets signicantly overlap.

September 4, 2006 20:51 00321

Target Detection Within Sea Clutter 197

4.1.4. Brief summary

It is evident from the above analyses that the seaclutter data should be treated as a random-walk-type process. Realizing this, one can then readilydevelop a method for detecting targets from seaclutter radar returns.

4.2. Detrended Fluctuation Analysisof Sea Clutter

Let us now discuss DFA of sea clutter. We also re-denote Fd(m) in Eq. (10) by Fd(u)(m), Fd(v)(m), orFd(w)(m), depending on whether DFA is applied tothe u(n), v(n), or w(n) time series.

4.2.1. Analysis of original sea clutter data

Let us start with applying DFA to the original seaclutter amplitude data u(n). We analyze the samemeasurement that has been studied by FA ear-lier. Representative results of log2 Fd(u)(m) versuslog2 m for the 14 range bins are shown in Fig. 9a.We observe that the curves for data with and with-out the target are all very similar to those obtainedby FA. We also identify the two time scales, oneabout 0.01 seconds and the other around a few sec-onds. Again we notice that the H values for the datawith the target are much larger than those for thedata without the target. This is explicitly shown inFig. 9b. It turns out that this feature is also gener-ically true for all the measurements.

Since the dierence between the results given byDFA and FA is very minor when the original seaclutter data u(n) is considered, we thus concludethat the sea clutter data does not contain any sig-nicant trends, especially linear.

Let us now carefully examine if the H param-eter estimated by DFA can be developed into arobust target detector within sea clutter. We havealso systematically studied 392 time series of thesea clutter data measured under various sea andweather conditions, by focusing on bins with pri-mary targets. The histograms for H under eachhypothesis (the bins without targets and those withprimary targets) for HH and VV datasets are shownin Figs. 10a and 10b, respectively. We observethat the detection accuracy for both HH and VVdatasets is very high, except for two measurements.Interestingly, those measurements correspond to thetwo measurements where FA works on the HH butfails in the VV measurements.

4.2.2. Analysis of the integrated data

Let us now apply DFA to the data v(n) obtainedby integrating u(n). Figure 9c shows a representa-tive example of log2 Fd(v)(m) versus log2 m for onesingle measurement, where the curves denoted byopen circles are for data with the target, while thecurves denoted by asterisks are for data withoutthe target. As will be discussed in Sec. 5, the seem-ingly good fractal scaling behavior is also an illu-sion. The H value for each curve is estimated bytting a straight line to log2 Fd(v)(m) versus log2 min the range of m = 24 to m = 212. The variation ofH versus the range-bin number for the 14 range binsis shown in Fig. 9d. We observe that the H valuecan be used for separating sea clutter data with andwithout the target.

Comparing Figs. 9b and 9d, one notices that theH values estimated from u(n) and v(n) time seriesdier by around 1. How may we understand thisfeature? Notice that when the process u(n) has aPSD of the form 1/f, where = 2H + 1, 0