Modeling and SAR Imaging of the Sea Surface: a Review of the State-of-the-Art with Simulations
11 Modeling and SAR Imaging of the Sea Surface:a Review of the State-of-the-Art with Simulations
Igor G. Rizaev, Oktay Karakus¸, S. John Hogan, Alin Achim
Abstract —Among other remote sensing technologies, syntheticaperture radar (SAR) has become firmly established in thepractice of oceanographic research. Despite solid experiencein this field, comprehensive knowledge and interpretation ofocean/sea and vessel wave signatures on radar images are stillvery challenging. This is not only due to the complex mechanismsinvolved in the SAR imaging of moving waves: Many technicalparameters and scanning conditions vary for different SARplatforms, which also imposes some restrictions on the cross-analysis of their respective images. Numerical simulation ofSAR images, on the other hand, allows the analysis of manyradar imaging parameters including environmental, ship, orplatform related. In this paper, we present a universal simulationframework for SAR imagery of the sea surface, which includesthe superposition of sea-ship waves. This paper is the first attemptto cover exhaustively all SAR imaging effects for the sea wavesand ship wakes scene. The study is based on well proven concepts:the linear theory of sea surface modeling, Michell thin-ship theoryfor Kelvin wake modeling, and ocean SAR imaging theory. Wedemonstrate the role of two main factors that affect imaging ofboth types of waves: (i) SAR parameters and (ii) Hydrodynamicrelated parameters such as wind state and Froude number. TheSAR parameters include frequency (X, C, and L-band), signalpolarization (VV, HH), mean incidence angle, image resolution(2.5, 5 and 10 m), variation by scanning platform (airborne orspaceborne) of the range-to-velocity (
R/V ) ratio, and velocitybunching with associated shifting, smearing and azimuthal cutoffeffects. We perform modeling for five wave frequency spectraand four ship models. We also compare spectra in two aspects:with Cox and Munk’s probability density function (PDF), andwith a novel proposed evaluation of ship wake detectability. Thesimulation results agree well with SAR imaging theory. The studygives a fuller understanding of radar imaging mechanisms forsea waves and ship wakes.
Index Terms —Synthetic aperture radar (SAR), SAR imagerysimulation, sea wave spectrum, Kelvin ship wake, velocity bunch-ing.
I. I
NTRODUCTION I N the field of remote sensing for oceanography and oceanicengineering in recent decades many studies have focused onSynthetic Aperture Radar (SAR) imaging technologies. SARimages provide useful information on ocean state, includingwind speed and direction, gravity waves, swells, currents, sea-ice structures, and meteorological and environmental condi-tions in bodies of water. SAR Imaging is used extensively toobserve vessels and their wave signatures, in order to monitor
This work was supported by the Engineering and Physical SciencesResearch Council (EPSRC) under grant EP/R009260/1 (AssenSAR).Igor G. Rizaev, Oktay Karakus¸ and Alin Achim are with the VisualInformation Laboratory, University of Bristol, Bristol BS1 5DD, U.K. (e-mail:[email protected]; [email protected]; [email protected])S. John Hogan is with the Department of Engineering Mathematics,University of Bristol, Bristol BS1 5DD, U.K. (e-mail: [email protected]) port traffic and so on. Although SAR has been used in oceanresearch for over fifty years, the understanding and applicationof ocean radar scenes are still being actively developed, andthere remain unresolved issues in the interpretation of SARimages. The physical principle of SAR imaging of movingwaves is complicated, involving multiple factors, which aredifficult to disambiguate. This problem can generally beapproached from two angles. The first approach involvesconsideration of the SAR configuration geometry such as thedirection of scanning (left or right-looking antenna), choiceof platform (airborne or satellite-based), in conjunction withthe range-to-velocity (
R/V ) ratio and scanning parametersincluding image resolution, incidence angle, and electromag-netic properties (i.e. frequency and polarization of the signal).The second approach involves a combination of linear andnonlinear radar imaging mechanisms such as real apertureradar (RAR), including tilt and hydrodynamic modulations(predominately in scanning or range direction), and specificSAR imaging process. The latter mainly relate to the azimuthdirection and arise from variation of the radial velocities ofthe sea surface scatterers. These are usually referred to in theliterature as facets. This mechanism is usually approximatedusing time-dependent or velocity bunching models [1]–[3],although other imaging models have also been proposed inthe past [4]. The contribution of the hydrodynamic componentof moving waves to the SAR image formation must also beconsidered, with larger orbital velocities of water particlescorresponding to greater image degradation. The existence ofship wakes signatures in SAR images is determined by vesselparameters and ambient sea state, as well as SAR scanningparameters. Higher wind velocity [5]–[8] or high sea waveamplitude reduces the visibility of transverse and divergentwaves, leaving only the central turbulent wake, and then theexternal boundary of the wake, which is formed by cuspwaves. All the parameters and processes described above areusually closely related and in combination can either distortor enhance SAR images of wakes. This creates difficulties inthe analysis and interpretation of real radar images. One of themain limitations of SAR systems, therefore, is their sensitivityto the motion of targets, which results in a distorted phasehistory of the backscattered signal.There is now a wide range of spaceborne SAR missions,such as TerraSARX, COSMO-SkyMed, NovaSAR-1, ICEYE-1, ALOS-2 and Sentinel-1, which are capable of producingvery high spatial resolutions. However, not all SAR missionscan provide the best wave imaging results, the reasons forwhich are complex and go beyond the single factor of ”imageresolution”. For example, higher satellite orbital altitude in- a r X i v : . [ phy s i c s . a o - ph ] F e b creases R/V ratio, as is the case for example with Sentinel-1,and this leads to additional image degradation and azimuthalcut-off of gravity waves with small wavelengths, which in-cludes vessel wakes. Although there are multiple sources ofreal SAR data, their general availability is limited as most datais proprietary, and not open-access. This incentivizes the simu-lation of SAR images. The use of simulators also advances thestate-of-the-art by developing new methods for processing realSAR images, and simulation is useful for the planning of newSAR missions. The central factor in forming radar images ofthe sea is an approximation of the backscattering power fromthe water surface. The methodological basis for simulatingreflections of a radar signal from the sea surface is derivedfrom the fundamental theoretical framework of electromag-netic (EM) wave scattering from rough surfaces, including:Geometrical Optics (GO) [9], [10], Phase Perturbation Model(PPM) [11], [12], Integral Equation Model (IEM) [13], FullWave Model (FWM) [14], Facet Approach (FA) [15], SmallSlope Approximation (SSA), for first order SSA-1 [16] andsecond order SSA-2 [17], and many others [18]. It is firstnecessary to note that the radar scattering mechanisms can bedivided into Bragg scattering and non-Bragg scattering, associ-ated with specular (or quasi-specular) reflection, and breakingwaves [19]. In the literature, for both scattering mechanisms agood amount of theoretical work has already been conducted.For the specular reflection region, with incidence angles about ◦ ∼ ◦ , the Kirchoff Approximation (KA) [20], also knownas the Physical Optics (PO) method, has been developed. Thecontribution of breaking waves in a normalized radar cross-section (NRCS) has been investigated in numerous works[19], [21]–[24]. At moderate incidence angles ( ◦ ∼ ◦ for VV, and ◦ ∼ ◦ for HH polarization), the dominantmechanism is Bragg scattering. Here, the Small PerturbationModel (SPM) [25] was developed, and then later the Two-Scale Model (TSM) [26]–[28], which is based on compositesurface Bragg theory [26], [29], [30]. The TSM is probablythe most prominent simulation method today. A three-scalemodel [31] has also been used, which took into account theimpact of intermediate-scale waves.In the TSM the wind wave spectrum is divided into tworegions with the purpose of separating the large-scale (longwaves) and small-scale (capillary waves) roughness elements.This partitions the wavefield into deterministic and statisticalcomputations. Unlike the SSA, there is no need for calculationof highly detailed sea surface models [32]. Although parallelcomputation solutions have been proposed that significantlyaccelerate the creation of electrically large sea surface models[33]–[36], these solutions require additional hardware devices,such as a graphics processing unit (GPU), combined withspecialist platforms like the Compute Unified Device Archi-tecture (CUDA). The TSM method avoids these requirementsand has the advantage of simplicity in its realization, and aconsiderably lower computational cost in both processing timeand computer memory. For example, in one study [22] it wasshown that the TSM is much more computationally efficientthan SSA-2, in a ratio of approximately 1:800.As the TSM has gained prominence many different im-provements and additions have been introduced: simulation of multiview SAR wave synchronization data for azimuth cutoffwavelength compensation [37], polarimetric TSM [38], seasurface velocity of wind retrieval [39], improved TSM [40],[41], facet-based modification of TSM [42], application ofmodified TSM to breaking waves simulation [23], and finallyinteresting approaches which, although not directly related toTSM, use the same scale-splitting principle to generate seasurface for infrared imaging [43] and for improving SSAcalculations [33], [44], [45].The various radar imaging mechanisms and scattering meth-ods mentioned above form the theoretical basis for SARimagery simulation of complex ambient sea wave and shipwake scenes. The first group of simulation studies focusedon the sea wave field [1]–[3], [46]–[57], while the secondincluded both sea and/or ship waves [6], [58]–[72].Many authors have compared simulated with actual imagespectra [1], [3], [46]–[48]. In [49] the simulations are relevantonly to spaceborne SAR imagery. In [2] attention was paidto the features of SAR image formation through a combi-nation of nonlinear and linear mechanisms. Simulations ofinterferometric SAR (InSAR) were proposed in other works[50], [51], [53]. Nunziata et al. [54] designed a simulator foreducational purposes, which, although simplified, still containsall the SAR image formation mechanisms. Time-domain SARsimulations of moving ocean waves were presented in [55].Simulators to produce SAR raw data, rather than actual imagesimulation, have also been developed, see [52], [56]. Recently,a simulator of SAR image spectra of ocean swell waves waspresented [57]. One of the earliest works on sea and ship wakeradar image simulation was developed in [58]. In [59] along-track interferometric SAR image simulation of ship wakeswas proposed. The simulation of the NRCS of Kelvin armswas compared with real SAR images for different spaceborneplatforms, in [61]. The bistatic configuration of the SAR imageformation models was also simulated in [63], [65], [68]. Somestudies focused in particular on ship wakes modeling. This wasthe case with [67] and, [72] where EM scattering simulationsfrom the ship-generated internal wave wake were presentedand for turbulent wake SAR imaging in [58], [62], [64].Linear and nonlinear sea surface models were considered andcompared in [66]. The polarized Doppler spectra of dynamicship wakes was investigated in [69]. In another work [70] amodel of far wakes was used. More general SAR simulatorsof ship wakes were designed in [6], [60], [71]. In [60] twofrequencies (X and L-band) are considered, while in [71]the visualization of ship wake features in SAR images wasevaluated. Zilman et al. [6], investigated the detection ofKelvin wakes in simulated SAR images, and assessed theinfluence of significant wave height on their detectability.Finally, other original modeling studies of radar scatteringfrom the ocean surface and ship wake include [73]–[78].Although all the works mentioned above, as a whole, forman important theoretical basis, individually they are somewhatincomplete, each concerning particular aspects of RAR/SARimaging, for example: (i) focusing only on spectra comparisonof simulated and real images, or (ii) simulating only one typeof vessel wave wake. The various works were also basedon particular SAR configurations, for example monostatic, bistatic, or interferometric. Further, it is important to note thatmost of these simulation methods have been proposed for fixedmodeling parameters only, for example, a single frequency,single scanning platform type, or single sea wave spectrum.Also, not all studies include the velocity bunching effect,which is crucial in forming SAR images of waves in motion.Further, some important modeling issues such as radar scenesize or azimuth cut-off effects are not always addressed. Thereproducibility of many of the published methods, and theirintegration, is also very laborious. All of these factors providea strong motivation for the design of a universal, relativelysimple simulation system, which can deal with the majority ofthe SAR imaging phenomena that arise when imaging movingwaves.In this paper, we present an extended study for modelingand simulation of SAR imagery corresponding to both ambientsea waves and superimposed ship wakes. The simulationmethod is based on the linear theory and stochastic conceptof sea surface modeling [79], [80], while the Kelvin wakemodel adopted here is based on Michell thin-ship theory, aspreviously considered in [6], [63].The fundamental contribution is to integrate and present,for the first time, the majority of the known phenomenaassociated with SAR imaging of Kelvin ship wakes. Theseinclude the effects of wind state and Froude number, andalso the effects of the SAR imaging characteristics of signalfrequency, incidence angle, polarization, the spatial resolutionof image cell, velocity bunching factors (image shifting andsmearing), and heading of moving ship relative to the radarline-of-sight. The numerical simulation results obtained are ingood agreement with the theory of SAR image formation formoving waves.The second original contribution is in the comparison ofdifferent sea surface spectra by detectability of ship wakein the resulting SAR images. The idea here is that an un-derstanding of the ’borderline condition’ when wakes aremaximally sensitive for imaging will make it possible toestimate the contribution of various spectra to the SAR image.This provides a greater understanding of how the SAR imagingof ship wakes will vary for different geographical areas. Thesimplest examples are coastal areas (fetch limited) or openwaters (fully developed seas). To the best of our knowledge, ananalysis in this respect has not been performed before. Thereare limited studies where various spectra are compared, butthen only in terms of ensuing absolute NRCS values [81]–[84]. By contrast we employ five well-known sea wave spectraincluding Pierson-Moskowitz [85], JONSWAP [86], Fung andLee [87], Elfouhaily et al. [88] and Romeiser et al. [73],and compare the contribution of these spectra to ship wakevisualization in SAR images. Additionally, the comparison ofthe utilized spectra in our work was also performed with Coxand Munk’s probability density function (PDF) [89], [90].The structure of this paper is as follows: Section II, startswith the generic presentation of the modeling of the seasurface, the governing equations are presented and specialattention is paid to the main modeling parameters. In Sec-tion III, the ship wake modeling approach is presented andcommon SAR vessel signatures are discussed. The SAR image formation within the TSM framework is formulated in SectionIV. Simulation experiments with discussions are included inSection V. Conclusions and potential directions for futureresearch are described in Section VI.II. S EA SURFACE MODELING
A. The random deep-sea surface model
The linear theory of sea surface gravity waves has beenused for describing ocean waves for more than 150 years[79]. In this theory, water is assumed to be inviscid andincompressible while its motion is irrotational. The basisfor describing three-dimensional moving sea waves relies onrandom phase modeling with summation of many independentharmonic waves, whereby they propagate within x - y and t spaces in direction θ , with Rayleigh distributed amplitude A ,and a uniformly distributed random phase (cid:15) ∈ (0 , π ) . Withinthe linear theory of surface waves, the sea wave elevationmodel is related to the fluid velocity potential Φ sea at freesurface through Z sea = − g ∂ Φ sea ∂t (cid:12)(cid:12)(cid:12)(cid:12) z =0 (1)where Φ sea ( x, y, z, t )= g (cid:88) i (cid:88) j A ij ω i e k i z sin [ k i ( x cos θ j + y sin θ j ) − ω i t + (cid:15) ij ] (2)Then, the irregular sea surface elevation model can be de-scribed using the double superimposition model as (3) Z sea ( x, y, z, t )= (cid:88) i (cid:88) j A ij cos [ k i ( x cos θ j + y sin θ j ) − ω i t + (cid:15) ij ] where k and ω are the wave wavenumber and wave circular(radian) frequencies, respectively. The general expression ofthe dispersion in deep water includes capillary and gravity-capillary waves as ω = gk (cid:0) k /k m (cid:1) (4)with k m = gρ/τ (5)where τ is the surface tension of water (N/m), ρ is the seawater density (kg/m ), and g is the gravitational acceleration.For the gravity range (short-gravity and gravity waves), theexpression in (4) is simplified as ω = (cid:112) gk (6)The amplitude A is expressed as: A ij = (cid:113) S ( k i ) D ( k i , θ j ) ∆ k i ∆ θ j (7)where S ( k ) is the wave spectrum as a function of angularspatial frequency (it can also be expressed as a function offrequency S ( f ) or circular frequency S ( ω ) ), D ( k, θ ) is the angular spreading function, while ∆ k , and ∆ θ are the sam-pling intervals of the wavenumbers and propagation angles.In simulations, two main parameters should be well-thought-out: (i) the minimum size of the model and (ii) the sizeof the scatterers (or facets) of the sea model. The minimumsize of the sea surface model affects the wavelength andthe accuracy of modeled large gravity waves, specifically theamplitude of large waves for the appropriate wind velocity V w , while the facet size affects the proportion of large andsmall-scale waves.Some studies [63], [91]–[93] have focused on determiningthe minimum size of the model for correctly transferring theenergy of the spectrum to the sea surface model. In [91] itwas argued that the size should be larger than the dominantwavelength of the sea waves to reflect their modulation effects.In [63], using Fung and Lee’s spectrum, it has been determinedthat the minimum size, L min , of the scene representing thesea surface can be approximated as L min = 3 . × V w . ,where V w . is the wind velocity at 19.5m above the meansea level. In another study [93], L min was determined viathe wavelength ( Λ p ) of the spectral peak (changes dependingon V w ) of the Pierson-Moskowitz spectrum as L min = 2Λ p .Another example based on Elfouhaily et al. spectrum [88],demonstrates the losing effect of the energy spectrum (de-creasing model heights) when the sea model size is reducedto 100 m (cf. Fig. 3 in [92]). In our calculations (SectionV), we select L min as 1000 m, which satisfies the correcttransferring of energy spectrum to the sea surface model in arange of V w = 3 . − m/s.The facet size determines the discretization level of thesea surface model or the high frequency component. Inthe simulator presented in [54], the facet size is half ofthe SAR resolution cell. In some studies [42], [65], it wasexperimentally determined that the average NRCS does notcrucially depend on the facet size in ranges from 0.5 to 2m [42] and from 0.5 to 1.5 m [65]. In [34] the facet sizewas used in the range of 0.1-1 m corresponding to differentSAR wave bands. However, choosing the facet size dependsprimarily on the modeling of the EM scattering. Thus, theSSA method requires sizes of orders less than one-eighth/one-tenth of the incident wavelength [33], [44], [45]. Withinthe two-scale model (TSM), the facet size must be largeenough when compared to the wavelength of Bragg waves,but sufficiently small compared to the wavelength of longwaves [94]. Hence, in [95], the facet size was determined as 10times the microwave wavelength at X and Ka bands. In [96],facet sizes of the order of 10 to 20 Bragg wavelengths wereconsidered. However, for the TSM, there are two versions thatdetermine division of the wavefield via separation scale (facetsize) into deterministic and statistical calculations [28]. For theelectromagnetic-hydrodynamic (EMH) two-scale model, theseparation wavenumber is approximated as k EMH ≈ k B / ,where k B is Bragg wavenumber. For the SAR two-scalemodel, the separation wavenumber is determined as k SAR ≈ π/p SAR , where p SAR is the size of the SAR resolution cell.In general, for SAR simulations a facet size equal to theSAR resolution cell is usually assumed [6], [56], [94], since itimplies a lower memory cost. On the other hand, taking into
Fetch = 25 km x, km y , k m -1-0.500.5h, m (a) Fetch = 80 km x, km y , k m -1.5-1-0.500.511.5h, m (b)Fig. 1. Effect of different fetch length (sea state) in the sea surface modelingfor JONSWAP spectrum. The size of the model is 0.5 × V w = 10 m/s, wind direction is ◦ . (a) fetch = 25 km. (b) fetch= 80 km. account the Nyquist criterion, the choice of SAR resolutionshould depend on the resolvable sea wave or ship wake size.Thus, in [97] the effect of the SAR resolution is demonstratedby considering the Kelvin wake system. Therefore, in thiswork we consider a facet size equal to the SAR imageresolution ( k SAR separation scale), which can be varied fordifferent SAR configurations. In simulations, we set the facetsize at 2.5 m, which is in the order of typical real SAR imageryand matches the details of both types of gravity waves: wind-and ship-induced.
B. Omnidirectional sea wave spectrum
The one-dimensional, omnidirectional wave spectrum, rep-resented in (7) is interpreted as the empirical interrelationbetween energy distribution and frequency. Many wind-wavemodels based on wind speed [82] have been developed andinvolve various additional parameters that control the energyof the spectrum.Different sea states can be reproduced by these models, withthe two main categories being (i) fully developed sea (wavesin equilibrium with the wind) and (ii) developing or young sea(modeled using fetch and/or inverse wave age parameter). Itshould be noted that, for the same wind speed, the wave heightor amplitude will differ for fully developed and developing seaconditions. This effect is demonstrated in Fig. 1.Another general characteristic is the existence of a rangeof wavelengths covered by the spectrum (e.g. from gravity toshort-gravity or gravity to capillary). A particular spectrumdescribes only gravity waves while another can go up tocapillary scale. Here for the first time we depict this effectin models. Fig. 2 displays small scale models where the sizefor both models was chosen as . × . cm (gravity-capillaryscale [6]) and the facet size is set to 0.1 mm (capillary scale).At the scale of small ripples the surface tension is the dominantrestoring force which aims to stretch the surface of the waterinto flatness. Even with a significant change in the scale ofthe sea model, its patterns (but not the heights) follow that ofgravitational waves, cf. Fig. 2-(a). That is to say, in the sense ofstructure, the model is scale-invariant, which confirms that thePierson-Moskowitz spectrum describes only the gravity part.This is further verified in Fig. 3. In contrast, the Elfouhaily et S PM x, cm y , c m (a) S E x, cm y , c m (b)Fig. 2. Effect of scale in the sea surface modeling. The size of the modelis 2.5 × V w = 8 m/s. (a) gravity Pierson-Moskowitz spectrum. (b) gravity-capillary Elfouhaily et al. spectrum. al. spectrum in Fig. 2-(b) shows smoother capillary patternswhich differ from the gravity scale.In the remainder of this section, we describe the mostwell-known spectra: Pierson-Moskowitz [85], JONSWAP [28],Fung-Lee [87], Elfouhaily et al. [88] and Romeiser et al.[73]. For ease of reference, henceforth we refer to them as S P M , S J , S F L , S E , and S R , respectively. Let us note thatwe only emphasize key parameters and general mathematicalformulations for all spectra, and we refer the readers to theoriginal publications for further details. It must be emphasizedthat in the literature, different spectra are often expressed interms of different frequency variables.Conversion to wavenumber spatial frequency is achievedthrough the following expression: S ( k ) = S ( f ) dfdk (8)where f and k are interconnected by dispersion relation (6).The first and second spectra considered include only thegravity component, while all others are implemented usingboth the gravity and the capillary regime of waves. The S P M spectrum is a well-known approximation for the gravity wavesand, for a fully developed sea state, is formulated in terms of k as [85] S P M ( k ) = α k exp (cid:20) − β (cid:16) gk (cid:17) V w . (cid:21) (9)where α = 0 . is a Phillips constant, β = 0 . , g is thegravitational acceleration, and V w . is the wind speed at 19.5m above the mean sea surface.The general form of the spectrum for developing sea statewas described in the Joint North Sea Wave Project (JON-SWAP) as [98] (10) S J ( k ) = α k − exp (cid:34) − . (cid:18) kk p (cid:19) − (cid:35) × exp ln γ exp − (cid:16)(cid:112) k/k p − (cid:17) σ where α = 0 . V w /F g ) . , and the parameter σ , whichdescribes the width of the spectrum is either σ = 0 . if k ≤ k p or σ = 0 . if k > k p , where k p is the peak wavenumber[86], [99] k p = (cid:34) π gV w √ g (cid:18) V w gF (cid:19) . (cid:35) (11)Here and above, V w is the wind speed at 10 m above themean sea surface and F is the fetch length in meters. The peakenhancement factor γ is usually set to 3.3, although sometimesother values are employed by certain authors [100].Another widely known spectrum was developed by Fungand Lee [87] specifically for the estimation of radar backscatterover L to Ku frequency bands. Their spectrum includes twoparts: gravity and capillary wave ranges, which join at theseparation point k j = 0 . rad/cm: S F L = S F Lg if k < k j and S F Lc if k > k j . The gravity spectrum S F Lg is the same asin (9), except for α = 2 . × − , while the capillary spectrumis given by (12) S F Lc ( k ) = 0 . π ) p − (cid:0) k /k m (cid:1) g (1 − p ) / × (cid:2) k (cid:0) k /k m (cid:1)(cid:3) − ( p +1) / with k m the same as in (5) and p = 5 − log ( V w ∗ ) . Thecalculation of V w ∗ is provided later within the paper, in (22),(23).Probably the most well known spectrum is Elfouhaily et al.,which, like the S F L spectrum, consists of both wave ranges. Ithas however an additional advantage in that, similar to the S J spectrum, it can also simulate different sea states. The generalexpression of the omnidirectional S E spectrum is S E ( k ) = k − ( B l + B h ) (13)where the gravity or long-wave part of the spectrum is B l = 12 α p c p c L P M J p exp (cid:34) − Ω √ (cid:32)(cid:115) kk p − (cid:33)(cid:35) (14)whereas the capillary or short-wave part of spectrum is givenby B h = 12 α m c m c L P M J p exp (cid:34) − (cid:18) kk m − (cid:19) (cid:35) (15)In the short-wave part, we included the shape spectrum factor L P M and the peak enhancement factor J P as in [93]. Allparameters of Elfouhaily’s spectrum can be found in theiroriginal article [88].Unlike in S P M and S J , Romeiser et al. spectrum has beendeveloped for the improved composite surface model for theradar backscattering [73], which is based on the Apel spectrum[101]. The main omnidirectional formulation is S R ( k ) = k − P L W H (cid:18) V w V n (cid:19) β (16)where P L is computed as P L = 0 . − k p k +0 .
53 exp − (cid:16) √ k − (cid:112) k p (cid:17) . k p . (17) The peak wavenumber, k p , is k p = 1 √ gV w (18)The wind speed exponent β is given by (19) β = (cid:20) − exp (cid:18) − k k (cid:19)(cid:21) exp (cid:18) − kk (cid:19) + (cid:20) − exp (cid:18) − kk (cid:19)(cid:21) exp (cid:34) − (cid:18) k − k k (cid:19) (cid:35) with the values of the constants being k = 183 rad/m, k =3333 rad/m, k = 33 rad/m, k = 140 rad/m, and k = 220 rad/m. W H is the shape of Bragg wave region of the spectrumand can be expressed as W H = (cid:104) k/k ) . (cid:105) . (cid:104) k/k ) . (cid:105) (cid:104) k/k ) . (cid:105) exp (cid:18) − k k (cid:19) (20)where k = 280 rad/m, k = 75 rad/m, k = 1300 rad/m, and k = 8885 rad/m.Since the S P M and S J spectra do not include high-frequency parts, the short wave Phillips spectrum is employedfor both spectra [102] W P ( k ) = βk − (21)where β = × − .In order to ensure the consistency of wind speed at differentaltitudes in different spectrum models and angular spreadingfunctions ( V w , V w . , and V w . ), we apply the Fung andLee method [87]. The same method was used for calculationof the wind friction velocity V w ∗ in (12) V w = ( V w ∗ / . z/Z ) (22)and Z = (0 . /V w ∗ ) + 4 . × − V w ∗ − . (23)where V w is the wind speed at altitude z above the mean seasurface in, cm.The sea surface elevation model is usually characterized interms of significant wave height H s (the average of 1/3 of thehighest presented waves) [80], which relates to Z sea or thesea wave spectrum as H s = 4 (cid:115)(cid:90) S ( k ) dk = 4 (cid:112) var ( Z sea ) (24)In Fig. 3, the elevation spectra for different wind speeds areshown. We further discuss in Section V the effect of ambientsea waves on the SAR image simulation. C. Directional spreading function
The underlying idea for the representation of wave direc-tionality is that gravity waves are aligned with the meanwind direction. However, shorter waves are more divergentand can propagate perpendicularly or even move against the wind direction and, then gravity-capillary waves which onceagain are aligned with wind direction [88], [103]. In [104], byusing buoy measurements it was concluded that the directionof waves changes when the size of the waves decreases.A detailed review on this has recently been published, see[105]. The frequency spectrum S ( k ) alone is not sufficientto adequately describe the propagation of sea waves in atwo-dimensional (2D) space. In order to represent the stateof superimposed directional components of the wave energytransferring along the wind direction, the angular spreadingfunction D ( k, θ ) (sometimes called angular/directional distri-bution function) also needs to be used. Although many spread-ing functions have been proposed (see for example [103]), asatisfactory model of the directional energy distribution stillneeds to be developed. Nonetheless, all types of the spreadingfunctions presented in (7) have to meet the condition of π (cid:90) − π D ( k, θ ) dθ = 1 (25)where the sea waves direction θ = θ − θ w is corrected by themean wind direction θ w , which is taken relative to the flightdirection of the SAR platform.One of the first practical and simple forms of the directionalspreading function is the cosine-squared spreading functionwith directional width around ◦ [60]. For | θ | > ◦ , D ( θ ) = 0, and for | θ | ≤ ◦ the function is D ( θ ) = 2 π cos θ (26)The main limitation of this function is that it does not explicitlydepend on frequency and wind speed, which means that allwave components propagate in the same direction. An alterna-tive to the cosine-squared spreading function [106] is presentedin Fig. 4-(a). In general, higher-order functions correspond tobetter directionality for wave propagation. Probably the bestknown and the most widely used is the Longuet-Higgins et al.cosine type spreading function (see [79]) based on empiricalfield measurements of pitch-and-roll buoys D ( θ ) = Γ ( S + 1)[Γ ( S + 0 .
5) 2 √ π ] cos S (cid:18) θ (cid:19) (27)where Γ( · ) is the gamma function, and the parameter S controls the width of the function and depends on k . Thisparameter was later refined in some studies [104], [107]. Forsimplicity, we keep S as a constant value, following [79].When S is increased, the directionality of waves also increases,which is shown in Fig. 4-(b). For example, in our simulations,we employed S = 20 in Fig. 10-(a), -(b) and S = 8 in Fig.11-(b) and Fig. 13-(e), -(f). Some specific spreading functionshave also been proposed together with their omnidirectionalspectra S F L , S E , and S R . Fung and Lee spreading functionin Fig. 4-(c) can be expressed as D ( k, θ ) = (2 π ) − + a (cid:16) − e − bk (cid:17) cos (2 θ ) (28) k, rad/m -15 -10 -5 E l e v a t i on s pe c t r u m , m / r ad / m S PM (a) k, rad/m -15 -10 -5 E l e v a t i on s pe c t r u m , m / r ad / m S J (b) k, rad/m -15 -10 -5 E l e v a t i on s pe c t r u m , m / r ad / m S FL (c) k, rad/m -15 -10 -5 E l e v a t i on s pe c t r u m , m / r ad / m S E (d) k, rad/m -15 -10 -5 E l e v a t i on s pe c t r u m , m / r ad / m S R (e)Fig. 3. Omnidirectional elevation spectra for different wind speeds ( V w ) corresponding to different models: (a) S PM , (b) S J , (c) S FL , (d) S E , and (e) S R . For S J the fetch is set to 80 km, while for S E the inverse wave age Ω is 0.84. The vertical-dashed line ( k = 370 rad/m) represents the boundarybetween the gravity and capillary waves. D PM cos cos cos cos (a) D J S = 1 S = 2.5 S = 5 S = 10 (b) D FL X-band C-band L-band k = 0.1 rad/m (c) D E X-band C-band L-band k = 0.1 rad/m (d) D R X-band C-band L-band k = 0.1 rad/m (e)Fig. 4. The angular spreading functions presented in polar coordinates at different wavenumbers: gravity-capillary waves (X-band), short gravity waves (Cand L-bands), gravity waves (0.1 rad/m). For wind depended functions, the wind speed is set to V w = 8 m/s. (a) Cosine-squared. (b) Longuet-Higgins etal. with S parameter. (c) Fung and Lee. (d) Elfouhaily et al. (e) Romeiser et al. where a = (1 − R ) / (1 + R ) π (1 − B ) (29)with B = (cid:0) /σ t (cid:1) ∞ (cid:90) k S F L ( k ) e − bk dk (30)and σ t = ∞ (cid:90) k S F L ( k ) dk (31)The parameter b = 1 . cm and R is based on Cox and Munk’sratio of slope variances as follows [89] R = 0 .
003 + 1 . × − V w . . × − V w . (32)where V w . (m/s) is the wind velocity at 12.5 m above themean sea surface.Elfouhaily et al. spreading function in Fig. 4-(d) takes theform: D ( k, θ ) = 12 π [1 + ∆ ( k ) cos (2 θ )] (33) where the value of ∆( k ) is determined as ∆ ( k ) = tanh (cid:104) a + a p ( c/c p ) . + a m ( c m /c ) . (cid:105) (34)For the way to set the remaining parameters, we refer thereader to the original paper [88].The D F L and D E functions are symmetric about π/ (Fig.4, (c) and (d)), which is more realistic for electromagneticmodeling [88]. However, the D F L function is not realistic forthe simulation of long waves [88], which is apparent from theisotropic shape of the function at k = 0 . rad/m (Fig. 4-(c)).Finally, Romeiser et al. [73] spreading function is based ona Gaussian kernel as follows D ( k, θ ) = exp (cid:18) − θ δ (cid:19) (35)with (36) δ = 0 .
14 + 0 . (cid:20) − exp (cid:18) − kV w c (cid:19)(cid:21) + 5 exp (cid:20) . − . (cid:18) V w V n (cid:19) − . (cid:18) kk n (cid:19)(cid:21) where c = 400 rad/s, and k n = 1 rad/m. This function hasmaximal directionality at wavenumber 0.1 rad/m as opposedto other k -depended functions (cf. Fig. 4-(e)). The effect of wind speed on spreading functions here hasbeen omitted. It should be noted that when wind speed isincreased, the differences between gravity-capillary and shortgravity waves are minimized. An example of this effect canbe found in [108], p. 10. All the spreading functions presentedin this section can potentially be applied in conjunction withdifferent omnidirectional spectra in order to offer exhaustivemodalities of simulating SAR images. This could also broadentheir applicability (e.g. some spectra may represent better realconditions in some geographical areas compared to others),for mitigating the distortions introduced during the SARimage simulation process for both sea waves and ship wakes.However, in our study for the S F L , S E and S R spectra kepttheir original D ( k, θ ) functions whereas S P M combined withcosine-squared function D P M and S J with Longuet-Higginset al. cosine type spreading function D J .III. S HIP WAKE MODELING
A moving vessel in the sea generates different types ofwave patterns called – wakes, which are usually classifiedinto Kelvin wakes, turbulent wake, internal-wave wakes andnarrow-V wakes [5], [109]. It is not always possible tosimultaneously identify all these wake types in a SAR image.This is due to variation in the SAR parameters, for examplethe orientation of the antenna, ship parameters, and back-ground sea condition or wind speed. It is well known thatthe most distinguishable feature of ship signatures in SARimages is the turbulent wake [97]. Although some simulationapproaches have been performed [62]–[64], [110], there isstill no complete understanding of this phenomena in SARimages. Approaches to the modeling of ship-generated internalwakes have also been formulated in some papers [23], [72],but these wakes have only been successfully visualized inSAR images where there is a shallow thermocline in the sea[109]. In the real SAR images, the narrow-V wave wakes(bright lines) presence is challenging and there are differentmechanisms explaining the reasons for their appearance ornon-appearance. This reason is usually associated with theturbulent wake or internal waves. A further complexity in themodeling of narrow-V wakes results from the fact that theyare only visible on SAR images at low wind speed ( V w < m/s) [111]. However, most sea spectra models are theoreticallyvalid only when the wind friction velocity V w ∗ > cm/s( V w > . m/s) [87], [112]. Therefore, the possibilities forthe simulation of SAR images with narrow-V wakes may belimited. The Kelvin wake is a ship wake structure which ismost often present in satellite SAR images [5]. The classicalKelvin wave system of a ship includes transverse and divergentwaves, and the cusp waves which are formed by interferenceat the wake edges (see further in [5], [112]) and which produceKelvin envelope lines [109] or cusp lines [6] on the SARimages. These are also known as Kelvin arms after theirdiscoverer Sir W. Thomson (Lord Kelvin) [113], and formopening angles of about ± Z ship ( x, y ) = Re π/ (cid:90) − π/ A ( θ ) e ik ( x cos θ + y sin θ ) dθ (37)where k = ν sec θ , ν = g/V s , V s is the velocity of the ship,and A ( θ ) is a function of the ship’s shape with parameters: B – beam (m), L – length at the waterline (m), D t – draft (m).By analogy with a sea surface elevation model (1), andreplacing ∂ Φ /∂t = − V s ∂ Φ /∂x , the Kelvin wake elevationsurface is related to the fluid velocity potential as: Z ship = V s g ∂ Φ ship ∂x (38)Here, Zilman et al.’s [6] approximated form of fluid velocitypotential is employed Φ ship ( x, y, z ) = − BLπ − F r Re ∞ (cid:90) C ( τ, x, z ) e iyτ dτ (39)where (40) C ( τ, x, z ) = (cid:0) − e − ναD t (cid:1) sin ( β − β cos β ) α / (cid:112) / τ /ν × cos (cid:0) xν √ α (cid:1) e zνα with α = (cid:16) (cid:112) τ /ν (cid:17) / β = √ α/ F r (41)and τ = ν (cid:112) sec θ − sec θ (42)The Froude number F r appearing in (39) is introduced laterin (57). Since the secant function in (42) is undefined forangles of − π/ and π/ , in implementation, a range for θ excluding the borders ( − π/ , π/ is used. Other numericalmethods for Kelvin wake simulation have been described invarious studies [59], [60], [115].IV. SAR IMAGING OF THE SEA SURFACE
The scattering of the microwave radiation from the disturbedsea surface is a complex process which includes: the physicalproperties of the surface, scanning platform geometry androughness conditions (e.g. local slopes), as well as microwavesignal properties. As mentioned in the Introduction I, there aretwo main scattering mechanisms: Bragg scattering and non-Bragg scattering (from breaking waves and specular reflection)[19]. SAR works mainly at wavelengths in the centimeters to decimeters range, and covers moderate incidence angles of ◦ ∼ ◦ for VV polarization (for HH polarization ◦ ∼ ◦ ),which is related to Bragg scattering. In this study we onlyconsider the Bragg scattering region, which is determined bythe sea surface roughness at the scale of the radar signal’swavelength. These short-scale waves are modulated in motionand orientation by long-scale waves (tilt and hydrodynamicmodulations), thus allowing the real aperture radar (RAR)to image wind- and ship-driven waves. The two-scale model(TSM) employed here is based on resonant Bragg scatteringtheory [26]–[28] and is a good compromise between thecalculation time and accuracy for approximation of scatter-ing. The Bragg scattering solution for the ensemble-averagednormalized radar cross-section (NRCS) with VV and HHpolarizations is represented as [6], [73] σ ( x, y ) = 8 πk e cos θ l W ( k Bx , k By ) | T | (43)where k e = 2 π/λ is the radar electromagnetic wavenum-ber and λ is the wavelength of the radar signal, θ l =cos − [cos( θ r − s p ) cos( s n )] is the local radar incidence an-gle, where θ r is the nominal radar incidence angle, while s n = tan − ( dZ/dx ) and s p = tan − ( dZ/dy ) are local slopesnormal and parallel to the radar look direction respectively, W ( · ) is the 2D wavenumber spectral density of the sea surfaceroughness determined as the short wave spectrum, with Braggscattering components k Bx , k By [6], T ( · ) is the complexscattering function which controls polarization of the radarsignal and depends on the relative dielectric constant e of thesea water. Here, e = 49 − . i for X-band, e = 60 − i for C-band and e = 72 − i for L-band according to [116].Taking into account the tilt and hydrodynamic modulations,NRCS is then defined by [53], [117] σ ( x, y ) = σ (cid:20) (cid:90) (cid:16) M ( k ) (cid:98) Z ( k ) e ikx + c.c. (cid:17) dk (cid:21) (44)where (cid:98) Z ( k ) is the 2D Fourier transform of the sea surfacemodel Z . The complex RAR modulation transfer function(MTF) is presented as a sum M ( k ) = M t ( k ) + M h ( k ) , wherefor the left looking SAR scanning geometry [47], [53] M t ( k ) = 4 cot θ r ± sin θ r ik y (45)with the plus sign related to VV and minus sign to HHpolarization, and M h ( k ) = − . ω k y | k | ω − iµω + µ (46)where µ is the hydrodynamic relaxation rate which dependson V w and signal frequency and was set according to [3], [54]:when V w ≤ m/s ⇒ µ = 0 . s − for X-band, µ = 0 . s − for C-band, and µ = 0 . s − for L-band, and when V w > m/s ⇒ µ = 1 . s − for X-band, µ = 0 . s − for C-band,and µ = 0 . s − for L-band. Hence, the NRCS with the MTFrepresents the RAR imaging, where MTF is at maximum forthe sea waves traveling in the range direction, and insignificantfor the azimuth moving direction. Once the RAR image iscalculated, it can be used for additional correction (specificSAR imaging mechanism) of the nonuniform displacements of water surface facets in the azimuthal direction via a velocitybunching (VB) mechanism [118], [119]. This leads to shiftedimaging of moving waves by the Doppler effect [2], [3] I ( x i , y i ) = (cid:90) (cid:90) δ ( y i − y ) σ ( x, y ) p a (cid:48) ( x, y ) × exp − π (cid:34) x i − x − RV ¯ U r ( x, y ) p a (cid:48) ( x, y ) (cid:35) dxdy (47)where I ( x i , y i ) is a speckle-free intensity SAR image, δ ( · ) is the SAR impulse response function in range direction,approximated by the Dirac delta function, R = H/ cos θ r isthe range distance between the antenna and the surface facets, H is the platform flight height, V is the platform velocity (itshould be noted that there is a difference between the platformvelocity and the footprint velocity [120]), while R/V is thealtitude-to-range ratio. ¯ U r is the mean radial velocity of thesurface facets in range direction and is written as ¯ U r = (cid:88) i (cid:88) j U rij B fij (48)where (49) B f = 2 k x ∆ x sin (cid:18) k x ∆ x (cid:19) k y ∆ y sin (cid:18) k y ∆ y (cid:19) × ωT i sin (cid:18) ωT i (cid:19) and the instantaneous radial velocity is given by U r = U z cos θ r − sin θ r ( U x sin θ w + U y cos θ w ) (50)where θ w is the angle between flight direction and wind di-rection or, if replaced by θ s , the angle between flight directionand ship moving direction. The additional filtering factor B f in(48) provides averaging of radial velocity over the integrationtime T i [47]. The large-scale orbital velocity components U x , U y and U z in (50) are calculated as the gradients of the fluidvelocity potential at the free surface ( z = 0 ) for both sea wavesand ship wake models U x = ∂ Φ ( x, y, ∂x U y = ∂ Φ ( x, y, ∂y U z = ∂ Φ ( x, y, ∂z (51)The degraded azimuthal resolution p (cid:48) a in (47) is equal to p (cid:48) a ( x, y ) = N p a (cid:20) π T i N λ ¯ A r ( x, y ) + 1 N T i τ c (cid:21) / (52)where N is the number of incoherent looks in azimuthdirection, λ is the wavelength of the radar signal, p a is thenominal single-look azimuthal resolution as follows p a = λR V T i (53)Since the integration time T i is one of the smearing factorsin SAR images and relates to the full azimuthal resolution p a , it follows that with decreasing azimuthal resolution theintegration time also decreases, which reduces the smearing effect [53]. The mean radial acceleration of the surface facets ¯ A r is calculated the same way as ¯ U r , but by inserting thelarge-scale acceleration components A x , A y , A z into (50),which in turn is calculated using (51) by replacing Φ( x, y, with ∂ Φ( x, y, /∂t . The scene coherence time τ c is relatedto the spreading of the facets motion within a SAR resolutioncell [96]. However, for simplicity, and assuming a Pierson-Moskowitz type wave spectrum, the coherence time is approx-imated as [121] τ c ≈ λV w . erf − / (cid:18) . p a V w . (cid:19) (54)It is worth noting that the velocity bunching method is notvalid for large values of the integration time T i [1], however,a recent idea about the time-divided velocity bunching modelcan help to overcome this limitation [122]. Finally, the inten-sity of the SAR image which includes a multiplicative noisecomponent can be presented as [6]: I n ( x i , y i ) = I ( x i , y i ) N ( x i , y i ) (55)where the noise sequence can be expressed with the ex-ponential distribution with a PDF of P ( N ) = exp( − N ) .Please also note that for some specific applications, variousother advanced intensity speckle models, such as Gamma, K , Gen-Rayleigh [123], Gen-Rician [124], Gen-Gamma [125]can be used. The important limitation of SAR imaging ofwaves moving in flight direction which is associated withthe velocity bunching is the azimuthal cut-off effect. Theminimal detectable wavelength of the surface waves can beapproximated as [126] λ min = RV (cid:112) H s (56)Here an increase in the R/V ratio and significant wave height H s , reduces the ability for the azimuth-traveling wave tobe imaged. In order to illustrate the azimuthal cut-off effectwe assume the S P M spectrum where the significant waveheight can be approximated [80] as H s = 0 . V w . /g and peak wavenumber k p = 0 . /H s , whence the dominantwavelength of the surface waves λ d = 2 π/k p . The resultinggraph-map is shown in Fig. 5 which is calculated for arange of V w . = 3 . − . m/s ( H s = 0 . − . m or λ d = 10 . − . m) and SAR ratios R/V = 10 − s. For example, hereinafter in Fig. 11, (a) and (b) two SARimages for R/V = 24 . s (a) and R/V = 113 . s (b) with V w . ≈ . m/s or λ d ≈ . m are presented (the lastvalues are given in terms of S P M , although S E is used inFig. 11, but the result is similar). Using the presented graph-map in Fig. 5, for the first SAR image in Fig. 11, (a) theminimal detectable wavelength λ min = 38 . m and for thesecond image in Fig. 11, (b) λ min = 176 . m. Thus, in the firstcase (a) the waves are clearly detectable because the condition λ min < λ d is satisfied, while in the second case (b) where λ min > λ d it is not. It is worth mentioning that the proposedgraph-map (Fig. 5) is suitable for a fully developed sea state.
25 50 75 100 125 150 175 200 225 250
R/V, s d , m min , m Fig. 5. The SAR azimuthal cut-off effect where the minimal detectablewavelength of the sea surface waves λ min is presented in terms of thedominant wavelength of waves λ d and the R/V ratios. The graph-map isbased on the S PM spectrum and valid for the range of V w . = 3 . − . m/s and R/V = 10 − s. V. S
IMULATION RESULTS
In this section, we show the results of our simulationexperiments involving SAR image formation, including fordifferent wave spectra. Detailed attention has been paid tospecific parameters of the radar image formation. Firstly, weconsider the factors that affect the imaging of both typesof waves, sea and ship. In general, these factors can be re-duced to hydrodynamic (surface modeling) and SAR imagingconstituents. There are hydrodynamic effects due to wavessuperposition: various sea surfaces and ship wakes with differ-ent Froude numbers are examined. The numerical simulationanalysis was performed in two steps. First, the geometry ofthe SAR platforms and scanning parameters was considered:different platforms,
R/V ratios, resolutions, incidence angles,signal frequencies, and polarizations were utilized. Second,the specific radar mechanisms in terms of ship wakes and seawaves imaging were evaluated.The comparison of different spectra has been performed intwo aspects: sea surface roughness validation and ship wakedetectability evaluation. Since the Bragg scattering directlydepends on surface roughness, we compared simulated spectrawith the well-known Cox and Munk’s probability densityfunction [90] of surface slopes. The detectability evaluation ofthe ship wake was also performed in two stages: determinationof the boundary condition, and qualitative and quantitativeassessment.Simulated SAR images were created corresponding to bothairborne and satellite platforms, with details shown in Table I.Airborne platforms were presented at low altitude (AI) andhigh altitude (AII). Satellite platforms were also separatedinto low altitude (SI) and high altitude (SII) types, withmain parameters similar to the TerraSAR-X and Sentinel-1respectively. The terms low and high should be understoodas relative to conventional Earth observation SAR scanners.It must be noted that the range-to-velocity ratio
R/V wasdetermined in accordance with the specified incidence angle, which may vary for some calculations. For simplicity, weneglected the differences of SAR resolution in range and az-imuth, which usually exist for real images, and used 2.5 m forboth resolutions. For the simulations, we utilized four differentship models with varying Froude numbers of F r = 0 . − . • Ship-I → L = 35 m, B = 5 m, D t = 2 . m, • Ship-II → L = 50 m, B = 6 . m, D t = 3 . m, • Ship-III → L = 65 m, B = 10 m, D t = 4 . m, • Ship-IV → L = 135 m, B = 25 m, D t = 10 m. A. Effect of different wind state
Two time-frozen elevation models, the sea surface andKelvin wake surface are based on a superposition of wind-and ship-generated waves and form the total sea wave-wakemodel as Z = Z sea + Z ship . The same summation principleis related to all derivatives of these models: surface slopes,fluid velocity, and acceleration components. When consideringsuperposition, it is important to select the correct wind velocityvalue for modeling. In many studies [5]–[8], [61], [71], [97],[127], [128], it is postulated that wind velocity (or conse-quently significant wave height) is one of the main factorsin ship wake visualization in SAR imagery. The Kelvin wakesystem (divergent and transverse waves) can be best observedwhen the wind velocity is about 3 m/s or less, as these levelsgive a very calm sea surface [7]. However, visualization of thecusp waves is actually more stable for relatively high windspeeds ( ∼ m/s) [7], [61], [97]. In Fig. 6, our modelingresults clearly demonstrate that at higher wind velocity theship wake visualization is significantly decreased, with largersea waves and ship wakes canceling each other out. It isalso important to note that with increasing amplitude of thesea waves, radial velocity and acceleration are also increased,which leads to greater smearing and shifting in the resultingSAR image. B. Effect of the Froude number
From Section III it follows that the main parameters whichinfluence the pattern and amplitude of the Kelvin wake arelength L , beam B , and draft D t of the modeled ship, as wellas its velocity V s . When modeling the Kelvin wake, it is usefulto apply a dimensionless quantity, the Froude number, whichis used to determine the resistance of a submerged movingvessel and is defined by F r = V s √ gL (57)From Fig. 7 it can be easily seen that with a change in F r , the amplitude of wakes also changes. However, shipsmay differ in length and velocity, but if
F r is the same,then the vessels will still produce a similar wave pattern(but not amplitude). Therefore, wakes of different ships withidentical
F r have fractal or scalable properties. For example,this property has been applied in work [61], where the wavepattern of a modeled ship was scaled to the wave pattern ofa real ship. However, a study [6] has also shown that a smallship with high Froude number and a bigger ship with lowFroude number may produce Kelvin wakes with the same wave amplitude. Also, the wake angle between cusp wavescan be decreased at large ship velocities or when
F r > . as shown in [114], with the consequence that the transversewaves become invisible [115]. In this case, the opening angle(rad) can be estimated using the following expression α w ≈ √ πF r (58) C. Effect of SAR imaging
In this subsection, we highlight the main features of SARimaging for both sea and ship waves. For most presentedmodels, we employed the S E spectrum, except for in Fig.11 where, in order to better illustrate the velocity bunchingeffect, the S E is combined with the Longuet-Higgins direc-tional function. Since a lower wind velocity is better forwake imaging [7], a wind velocity of V w = 3 . m/s wasapplied, with wind direction of D E = 45 ◦ (except whereother parameters are specified). The specific visualization ofa Kelvin wake pattern has been studied before, and it is wellknown that better imaging of the ship wake results can beobtained when local waves travel along the radar line-of-sight.Fig. 8 illustrates RAR images for three main ship headingdirections relative to the platform flight direction, and thevisualization results match the results obtained in previousstudies [7], [61]. It has traditionally been considered that HHpolarization improves the detection of wakes [7], [61], [97],[129], though the larger difference in signal intensity betweenVV and HH polarization occurs at high incidence angles andless at low incidence angles. However, since a low incidenceangle generally provides much better imaging of wake, thisreduces the importance of the difference in signal intensitybetween VV and HH polarization.The next illustration in Fig. 9 shows the effect of thedifferent radar incidence angles θ r on the detectability ofthe ship wake. It can be seen that the best visualization ofship wake (divergent and transverse waves) occurs when theincidence angle is smaller (Fig. 9-(a)) and that cusp waves aremainly visible at high incidence angles (Fig. 9-(c)). This effecthas previously been noted in studies [61], [97], [128].Although much research has been done in the field of SARimaging of vessel signatures, there is no clear understandingof the effect of radar frequency on the imaging of shipwakes. In some studies [128], [130], [131], real SAR datafor different bands were analyzed, and an overall conclusionhas been reached that ship wakes are better imaged in theX-band SAR images (e.g. TerraSAR-X) instead of the C/L-bands. It is clear that the best detectability of ship wakes isassociated with the lower altitude of TerraSAR-X, comparedwith other satellites, which reduces the R/V ratio. It isimportant to note that for an objective comparison, it wouldbe necessary that all parameters be fixed, i.e. the same satelliteplatform, the same ship parameters, the same ambient seawaves amplitude, and direction, etc., which in practice is veryhard to achieve. The comparison of different frequencies in[129] has shown that Kelvin arms are best visible at high radarfrequencies. In contrast, the simulation results in [60] haveshown better observation of ship wake at L-band rather than (a) (b) (c) (d)Fig. 6. Composite sea-ship surface elevation models Z for different velocities of wind V w as follows: (a) 3.5 m/s. (b) 6 m/s. (c) 8.5 m/s. (d) 11 m/s. ShipII , Fr = 0.5. Sea surface parameters: S E spectrum with D E = ◦ .(a) (b) (c) (d)(e) (f) (g) (h)Fig. 7. Simulated Kelvin wake models for the Ship I (c), (g);
Ship III (d), (h) and
Ship IV (a), (b), (e), (f) with different Froude numbers. Upper images3D representation of wake models. Lower images superposition of ship wake models with ambient sea waves under constant V w = 8.5 m/s.(a) (b) (c) (d)Fig. 8. Simulated NRCS images (X-band, θ r = 50 ◦ ) with RAR modulations included for different ship heading direction relative to flight direction: (a) ◦ .(b) ◦ . (c), (d) ◦ . Ship IV with Fr = 0.25; V w = 3.5 m/s and D E = ◦ . TABLE IM
AIN PARAMETERS OF SCANNING PLATFORMS FOR THE SIMULATED
SAR
IMAGES . Parameter
AI AII SI SII
Frequency f , [GHz] 9.65 (X-band), 5.3 (C-band), 1.275 (L-band)Wavelength λ , [m] 0.031 (X-band), 0.057 (C-band), 0.235 (L-band)Incidence angle θ r , [deg] 20-70Polarisation VV, HHPlatform altitude H , [km] 2.5 7 514 705Platform velocity V , [m/s] 125 160 7600 7600 R/V , [s] 21-59 47-128 72-198 99-271( θ r = 20 ◦ − ◦ )Integration time T i , [s] 0.13-0.36 (X-band) 0.29-0.79 (X-band) 0.45-1.23 (X-band) 0.61-1.68 (X-band)( θ r = 20 ◦ − ◦ ) 0.24-0.66 (C-band) 0.53-1.45 (C-band) 0.82-2.24 (C-band) 1.12-3.07 (C-band)1-2.75 (L-band) 2.19-6.01 (L-band) 3.38-9.29 (L-band) 4.64-12.75 (L-band)Coherence time τ c , [s] 0.035-0.034 (X-band)( V w = 3 . − m/s) 0.064-0.062 (C-band)0.266-0.256 (L-band)Azimuth resolution, [m] 2.5Range resolution, [m] 2.5 (a) (b) (c) (d) (e)Fig. 9. Effect of the incidence angle θ r on the SAR imaging (VV polarization) of the ship wake at different bands for AI platform: (a), (b), (c) X-band. (d)C-band. (e) L-band. Ship I with Fr = 0.35. V w = 3.5 m/s and D E = ◦ . X-band. In studies such as [5], [61], it is stated that variationin wake visualization had little dependence on radar frequency.According to our results, there is no significant difference inwake visualization for changes in radar frequency, except forat high incidence angle for L-band in Fig. 9-(e) where thevisibility of divergent and transverse waves is a little betterthan for X-band and C-band (but not cusp waves) at theequivalent angle (Fig. 9, (c), (d)) and is close to X-band atlower incidence angle in Fig. 9-(b). However, this result is nota general conclusion and further attention is needed here.In order to clearly demonstrate the effect of the SARresolution cell, the SAR parameters were fixed and realizedfor the airborne (AI) scanning platform. It can be seen thatat the detailed resolution in Fig. 10-(a) all wake componentsare easily distinguishable. However, when the resolution is de-graded (Fig. 10-(c)), the wake becomes difficult to distinguish.For the RAR imaging (Fig. 8) the transverse and divergentwaves are most visible when the ship is moving in theradar range direction, and only divergent waves exist for theazimuth direction. This effect generally also holds for theSAR imaging, although it is sometimes modified accordingto the SAR parameters relating to the properties of the seaand ship waves. For example, the transverse waves may stillappear in radar images for a ship heading parallel to azimuth direction, at the small values of
R/V ratio and lower V w (Fig. 10, (a)), mainly for airborne SAR platforms. This isdue to the contribution of the orbital motion of the gravitywaves to the VB imaging, and particularly where the minimumdiscernible azimuthal wavelength λ min in eq. (56) is lessthan the wavelength of transverse waves. However, for thesatellite platforms, transverse waves begin to show smearing,for instance, in Fig. 10-(e), and are completely invisible atthe large R/V ratio in Fig. 10-(f). This also well explainswhy for the Sentinel-1 satellite, which has a large
R/V ratio, the images mainly display cusp waves only. This hadpreviously been ascribed to the lower SAR resolution of theSentinel-1 images, but our results show that this is not the solecontributor. Looking ahead, the contribution of the sea waveamplitude factor can also influence visualization of ship wake,which is demonstrated in Fig. 14-(a), (e) and (b), (f).Importantly, the SAR imaging of moving waves is relatedto the velocity bunching mechanism. The main image degra-dation effects consist in azimuthal shifting determined as [47] ∆ X = RU r /V (59)and smearing given as δX = 2 Rσ u /V (60) (a) (b) (c)(d) (e) (f)Fig. 10. Effect of the different SAR image resolutions p a (a), (b), (c) and R/V ratios (a), (d), (e), (f) on the SAR imaging (X-band, θ r = 25 ◦ , HHpolarization) of the ship wake. Ship I with Fr = 0.5. V w = 3.5 m/s and D E = ◦ . where σ u is the standard deviation of the radial velocitieswithin a SAR resolution cell. In order to better demonstratethese effects, the simulations are presented as speckle-freeimages in Fig. 11. Two different sea states with low V w = 3 . m/s and relatively high V w = 10 m/s with waves travelingin the azimuth direction are presented in Fig. 11-(c), (d) and(a), (b), respectively, based on the S E spectrum and Longuet-Higgins directional function. While keeping the same seasurface model ( V w = 10 m/s), the smearing is minimal forAI platform (a), and increased for SII platform (b). Next,the shifting is associated with the contribution of the radialvelocity of the model’s facets, and shifting direction can beeasily determined [119]. Here an important role is played bywhether one has a right or left looking SAR configuration.When facets move toward the SAR platform, then they areimaged as shifted in the flight direction (Fig. 11-(c)), andreversely (Fig. 11, (d)). Additionally, the azimuth cut-off effectis well demonstrated on the example of sea waves in Fig. 11-(a) and (b) (an example for the ship wake is in Fig. 10-(a) and(f)). It should be noted that shifting and smearing effects existfor all simulated scenarios, and a separate illustration of theseeffects in Fig. 11 is made to facilitate understanding. D. Spectra comparison: sea surface slopes vs PDF
Many studies devoted to sea surface modeling and SARimagery simulation do not pay due attention to the correctness of the sea surface modeling step. It is assumed, for example,that the real sea surface roughness is well enough approxi-mated by a spectrum model. Therefore, we present a basicapproach to validate the spectra models. A comparison of theprobability distribution of sea surface slopes with the wellknown Cox and Munk probability density function (PDF) [89],[90] was performed. This PDF is based on the Gram-Charlierdistribution as: (61) p = (2 πσ c σ u ) − exp (cid:20) − (cid:0) ξ + η (cid:1)(cid:21)(cid:26) − c (cid:0) ξ − (cid:1) η − c (cid:0) η − η (cid:1) +124 c (cid:0) ξ − ξ + 3 (cid:1) + 14 c (cid:0) ξ − (cid:1) (cid:0) η − (cid:1) + 124 c (cid:0) η − η + 3 (cid:1) (cid:27) where ξ = Z y /σ c and η = Z x /σ u are normalized crosswindand upwind slope components, respectively (up to ξ = η =2 . ), with local surface slopes Z x and Z y (here we assume x is upwind, and y is crosswind, direction as in [132]), and rmsslope components are σ u and σ c . The latter are expressed as σ u = 3 . × − V w . (62) σ c = 0 .
003 + 1 . × − V w . (63) (a) (b) (c) (d)Fig. 11. The smearing δX and shifting ∆ X effects in speckle-free simulated SAR images (X-band, θ r = 35 ◦ . VV polarization). (a) AI platform and (b)SII platform with V w = 10 m/s, S E with D J = ◦ where parameter S = 20. (c) and (d) AI platform with V w = 3.5 m/s, S E with D J = ◦ , and ShipIII with Fr = 0.4.
The skewness coefficients for the clean sea surface are pre-sented c = 0 . − . V w . (64) c = 0 . − . V w . (65)and the peakedness coefficients are given c = 0 . c = 0 . c = 0 . (66)It is clear that with decreasing facet size the coherency of theslope model also deteriorates; this effect is explained in [92].However, if the facet size is increased, for example, to 0.5 m asin [133], the distribution of slopes will generally correspond tothe PDF model, which can be used for verification. When thefacet size is increased to a higher order, for example 3 mm size(capillary waves) then the slopes are consistent with the PDFwith sufficient accuracy [32]. On the other hand, since we usethe two-scale SAR model, and the high-frequency part (shortwaves) is modeled statistically, there is no need for a high-resolution sea surface model calculation. Therefore, here weapply the facet size of 0.5 m following [133] to test the overallconsistency of all spectra of the generated slope models to Coxand Munk’s PDF. The wind speed is set as 8 m/s and the sizeof the surfaces is 0.25 × L min sizeselection). The result of this comparison is demonstrated inFig. 12. As is shown, not all spectra give a reasonable match tothe Cox and Munk PDF. Only the wave spectra of S E and S R are matched well enough. However, for the other spectra, thisagreement was not observed where S P M and S J showed betteragreement than S F L . The results, in general, are consistentwith results shown in previous studies [34], [81], [133]. Forexample, in [81] the slope variance of different spectra has alsobeen compared with the slope variances according to the Coxand Munk’s model. Also, the agreement with Cox and Munk’sPDF likely may depend on the spreading functions, as statedin [81], [134]. It should be recalled that for the S E , S F L and S R spectra, their original spreading functions were applied,while for the S P M cosine-squared function, and for S J theLonguet-Higgins et al. function was used (Section II-C). E. Spectra comparison: SAR imaging of ship wakes
In this subsection, we applied quantitative analysis of vi-sualization of ship wakes in SAR images for different seawave spectra models. The idea behind the comparison consistsin the fact that all the spectra are created by empirical orsemi-empirical approaches for particular geographical placesand conditions. It means that even under the same modelingparameters (first of all, the velocity of wind), the amplitudeand as a consequence significant wave height H s may vary.As is mentioned in Section V-A, and, in various studies in theliterature, the ambient amplitude of sea waves can dramaticallyreduce or reversely increase the visibility of vessel signaturesin SAR images. On the other hand, different spectra modelshave been applied for SAR image simulation of ship wake[6], [58], [60], [63], [65]–[69], [71], [72]. It is important tonote that the results presented here reveal a relative differencein spectra, in terms of a superposition of sea waves withship wake and are not compared in terms of absolute SARintensity values. The comparison is presented in two stages:(i) Determination of the borderline condition; (ii) Assessmentin terms of standard statistical measures of imagery.In order to provide an evaluation of the contribution ofdifferent sea spectra to the SAR imaging of ship wake, we firstdetermined the boundary condition where the wake signaturesdue to ambient sea waves can disappear or be less noticeablein the SAR image. A recently proposed ship wake detectionmethod based on sparse regularization, and successfully testedon both type of real and simulated SAR images, [130] wasapplied. Noting that the Pierson-Moskowitz spectrum has beenthe basis of many presented spectra, it was selected as areference to generate the sea wave models within a range of V w = 3 . − m/s. As was mentioned in Section II-B, theconformity of V w for different spectra was performed by usingthe Fung and Lee method [87]. For the ship model, we selectedShip-I with F r = 0 . , and calculations are done for theairborne AI platform as the best available SAR configurationwith parameters: X-band, incidence angle θ r = 35 ◦ and VVpolarization. The speckle model is fixed to an exponentialdistribution with unit mean for all scenarios. The wind andship moving directions are tuned to ◦ (azimuth direction).We considered the boundary condition as valid if at least one -0.4 -0.2 0 0.2 0.4 Slope in the upwind direction P D F S PM CM (a) -0.5 0 0.5 Slope in the upwind direction P D F S J CM (b) -0.2 -0.1 0 0.1 0.2 Slope in the upwind direction P D F S FL CM (c) -0.5 0 0.5 Slope in the upwind direction P D F S E CM (d) -0.5 0 0.5 Slope in the upwind direction P D F S R CM (e)Fig. 12. Density comparison for different spectra (a) S PM , (b) S J with fetch = 80 km, (c) S FL , (d) S E with Ω = 0.84, and (e) S R . Dashed-lines in eachsub-figures refer to the Cox and Munk’s (CM) model. The wind velocity V w = 8 m/s and the size of the sea surface model is 0.25 × of the Kelvin arms has not been detected using the method[130]. Although, strictly speaking, the boundary condition maychange for other SAR platforms and/or ship parameters, thisdoes not reduce the significance of the experiment, since weinvestigated this approach as a starting point. The results of thedetection of ship wake are given in Fig. 13 where it is shownthat when the sea amplitude reaches H s = 1 . m ( V w = 8 . m/s), the visualization of the ship’s signature in the SAR imagebecomes difficult, which also affects the detection efficiency.After the boundary condition ( V w = 8 . m/s) was de-termined (as the point where the visualization of wakes inSAR images becomes unstable), the same ship model underthe same velocity wind speed was applied for all spectramodels. The benefit of the simulation is that it is possible togenerate SAR images and reference SAR images to evaluatethe comparison. The reference image is considered to be theSAR image with the same sea surface model but withouta ship wake model. The analysis is then carried out bycomparison, for each spectrum of SAR image (intensity valuesof superimposed sea-ship waves I s ), with the reference image(intensity values of ambient sea waves I w ) in terms of peaksignal-to-noise (PSNR), signal-to-noise (SNR), mean-squarederror (MSE), standard deviation (STD) and the StructuralSimilarity (SSIM) index [135]. Since for different spectramodels the absolute NRCS values change (as confirmed inmany studies [81]–[84]) and because we are interested inrelative changes of intensity values, all images and referenceimages were normalized before calculating the PSNR, SNR,MSE, STD, and SSIM measures. The STD measure wascalculated for the difference image, which was determinedas ∆ I = I s − I w (in Fig. 14, for better visualization onlythe positive part of ∆ I values is shown, while in Table IIthe whole STD values are shown). Also, all statistics areprovided for speckle-free images. It is useful to look at theeffect of noise in the formation of a SAR image on pairs ofspeckle-free ( I n ) and speckle ( I ) intensity images, which arepresented in Fig. 14, (a)-(b), (e)-(f), (i)-(j), (m)-(n), (q)-(r).It can be seen that the noise cancels out the details of bothsea and ship waves. In contrast to the standard image analysisinterpretation, where for example higher PSNR value indicatesbetter denoising performance, here the intensities from shipwakes are considered as a ’positive’ noise. In practice, itmeans that when the noise power is higher (bigger MSE and STD), and hence a lower PSNR value (and SNR), the wakesvisualization is improved. The same argument applies to theSSIM, where the lower global values in Table II and morenegative local values in Fig. 14, (d), (h), (l), (p), (t), correspondto better identification of ship wake signatures.According to the results in Table II, for all spectra exceptthe S F L spectrum, the better visualization of wakes on mostmeasures generally follows a decrease of the significant waveheight H s , which is consistent with previous simulation studies[6], [71], [136] or real images analysis [7], [61], [128]. Thebest performance is shown by S J model (the fetch size is25 km), where the H s is around half compared to most ofthe presented spectra, so the amplitude of the wake is notwashed out by the surrounding waves. This spectrum alsocreates a shorter wavelength of waves on the sea surfacethereby allowing better preservation of the ship wake pattern.It is also important to note that there is some variation in thecalculated values (Table II) which indicates that the visibilityof wake is also associated with the ambient sea wave pattern,which is in turn dependent on the properties of the spectra.For example, we can observe that the pair S P M - S R givessimilar results. Also, the significant wave height is much lowerfor the S F L where better wake imaging was expected. Onepossibility is that this may be attributable to the contributionof the D F L spreading function. Indeed, the D F L spreadingfunction has an isotropic shape for large-scale waves (Fig.4) and at this scale it does not realistically represent the seawave propagation [88]. This may be the cause of reduced wakevisibility in the results, but this effect requires a separate study.This means that at the boundary condition the contribution of H s is not of primary importance for modeled SAR images andother modeling parameters can also play a role. Obviously, theobtained results of the comparison of the spectra indicate thatfor different marine areas and at the same wind speed, thevisualization of wakes will differ. However, these results arevalid for the boundary condition when the ambient sea wavesare comparable in size (wavelength) and amplitude with shipwake (Fig. 13-(c)).In Fig. 14 we show the local maps of the difference image ∆ I and SSIM index. For better visualization, only the positivevalues of ∆ I are shown, which correspond to the high NRCSbackscattering signal. It can be seen that the wake patternsfor S P M , S E , and S R spectra are more smeared compared to (a) (b) (c)Fig. 13. The wake detection results for simulated SAR images (X-band, θ r = 35 ◦ , VV polarization) using method based on sparse regularization [130]under different velocities of wind: (a) V w = 3.5 m/s. (b) V w = 6 m/s. (c) V w = 8.5 m/s. Ship I with Fr = 0.5. S J and S F L . In particular, individual divergent and transversewaves are difficult to distinguish because they are mixed withambient waves. It is interesting to point out that although thestatistical results for S F L spectrum are not the best (fromTable II), the wake pattern in the difference image ∆ I isbetter compared to all but the S J spectra. Indeed, the details oftransverse and divergent waves are better visualized. However,the relative contrast between backscattering from sea wavesand ship wake is bigger for the S J spectrum (Fig. 14-(g)),which is also well supported by the objective results (Table II,MSE, and STD). The same is true for the SSIM index results,where greater difference (or lower global value in Table II) isrelated to the S J spectrum which means better visualizationand detectability of wakes. TABLE IIC
OMPARISON OF DIFFERENT SPECTRA MODELS IN SIMULATEDSPECKLE - FREE
SAR
IMAGERY BY VARIOUS MEASURES . ∗ Spectrum H s PSNR SNR MSE STD SSIM S P M S J S F L S E S R ∗ Since ship wake considered as a positive noise, the statistics must beinterpreted inverted, explanation in the text above.
VI. C
ONCLUSION
In this paper, we reviewed the state-of-the-art and presenteda comprehensive SAR imagery simulation framework forcomplex sea-ship waves scenario. This study was performedon the back of decades of research and knowledge that spansa very diverse range of fields including the linear theory [79]and stochastic modeling [80] of the sea surface, Kelvin shipwake modeling [6], [59], [60], [63], [113] and the theory ofSAR imaging oceans [26]–[28], [73], [117], [119]. In contrastto existing simulation studies for both wind- and ship-induced waves, we have extended our framework to include the analy-sis of hydrodynamical effects as well as SAR imaging effects.For the first time, the most common SAR imaging phenomenarelated to ship wakes have been integrated and presented. Inparticular, we investigated different SAR parameters, includingfrequencies (X, C, and L-band), incidence angles, signal polar-izations (VV, HH), image resolutions (2.5, 5 and 10 m), fourdifferent platform configurations (two for airborne and two forspaceborne), and velocity bunching effects (azimuthal cut-off,shifting and smearing). Numerical simulations were performedfor five sea wave spectra ( S P M , S J , S F L , S E , and S R ) andfor four different ship models. The experimental simulationresults demonstrated a fine agreement with the theory. Theproposed simulator is implemented in MATLAB.Our analysis of ship wake detectability for evaluating thevarious spectra supports the accepted rule that wake ob-servability decreases with the increase of significant waveheight. However, there is some variation, for example thelower detectability for the S F L spectrum at lower H s , whichmeans that other modeling parameters could contribute towake visualization. The important conclusion is not howeverthe perceived differences among spectra, but the very existenceof a difference that was not investigated earlier in the SARsimulation of composite sea-ship scenes.The presented versatile SAR imaging methodology may bemore convenient as it allows the selection of different spectrafor different tasks, for example, when considering differentmaritime regions, or to obtain greater consistency betweenspectrum model and real sea surface roughness. Our simulationframework can be employed for a better understanding of thevisibility and detectability of ship wakes in real SAR images.Work along these lines has been initiated, in particular for shipwake detection [130], despeckling of simulated SAR images[137], and for studying the effect of sea state [138].Finally, we would like to outline some interesting directionsfor future research. Although it is believed that wind directionhas no effect on radar wake imaging, see for example [61],this conclusion is based on satellite observations and is mainlytrue for cusp waves only. In this respect, not only the directionbut the angular spreading function D ( k, θ ) could possibly (a) (b) (c) (d)(e) (f) (g) (h)(i) (j) (k) (l)(m) (n) (o) (p)(q) (r) (s) (t)Fig. 14. Simulated SAR images (X-band, θ r = 35 ◦ , VV polarization) for different spectra models: (a)-(d) S PM . (e)-(h) S J . (i)-(l) S FL . (m)-(p) S E . (q)-(t) S R . For S J the fetch is 25 km, for S E the inverse wave age Ω is 0.84. For all spectra V w = 8.5 m/s and Ship I with Fr = 0.5. influence wave-wave interaction, which is directly related tothe imaging of a vessel’s signature. Thus, a more detailedinvestigation in this respect is needed. Another interestingdirection is a comparison of simulated SAR images of shipwakes with real SAR images, for example using spectraldecomposition approaches [77]. There have been good resultswhere the simulated SAR spectra of ocean waves have beencompared by spectral analysis with the real SAR measure-ments [1], [3], [96]. However, the methodological basis fordirect comparisons of simulated ship wake images with realimages has still not been fully explored [6], [58] and newapproaches can be developed in this direction, based on thesimulation framework presented in our study.A CKNOWLEDGMENT
This work was supported by the Engineering and Phys-ical Sciences Research Council (EPSRC) under grantEP/R009260/1 (AssenSAR). The authors would like to thankProfessor G. Zilman for scientific discussion.R
EFERENCES[1] L. Zurk and W. J. Plant, “Comparison of actual and simulated syntheticaperture radar image spectra of ocean waves,”
Journal of GeophysicalResearch: Oceans , vol. 101, no. C4, pp. 8913–8931, 1996.[2] C. Br¨uning, W. Alpers, and K. Hasselmann, “Monte-carlo simulationstudies of the nonlinear imaging of a two dimensional surface wavefield by a synthetic aperture radar,”
International Journal of RemoteSensing , vol. 11, no. 10, pp. 1695–1727, 1990.[3] C. Br¨uning, R. Schmidt, and W. Alpers, “Estimation of the oceanwave–radar modulation transfer function from synthetic aperture radarimagery,”
Journal of Geophysical Research: Oceans , vol. 99, no. C5,pp. 9803–9815, 1994.[4] D. P. Kasilingam and O. H. Shemdin, “Models for synthetic apertureradar imaging of the ocean: A comparison,”
Journal of GeophysicalResearch: Oceans , vol. 95, no. C9, pp. 16 263–16 276, 1990.[5] W. G. Pichel, P. Clemente-Col´on, C. Wackerman, and K. S. Friedman,“Ship and wake detection,” pp. 277–303, 2004.[6] G. Zilman, A. Zapolski, and M. Marom, “On detectability of a ship’skelvin wake in simulated sar images of rough sea surface,”
IEEETransactions on Geoscience and Remote Sensing , vol. 53, no. 2, pp.609–619, 2014.[7] A. Panico, M. D. Graziano, and A. Renga, “Sar-based vessel velocityestimation from partially imaged kelvin pattern,”
IEEE Geoscience andRemote Sensing Letters , vol. 14, no. 11, pp. 2067–2071, 2017.[8] B. Tings, A. Pleskachevsky, D. Velotto, and S. Jacobsen, “Extension ofship wake detectability model for non-linear influences of parametersusing satellite based x-band synthetic aperture radar,”
Remote Sensing ,vol. 11, no. 5, p. 563, 2019.[9] R. Kodis, “A note on the theory of scattering from an irregular surface,”
IEEE Transactions on Antennas and Propagation , vol. 14, no. 1, pp.77–82, 1966.[10] D. Barrick, “Rough surface scattering based on the specular pointtheory,”
IEEE Transactions on Antennas and Propagation , vol. 16,no. 4, pp. 449–454, 1968.[11] S. L. Broschat, “The phase perturbation approximation for roughsurface scattering from a pierson-moskowitz sea surface,”
IEEE trans-actions on geoscience and remote sensing , vol. 31, no. 1, pp. 278–283,1993.[12] K.-S. Chen and A. K. Fung, “A comparison of backscattering modelsfor rough surfaces,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 33, no. 1, pp. 195–200, 1995.[13] A. K. Fung, Z. Li, and K.-S. Chen, “Backscattering from a randomlyrough dielectric surface,”
IEEE Transactions on Geoscience and remotesensing , vol. 30, no. 2, pp. 356–369, 1992.[14] E. Bahar and Y.-F. Li, “Scattering cross sections for non-gaussian roughsurfaces: unified full wave approach,”
IEEE transactions on antennasand propagation , vol. 39, no. 12, pp. 1777–1781, 1991. [15] M. P. Clarizia, C. Gommenginger, M. Di Bisceglie, C. Galdi, and M. A.Srokosz, “Simulation of l-band bistatic returns from the ocean surface:A facet approach with application to ocean gnss reflectometry,”
IEEETransactions on Geoscience and Remote Sensing , vol. 50, no. 3, pp.960–971, 2011.[16] A. Voronovich, “Small-slope approximation for electromagnetic wavescattering at a rough interface of two dielectric half-spaces,”
Waves inrandom media , vol. 4, no. 3, pp. 337–368, 1994.[17] A. Voronovich and V. Zavorotny, “Theoretical model for scattering ofradar signals in k u-and c-bands from a rough sea surface with breakingwaves,”
Waves in Random Media , vol. 11, no. 3, pp. 247–269, 2001.[18] T. M. Elfouhaily, C.-A. Gu´erin et al. , “A critical survey of approximatescattering wave theories from random rough surfaces,”
Waves inrandom media , vol. 14, no. 4, pp. R1–R40, 2004.[19] V. Kudryavtsev, D. Hauser, G. Caudal, and B. Chapron, “A semiem-pirical model of the normalized radar cross-section of the sea surface1. background model,”
Journal of Geophysical Research: Oceans , vol.108, no. C3, pp. FET–2, 2003.[20] L. Brekhovskikh, “Wave diffraction by a rough surface,”
Experimentaland Theoretical Physics Journal , vol. 23, no. 3, p. 9, 1952.[21] O. Phillips, “Radar returns from the sea surface—bragg scattering andbreaking waves,”
Journal of physical oceanography , vol. 18, no. 8, pp.1065–1074, 1988.[22] P. A. Hwang and F. Fois, “Surface roughness and breaking waveproperties retrieved from polarimetric microwave radar backscattering,”
Journal of Geophysical Research: Oceans , vol. 120, no. 5, pp. 3640–3657, 2015.[23] J. Li, M. Zhang, W. Fan, and D. Nie, “Facet-based investigation onmicrowave backscattering from sea surface with breaking waves: Seaspikes and sar imaging,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 55, no. 4, pp. 2313–2325, 2017.[24] X. Zhang, X. Su, and Z.-S. Wu, “A volume-surface composite scat-tering model for nonlinear ocean surface with breaking waves andfoam layers under high wind conditions,”
Progress In ElectromagneticsResearch , vol. 85, pp. 125–142, 2019.[25] S. O. Rice, “Reflection of electromagnetic waves from slightly roughsurfaces,”
Communications on pure and applied mathematics , vol. 4,no. 2-3, pp. 351–378, 1951.[26] J. Wright, “A new model for sea clutter,”
IEEE Transactions onantennas and propagation , vol. 16, no. 2, pp. 217–223, 1968.[27] G. R. Valenzuela, “Theories for the interaction of electromagnetic andoceanic waves—a review,”
Boundary-Layer Meteorology , vol. 13, no.1-4, pp. 61–85, 1978.[28] K. Hasselmann, R. Raney, W. Plant, W. Alpers, R. Shuchman, D. R.Lyzenga, C. Rufenach, and M. Tucker, “Theory of synthetic apertureradar ocean imaging: A marsen view,”
Journal of Geophysical Re-search: Oceans , vol. 90, no. C3, pp. 4659–4686, 1985.[29] F. Bass, I. Fuks, A. Kalmykov, I. Ostrovsky, and A. Rosenberg, “Veryhigh frequency radiowave scattering by a disturbed sea surface part ii:Scattering from an actual sea surface,”
IEEE Transactions on Antennasand Propagation , vol. 16, no. 5, pp. 560–568, 1968.[30] G. Valenzuela, “Scattering of electromagnetic waves from a tiltedslightly rough surface,”
Radio Science , vol. 3, no. 11, pp. 1057–1066,1968.[31] R. Romeiser, A. Schmidt, and W. Alpers, “A three-scale compositesurface model for the ocean wave–radar modulation transfer function,”
Journal of Geophysical Research: Oceans , vol. 99, no. C5, pp. 9785–9801, 1994.[32] S. Kay, J. Hedley, S. Lavender, and A. Nimmo-Smith, “Light transferat the ocean surface modeled using high resolution sea surface realiza-tions,”
Optics express , vol. 19, no. 7, pp. 6493–6504, 2011.[33] W.-Q. Jiang, M. Zhang, P.-B. Wei, and D. Nie, “Spectral decompositionmodeling method and its application to em scattering calculation oflarge rough surface with ssa method,”
IEEE Journal of Selected Topicsin Applied Earth Observations and Remote Sensing , vol. 8, no. 4, pp.1848–1854, 2015.[34] L. Linghu, J. Wu, B. Huang, Z. Wu, and M. Shi, “Gpu-acceleratedmassively parallel computation of electromagnetic scattering of a time-evolving oceanic surface model i: Time-evolving oceanic surface gener-ation,”
IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing , vol. 11, no. 8, pp. 2752–2762, 2018.[35] L. Linghu, J. Wu, Z. Wu, and X. Wang, “Parallel computation of embackscattering from large three-dimensional sea surface with cuda,”
Sensors , vol. 18, no. 11, p. 3656, 2018.[36] L. Linghu, J. Wu, Z. Wu, G. Jeon, and X. Wang, “Gpu-acceleratedcomputation of time-evolving electromagnetic backscattering field from large dynamic sea surfaces,” IEEE Transactions on Industrial Informat-ics , vol. 16, no. 5, pp. 3187–3197, 2019.[37] Y. Wan, X. Zhang, Y. Dai, and X. Shi, “Research on a method forsimulating multiview ocean wave synchronization data by networkedsar satellites,”
Journal of Marine Science and Engineering , vol. 7, no. 6,p. 180, 2019.[38] G. Di Martino, A. Iodice, and D. Riccio, “Closed-form anisotropic po-larimetric two-scale model for fast evaluation of sea surface backscat-tering,”
IEEE Transactions on Geoscience and Remote Sensing , vol. 57,no. 8, pp. 6182–6194, 2019.[39] X. Ye, M. Lin, Q. Song, and Q. Wang, “The optimized small incidenceangle setting of a composite bragg scattering model and its applicationto sea surface wind speed retrieval,”
IEEE Journal of Selected Topicsin Applied Earth Observations and Remote Sensing , vol. 13, pp. 1248–1257, 2020.[40] D. Li, Z. Zhao, C. Qi, Y. Huang, Y. Zhao, and Z. Nie, “An improvedtwo-scale model for electromagnetic backscattering from sea surface,”
IEEE Geoscience and Remote Sensing Letters , 2019.[41] T. Wang and C. Tong, “An improved facet-based tsm for electromag-netic scattering from ocean surface,”
IEEE Geoscience and RemoteSensing Letters , vol. 15, no. 5, pp. 644–648, 2018.[42] M. Zhang, H. Chen, and H.-C. Yin, “Facet-based investigation on emscattering from electrically large sea surface with two-scale profiles:Theoretical model,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 49, no. 6, pp. 1967–1975, 2011.[43] S. He, X. Wang, R. Xia, W. Jin et al. , “Polarimetric infrared imagingsimulation of a synthetic sea surface with mie scattering,”
Appliedoptics , vol. 57, no. 7, pp. B150–B159, 2018.[44] P.-B. Wei, M. Zhang, D. Nie, and Y.-C. Jiao, “Improvement of ssaapproach for numerical simulation of sea surface scattering at highmicrowave bands,”
Remote Sensing , vol. 10, no. 7, p. 1021, 2018.[45] J. Li, L. Wang, M. Zhang, W. Fan, and P. Wei, “Effective approachto generate electrically large rough surfaces and the application toem scattering problems,”
IET Microwaves, Antennas & Propagation ,vol. 14, no. 4, pp. 233–240, 2019.[46] W. Alpers, “Monte carlo simulations for studying the relationshipbetween ocean wave and synthetic aperture radar image spectra,”
Journal of Geophysical Research: Oceans , vol. 88, no. C3, pp. 1745–1759, 1983.[47] D. R. Lyzenga, “Numerical simulation of synthetic aperture radarimage spectra for ocean waves,”
IEEE Transactions on Geoscienceand Remote Sensing , no. 6, pp. 863–872, 1986.[48] R. O. Harger and C. E. Korman, “Comparisons of simulated and actualsynthetic aperture radar gravity wave images,”
Journal of GeophysicalResearch: Oceans , vol. 93, no. C11, pp. 13 867–13 882, 1988.[49] P. W. Vachon, R. K. Raney, and W. Emergy, “A simulation for space-borne sar imagery of a distributed, moving scene,”
IEEE Transactionson Geoscience and Remote Sensing , vol. 27, no. 1, pp. 67–78, 1989.[50] L. Shemer and E. Kit, “Simulation of an interferometric syntheticaperture radar imagery of an ocean system consisting of a current anda monochromatic wave,”
Journal of Geophysical Research: Oceans ,vol. 96, no. C12, pp. 22 063–22 073, 1991.[51] M. Bao, C. Bruning, and W. Alpers, “Simulation of ocean wavesimaging by an along-track interferometric synthetic aperture radar,”
IEEE transactions on geoscience and remote sensing , vol. 35, no. 3,pp. 618–631, 1997.[52] G. Franceschetti, M. Migliaccio, and D. Riccio, “On ocean sar rawsignal simulation,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 36, no. 1, pp. 84–100, 1998.[53] J. Schulz-Stellenfleth and S. Lehner, “Ocean wave imaging using anairborne single pass across-track interferometric sar,”
IEEE Transac-tions on Geoscience and Remote Sensing , vol. 39, no. 1, pp. 38–45,2001.[54] F. Nunziata, A. Gambardella, and M. Migliaccio, “An educational sarsea surface waves simulator,”
International Journal of Remote Sensing ,vol. 29, no. 11, pp. 3051–3066, 2008.[55] T. Yoshida and C.-K. Rheem, “Sar image simulation in the time domainfor moving ocean surfaces,”
Sensors , vol. 13, no. 4, pp. 4450–4467,2013.[56] B. Liu and Y. He, “Sar raw data simulation for ocean scenes usinginverse omega-k algorithm,”
IEEE Transactions on Geoscience andRemote Sensing , vol. 54, no. 10, pp. 6151–6169, 2016.[57] F. M. Santos, A. L. Santos, N. Violante-Carvalho, L. M. Carvalho,Y. O. Brasil-Correa, J. Portilla-Yandun, and R. Romeiser, “A simulatorof synthetic aperture radar (sar) image spectra: the applications onoceanswell waves,”
International Journal of Remote Sensing , vol. 42,no. 8, pp. 2981–3001, 2021. [58] J. K. Tunaley, E. H. Buller, K. Wu, and M. T. Rey, “The simulation ofthe sar image of a ship wake,”
IEEE transactions on geoscience andremote sensing , vol. 29, no. 1, pp. 149–156, 1991.[59] L. Shemer, L. Kagan, and G. Zilman, “Simulation of ship wakes imageby an along-track interferometric sar,”
International Journal of RemoteSensing , vol. 17, no. 18, pp. 3577–3597, 1996.[60] K. Oumansour, Y. Wang, and J. Saillard, “Multifrequency sar observa-tion of a ship wake,”
IEE Proceedings-Radar, Sonar and Navigation ,vol. 143, no. 4, pp. 275–280, 1996.[61] I. Hennings, R. Romeiser, W. Alpers, and A. Viola, “Radar imaging ofkelvin arms of ship wakes,”
International Journal of Remote Sensing ,vol. 20, no. 13, pp. 2519–2543, 1999.[62] A. Wang and M. Zhu, “Simulation of ship generated turbulent andvortical wake imaging by sar,”
Journal of Electronics (China) , vol. 21,no. 1, pp. 64–71, 2004.[63] A. Arnold-Bos, A. Khenchaf, and A. Martin, “Bistatic radar imagingof the marine environment—part ii: Simulation and results analysis,”
IEEE Transactions on Geoscience and Remote Sensing , vol. 45, no. 11,pp. 3384–3396, 2007.[64] R.-Q. Sun, G. Luo, M. Zhang, and C. Wang, “Electromagnetic scatter-ing model of the kelvin wake and turbulent wake by a moving ship,”
Waves in Random and Complex Media , vol. 21, no. 3, pp. 501–514,2011.[65] H. Chen, M. Zhang, and H.-C. Yin, “Facet-based simulator for bistaticscattering of the maritime scene with electrically large ships: Slopesummation facet model,”
International journal of remote sensing ,vol. 33, no. 21, pp. 6927–6941, 2012.[66] R.-Q. Sun, M. Zhang, C. Wang, and X.-F. Yuan, “Scattering analysisfor ship kelvin wakes on two-dimensional linear and nonlinear seasurfaces,”
Progress In Electromagnetics Research , vol. 52, pp. 405–423, 2013.[67] N. Li, M. Zhang, D. Nie, and R.-Q. Sun, “Study of electromagneticcomposite scattering from a ship-generated internal wave wake andits underlying sea surface,”
Waves in Random and Complex Media ,vol. 25, no. 4, pp. 628–643, 2015.[68] Y. Zhao, M. Zhang, Y.-W. Zhao, and X.-P. Geng, “A bistatic sarimage intensity model for the composite ship–ocean scene,”
IEEETransactions on Geoscience and Remote Sensing , vol. 53, no. 8, pp.4250–4258, 2015.[69] W.-Q. Jiang, M. Zhang, D. Nie, and R.-Q. Sun, “Ship motion estimationfrom polarized doppler spectra from ship wakes on two-dimensionalsea surfaces,”
Waves in Random and Complex Media , vol. 26, no. 3,pp. 284–300, 2016.[70] A. Fujimura, A. Soloviev, S. H. Rhee, and R. Romeiser, “Coupledmodel simulation of wind stress effect on far wakes of ships insar images,”
IEEE Transactions on Geoscience and Remote Sensing ,vol. 54, no. 5, pp. 2543–2551, 2016.[71] J.-K. Wang, M. Zhang, J.-L. Chen, and Z. Cai, “Application of facetscattering model in sar imaging of sea surface waves with kelvin wake,”
Progress In Electromagnetics Research , vol. 67, pp. 107–120, 2016.[72] J.-K. Wang, M. Zhang, Z.-H. Cai, and J.-L. Chen, “Sar imagingsimulation of ship-generated internal wave wake in stratified ocean,”
Journal of Electromagnetic Waves and applications , vol. 31, no. 11-12,pp. 1101–1114, 2017.[73] R. Romeiser, W. Alpers, and V. Wismann, “An improved compositesurface model for the radar backscattering cross section of the oceansurface: 1. theory of the model and optimization/validation by scat-terometer data,”
Journal of Geophysical Research: Oceans , vol. 102,no. C11, pp. 25 237–25 250, 1997.[74] W. J. Plant, “A stochastic, multiscale model of microwave backscatterfrom the ocean,”
Journal of Geophysical Research: Oceans , vol. 107,no. C9, pp. 3–1, 2002.[75] P. Yang and L. Guo, “Polarimetric scattering from two-dimensionaldielectric rough sea surface with a ship-induced kelvin wake,”
Inter-national Journal of Antennas and Propagation , vol. 2016, 2016.[76] T. Qiao, L. Tsang, D. Vandemark, S. H. Yueh, T.-H. Liao, F. Nouguier,and B. Chapron, “Sea surface radar scattering at l-band based onnumerical solution of maxwell’s equations in 3-d (nmm3d),”
IEEETransactions on Geoscience and Remote Sensing , vol. 56, no. 6, pp.3137–3147, 2018.[77] Y.-X. Sun, P. Liu, and Y.-Q. Jin, “Ship wake components: isolation,reconstruction, and characteristics analysis in spectral, spatial, andterrasar-x image domains,”
IEEE Transactions on Geoscience andRemote Sensing , vol. 56, no. 7, pp. 4209–4224, 2018.[78] M. Zhang and J. Li, “Numerical simulation and analyses of sarimages from ship wakes,” in . IEEE, 2019, pp. 1–2. [79] L. H. Holthuijsen, Waves in oceanic and coastal waters . Cambridgeuniversity press, 2010.[80] M. K. Ochi,
Ocean waves: the stochastic approach . CambridgeUniversity Press, 2005, vol. 6.[81] H. Zheng, A. Khenchaf, Y. Wang, H. Ghanmi, Y. Zhang, and C. Zhao,“Sea surface monostatic and bistatic em scattering using ssa-1 anduavsar data: Numerical evaluation and comparison using different seaspectra,”
Remote Sensing , vol. 10, no. 7, p. 1084, 2018.[82] M. Ryabkova, V. Karaev, J. Guo, and Y. Titchenko, “A review of wavespectrum models as applied to the problem of radar probing of the seasurface,”
Journal of Geophysical Research: Oceans , vol. 124, no. 10,pp. 7104–7134, 2019.[83] D. Xie, K.-S. Chen, and X. Yang, “Effects of wind wave spectraon radar backscatter from sea surface at different microwave bands:A numerical study,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 57, no. 9, pp. 6325–6334, 2019.[84] B. Zhang, X. Zhao, W. Perrie, and V. Kudryavtsev, “On modeling ofquad-polarization radar scattering from the ocean surface with breakingwaves,”
Journal of Geophysical Research: Oceans , vol. 125, no. 8, p.e2020JC016319, 2020.[85] W. J. Pierson Jr and L. Moskowitz, “A proposed spectral form for fullydeveloped wind seas based on the similarity theory of sa kitaigorod-skii,”
Journal of geophysical research , vol. 69, no. 24, pp. 5181–5190,1964.[86] K. Hasselmann, T. Barnett, E. Bouws, H. Carlson, D. Cartwright,K. Enke, J. Ewing, H. Gienapp, D. Hasselmann, P. Kruseman et al. ,“Measurements of wind-wave growth and swell decay during the jointnorth sea wave project (jonswap),”
Erg¨anzungsheft 8-12 , 1973.[87] A. Fung and K. Lee, “A semi-empirical sea-spectrum model for scat-tering coefficient estimation,”
IEEE Journal of Oceanic Engineering ,vol. 7, no. 4, pp. 166–176, 1982.[88] T. Elfouhaily, B. Chapron, K. Katsaros, and D. Vandemark, “A unifieddirectional spectrum for long and short wind-driven waves,”
Journal ofGeophysical Research: Oceans , vol. 102, no. C7, pp. 15 781–15 796,1997.[89] C. Cox and W. Munk, “Measurement of the roughness of the seasurface from photographs of the sun’s glitter,”
Josa , vol. 44, no. 11,pp. 838–850, 1954.[90] ——, “Slopes of the sea surface deduced from photographs of sunglitter,” 1956.[91] J. Wang and X. Xu, “Doppler simulation and analysis for 2-d seasurfaces up to ku-band,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 54, no. 1, pp. 466–478, 2015.[92] C. D. Mobley, “Polarized reflectance and transmittance properties ofwindblown sea surfaces,”
Applied Optics , vol. 54, no. 15, pp. 4828–4849, 2015.[93] ——, “Modeling sea surfaces: A tutorial on fourier transform tech-niques,” 2016.[94] C. Rufenach and W. Alpers, “Imaging ocean waves by syntheticaperture radars with long integration times,”
IEEE Transactions onAntennas and Propagation , vol. 29, no. 3, pp. 422–428, 1981.[95] W. J. Plant, E. A. Terray, R. A. Petitt Jr, and W. C. Keller, “Thedependence of microwave backscatter from the sea on illuminatedarea: Correlation times and lengths,”
Journal of Geophysical Research:Oceans , vol. 99, no. C5, pp. 9705–9723, 1994.[96] W. R. Alpers and C. Bruening, “On the relative importance of motion-related contributions to the sar imaging mechanism of ocean surfacewaves,”
IEEE transactions on Geoscience and Remote Sensing , no. 6,pp. 873–885, 1986.[97] M. D. Graziano, M. Grasso, and M. D’Errico, “Performance analysis ofship wake detection on sentinel-1 sar images,”
Remote Sensing , vol. 9,no. 11, p. 1107, 2017.[98] M. B. Kanevsky,
Radar imaging of the ocean waves . Elsevier, 2008.[99] S. R. Massel,
Ocean surface waves: their physics and prediction .World scientific, 2017, vol. 45.[100] H. Mitsuyasu, F. Tasai, T. Suhara, S. Mizuno, M. Ohkusu, T. Honda,and K. Rikiishi, “Observation of the power spectrum of ocean wavesusing a cloverleaf buoy,”
Journal of Physical Oceanography , vol. 10,no. 2, pp. 286–296, 1980.[101] J. R. Apel, “An improved model of the ocean surface wave vectorspectrum and its effects on radar backscatter,”
Journal of GeophysicalResearch: Oceans , vol. 99, no. C8, pp. 16 269–16 291, 1994.[102] N. W. Guinard, J. T. Ransone Jr, and J. C. Daley, “Variation of thenrcs of the sea with increasing roughness,”
Journal of GeophysicalResearch , vol. 76, no. 6, pp. 1525–1538, 1971.[103] N. E. Huang, C.-C. Tung, and S. R. Long, “6. wave spectra,”
Oceanengineering science , p. 197, 1990. [104] H. Mitsuyasu, F. Tasai, T. Suhara, S. Mizuno, M. Ohkusu, T. Honda,and K. Rikiishi, “Observations of the directional spectrum of oceanwavesusing a cloverleaf buoy,”
Journal of Physical Oceanography ,vol. 5, no. 4, pp. 750–760, 1975.[105] Y. Du, X. Yang, K.-S. Chen, W. Ma, and Z. Li, “An improved spectrummodel for sea surface radar backscattering at l-band,”
Remote Sensing ,vol. 9, no. 8, p. 776, 2017.[106] K. Nagai, “Computation of refraction and diffraction of irregular sea,”
Report of the Port and Harbour Research Institute , vol. 11, no. 2, 1972.[107] D. E. Hasselmann, M. Dunckel, and J. Ewing, “Directional wave spec-tra observed during jonswap 1973,”
Journal of physical oceanography ,vol. 10, no. 8, pp. 1264–1280, 1980.[108] X. Zhou, J. Chong, H. Bi, X. Yu, Y. Shi, and X. Ye, “Directionalspreading function of the gravity-capillary wave spectrum derived fromradar observations,”
Remote Sensing , vol. 9, no. 4, p. 361, 2017.[109] J. D. Lyden, R. R. Hammond, D. R. Lyzenga, and R. Shuchman,“Synthetic aperture radar imaging of surface ship wakes,”
Journal ofGeophysical Research: Oceans , vol. 93, no. C10, pp. 12 293–12 303,1988.[110] P. Liu, W.-J. Ren, and Y.-Q. Jin, “Sar image simulation of ship turbulentwake using semi-empirical energy spectrum,” in .IEEE, 2020, pp. 47–49.[111] T. Soomere, “Nonlinear Components of Ship Wake Waves,”
AppliedMechanics Reviews , vol. 60, no. 3, pp. 120–138, 05 2007.[112] A. Arnold-Bos, A. Khenchaf, and A. Martin, “Bistatic radar imagingof the marine environment—part i: Theoretical background,”
IEEETransactions on Geoscience and Remote Sensing , vol. 45, no. 11, pp.3372–3383, 2007.[113] W. Thomson, “On ship waves,”
Proceedings of the Institution ofMechanical Engineers , vol. 38, no. 1, pp. 409–434, 1887.[114] M. Rabaud and F. Moisy, “Ship wakes: Kelvin or mach angle?”
Physical review letters , vol. 110, no. 21, p. 214503, 2013.[115] ——, “Narrow ship wakes and wave drag for planing hulls,”
Oceanengineering , vol. 90, pp. 34–38, 2014.[116] M. A. Donelan and W. J. Pierson Jr, “Radar scattering and equilibriumranges in wind-generated waves with application to scatterometry,”
Journal of Geophysical Research: Oceans , vol. 92, no. C5, pp. 4971–5029, 1987.[117] W. R. Alpers, D. B. Ross, and C. L. Rufenach, “On the detectability ofocean surface waves by real and synthetic aperture radar,”
Journal ofGeophysical Research: Oceans , vol. 86, no. C7, pp. 6481–6498, 1981.[118] W. Alpers and C. Rufenach, “The effect of orbital motions on syn-thetic aperture radar imagery of ocean waves,”
IEEE Transactions onAntennas and Propagation , vol. 27, no. 5, pp. 685–690, 1979.[119] D. Lyzenga, R. Shuchman, J. Lyden, and C. Rufenach, “Sar imagingof waves in water and ice: Evidence for velocity bunching,”
Journal ofGeophysical Research: Oceans , vol. 90, no. C1, pp. 1031–1036, 1985.[120] P. W. Vachon, H. E. Krogstad, and J. S. Paterson, “Airborne andspaceborne synthetic aperture radar observations of ocean waves,”
Atmosphere-Ocean , vol. 32, no. 1, pp. 83–112, 1994.[121] S. J. Frasier and A. J. Camps, “Dual-beam interferometry for oceansurface current vector mapping,”
IEEE Transactions on Geoscience andRemote Sensing , vol. 39, no. 2, pp. 401–414, 2001.[122] J.-W. Rim and I.-S. Koh, “Sar image generation of ocean surface usingtime-divided velocity bunching model,”
Journal of ElectromagneticEngineering and Science , vol. 19, no. 2, pp. 82–88, 2019.[123] E. E. Kuruoglu and J. Zerubia, “Modeling SAR images with a gen-eralization of the Rayleigh distribution,”
IEEE Transactions on ImageProcessing , vol. 13, no. 4, pp. 527–533, 2004.[124] O. Karakus¸, E. E. Kuruoglu, and A. Achim, “A generalized Gaussianextension to the Rician distribution for SAR image modeling,” arXivpreprint arXiv:2006.08300 , 2020.[125] H.-C. Li, W. Hong, Y.-R. Wu, and P.-Z. Fan, “An efficient and flexiblestatistical model based on generalized gamma distribution for amplitudeSAR images,”
IEEE Transactions on Geoscience and Remote Sensing ,vol. 48, no. 6, pp. 2711–2722, 2010.[126] P. Vachon, F. Monaldo, B. Holt, S. Lehner, C. Jackson, and J. Apel,“Ocean surface waves and spectra,”
Synthetic Aperture Radar MarineUser’s Manual , pp. 139–169, 2004.[127] A. M. Reed and J. H. Milgram, “Ship wakes and their radar images,”
Annual Review of Fluid Mechanics , vol. 34, no. 1, pp. 469–502, 2002.[128] B. Tings and D. Velotto, “Comparison of ship wake detectability on c-band and x-band sar,”
International Journal of Remote Sensing , vol. 39,no. 13, pp. 4451–4468, 2018. [129] W. Alpers, R. Romeiser, and I. Hennings, “On the radar imagingmechanism of kelvin arms of ship wakes,” in IGARSS’98. Sensingand Managing the Environment. 1998 IEEE International Geoscienceand Remote Sensing. Symposium Proceedings.(Cat. No. 98CH36174) ,vol. 4. IEEE, 1998, pp. 1932–1934.[130] O. Karakus¸, I. Rizaev, and A. Achim, “Ship wake detection in SARimages via sparse regularization,”
IEEE Transactions on Geoscienceand Remote Sensing , vol. 58, no. 3, pp. 1665–1677, 2020.[131] D. Velotto, C. Bentes, B. Tings, and S. Lehner, “First comparison ofsentinel-1 and terrasar-x data in the framework of maritime targetsdetection: South italy case,”
IEEE Journal of Oceanic Engineering ,vol. 41, no. 4, pp. 993–1006, 2016.[132] H. Chan and A. Fung, “A theory of sea scatter at large incident angles,”
Journal of Geophysical Research , vol. 82, no. 24, pp. 3439–3444, 1977.[133] X. Zhang, Z.-S. Wu, and X. Su, “Electromagnetic scattering fromdeterministic sea surface with oceanic internal waves via the variable-coefficient gardener model,”
IEEE Journal of Selected Topics in AppliedEarth Observations and Remote Sensing , vol. 11, no. 2, pp. 355–366, 2017.[134] S. R. Massel, “On the geometry of ocean surface waves,”
Oceanologia ,vol. 53, no. 2, pp. 521–548, 2011.[135] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Imagequality assessment: from error visibility to structural similarity,”
IEEEtransactions on image processing , vol. 13, no. 4, pp. 600–612, 2004.[136] Y. Zhang and L. Jiang, “A novel data-driven scheme for the ship wakeidentification on the 2-d dynamic sea surface,”
IEEE Access , vol. 8,pp. 69 593–69 600, 2020.[137] O. Karakus¸, I. Rizaev, and A. Achim, “A simulation study to evaluatethe performance of the Cauchy proximal operator in despeckling SARimages of the sea surface,” in . IEEE, 2020, pp. 1–4.[138] I. Rizaev, O. Karakus¸, J. S. Hogan, and A. Achim, “The effect of seastate on ship wake detectability in simulated SAR imagery,” in2020IEEE International Conference on Image Processing (ICIP)