232 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 14, NO. 2, FEBRUARY 2017 Target Extraction and Imaging of Maritime Targets in the Sea Clutter Spectrum Using Sparse Separation Masoud Farshchian Abstract— Detection and imaging of endoclutter maritime targets is a difficult problem, because the signal return of the target and the clutter overlap in both Doppler frequency and time. Consequently, traditional Doppler filtering is severely limited in terms of extracting the target signature for further processing. This letter proposes a new sparsity-based Doppler filtering and extraction paradigm for maritime radar where the different backscattered signals are separated in the endoclutter region based on their Doppler bandwidths rather than their Doppler frequencies. The novelty of this letter is successful extraction of the target profiles in the endoclutter region using the application of morphological component analysis with time–frequency transforms. Illustrative results of the proposed approach for boat Doppler profile extraction and inverse synthetic aperture image formation outperform traditional range-Doppler processing in endoclutter. The method represents a new paradigm for target detection and imaging in sea clutter. Index Terms— Inverse synthetic aperture radar, radar detection, radar imaging. I. I NTRODUCTION ADAR target detection using Fourier Doppler filtering is optimal for a constant-velocity point target in the exoclutter region (when the target’s Doppler profile has a different frequency support than the Doppler spectrum of the clutter) [1]. However, the detection performance of the same technique degrades significantly in the endoclutter region (when the target’s Doppler frequencies overlap with the Doppler spectrum of the clutter) relative to a constant false alarm rate (CFAR). Therefore, other signal processing methods to differentiate endoclutter targets from the clutter background are needed. New approaches using adaptive filter detectors have been proposed (for an extensive review see [2]). In general, these adaptive filter detectors rely on first whitening the clutter using covariance estimates derived from secondary samples that are ideally not contaminated by the target and are representative of the cell under test (CUT). Afterward, a test statistic for target detection for each assumed Doppler frequency of the target is calculated and compared with a threshold. A variety of these types of detectors have recently been proposed using the three combinations of [2]: 1) an approximate or exact generalized likelihood ratio test (e.g., Kelly detector, adaptive R Manuscript received April 1, 2016; revised September 14, 2016; accepted November 29, 2016. Date of publication December 28, 2016; date of current version January 19, 2017. The author is with Empyreal Waves LLC, Fairfax Station, VA 22039 USA (e-mail: mfarsh@gmail.com). Color versions of one or more of the figures in this letter are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2016.2636253 matched filter detector, adaptive normalized matched filter detector, and so on); 2) proposed different clutter models (especially spherically invariant random processes models) [3]; and 3) different techniques for the estimation of the timevarying clutter covariance matrix. Based on the conclusion of Carretero-Moya et al. [2], these methods have been shown to be unreliable due to their large false alarm deviation and probability of detection losses for small target detection. A major reason for this degradation is the estimated covariance matrix whose statistics might differ greatly from the CUT [2], [4]. The unreliable estimates are due to the nonstationarity of sea clutter itself [2]. Consequently, as pointed out in [2], the approaches of these methods may not be appropriate for the detection of targets in sea clutter. Carretero-Moya et al. [2] point out that: the main conclusion of this letter is that CFAR detection in heterogeneous high-resolution sea-clutter is clearly still an open problem. We propose a novel approach for endoclutter target detection. The proposed nonlinear filtering technique, new in the maritime radar framework, allows signals to overlap in time and Doppler frequency, yet separates (decomposes) them based on their differing Doppler bandwidth [5] rather than Doppler frequency (as in traditional range-Doppler processing). Furthermore, our emphasis is to perform such a separation (decomposition) without utilizing any adjacent or target cell covariance estimates or probability distributions [2], [4]. As will be explained and demonstrated by the experimental results based upon real measured data, the method of this letter is well-suited for discriminating signals based on their Doppler bandwidth and does not require the covariance estimate of the clutter. The ability to extract target signatures in endoclutter is a salient feature of the approach of this letter compared with the adaptive filtering algorithms that assume quasi-stationarity of sea clutter [2]. Morphological component analysis (MCA), first applied to image processing, is a recent sparsity-based separation technique [6] that decomposes a signal mixture into several components where the components are assumed to be sparse in different transform domains [6]. Two requirements for the effective application of MCA are: 1) the sparsifying dictionaries and 2) the optimization algorithm. In this letter—as far as we are aware for the first time—we apply MCA to the radar problem of: 1) boat Doppler profile extraction in endoclutter and 2) inverse synthetic aperture radar (ISAR) endoclutter image formation. In particular, our objective is to work on the raw I /Q data and separate the target from the sea clutter 1545-598X © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. FARSHCHIAN: TARGET EXTRACTION AND IMAGING OF MARITIME TARGETS IN THE SEA CLUTTER SPECTRUM using their different Doppler bandwidths. The novelty here is the application of MCA with the specific dictionaries for the successful extraction of maritime targets. An important emphasis of this letter is the utilization of dictionaries that lead to the sparsification of the target in the endoclutter and the exoclutter regions. The dictionaries used in this letter are the invertible short-time Fourier transform (STFT) with different window-lengths. The first application of such dictionaries in radar was the attenuation of Doppler-steaks due to transient noise and target spread for low frequency radar [7]. In [8], an MCA approach using the resonance-based dictionaries [9] was applied to the maritime discrimination of boats from bird flock (bird clutter) in sea clutter. It was noted that the spectrum of the bird flock has a broad Doppler spectrum while that of the boat has a narrow Doppler spectrum. The recent paper by Nguyen and Al-Ashwal [10] also utilize the resonance-based dictionary [9] to improve exoclutter detection of boat targets in sea clutter. However, despite the improved results discussed in [8] and [10], the utilization of the resonance-based dictionaries in [8] and [10] is not suited for the endoclutter region. This is because both the target and the clutter will have a small number of oscillations. Consequently both signals will be sparsified by the same low-resonance dictionary (which sparsifies signal with small number of oscillations as opposed to the high-resonance dictionary) [9], leading to poorer separation. Obviously, this is not the case for exoclutter targets which will have a sparse high-resonance (large number of oscillations) component, and subsequently can be successfully separated from low-resonance signals [8], [10]. Since the focus of this letter is on endoclutter targets rather than exoclutter targets, the STFT is preferable to constant-Q transforms used in [8] and [10]. In Section II, a brief overview of MCA [6] is provided. The dictionaries in this letter are the STFT frames which are discussed in Section III. In Section IV, we apply the MCA approach to extract a small radar cross section (RCS) boat from a high sea state clutter, and also to ISAR image formation. Section V summarizes the contributions of this letter and future research directions. II. M ORPHOLOGICAL C OMPONENT A NALYSIS We utilize the MCA approach to extract a narrow-band component from a wideband background over all Doppler frequency bins even when both components overlap in time and frequency. Given an observed signal x = x1 + x2 , where x1 , x2 ∈ C N represent the component signals of interest, the goal of MCA is to estimate x1 and x2 individually. We denote by ∗ that the Hermitian transpose of operator . We assume that the over-complete transforms ∗1 and ∗2 enable sparse representations of the vectors x1 and x2 , with coefficient vectors w1 ∈ C M1 ×K 1 and w2 ∈ C M2 ×K 2 , respectively. Note that ∗i has its domain as an Mi × K i sized vector (which can be written as a matrix) and its range as an N sized complex vector. That is, ∗i : C Mi ×K i → C N . A sparse representation of a signal is one where a small number of representation coefficients account for most of the signal energy. According to one form of MCA, given x, we estimate x1 and x2 by 233 solving the following minimization problem: arg min λw1 1 + (1 − λ)w2 1 (1a) ∗1 w1 (1b) w1 ,w2 such that x = + ∗2 w2 . Here, λ is a positive scalar regularization parameter although it can readily be extended to a vector to emphasize different atoms within a dictionary. One method to obtain the sparse w1 and sparse w2 coefficients is the split augmented Lagrangian shrinkage algorithm [11] as described in [9] and is also widely known as the alternating direction method of multipliers optimization [11]. The method relies upon breaking a large optimization problem into smaller subproblems that can be solved more easily. An algorithm to minimize (1) is given by [9] and reproduced here for completeness initialize: μ > 0, d, w (2a) w1 d1 0.5λ/μ d1 u1 ← soft + , − (2b) u2 w2 d2 d2 0.5(1 − λ)/μ u1 1 1 d1 ← x − ∗1 ∗2 (2c) d2 u2 2 2 d u w1 ← 1 + 1 (2d) w2 d2 u2 repeat where soft(y, T ) is the soft-threshold rule with threshold T defined by soft(y, T ) = y max(0, 1 − T /|y|), y ∈ C, T ∈ R+ . (3) Because of the high clutter-to-noise ratio (measured above 25 dB) in the experimental data, we have used an equality constraint in (1b) leading to signal representation rather than signal approximation (as in other forms of MCA). III. A PPROPRIATE D ICTIONARY Choosing the appropriate dictionary i is one of the main requirements in the successful application of MCA. Under the narrow-band assumption, the baseband return signal r (t) of a transmitted signal s(t) for a constant point scatterer moving with a constant velocity is r (t) = αs(t − τ ) exp(i ωd (t − τ )) (4) where ωd is the Doppler frequency, τ is delay, and α is a constant that is proportional to the RCS. Even with a nonconstant RCS, the amplitude modulation of the RCS does not induce a large Doppler bandwidth relative to sea clutter. Consequently, the narrow-band signal of a point target is sparsely represented with a few coefficients for a long-window STFT or discrete Fourier transform (DFT). In contrast, an accurate model for the background sea clutter is not available. Since sea clutter returns from the ocean are scattered from a distributed surface which is undergoing complex and dynamic wave motion, sea clutter has a broad time-varying spectrum. Qualitatively (as is apparent from the STFT), and quantitatively, the mean Doppler bandwidth of the sea clutter is much wider than those of point scattered targets [5] (also see Fig. 4 where the clutter and target Doppler bandwidths are about 200 and 10 Hz, respectively). A wideband signal which is equivalent to a more transient 234 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 14, NO. 2, FEBRUARY 2017 signal (e.g., sea clutter and spray generated by the ocean) is more sparsely represented with a short-window STFT than a long-window STFT. We can write the discrete STFT S(k, m) as R−1 (5) S(k, m) = DFT M x(n − Lm)v(n)|n=0 , 0 M−R where m is the time index, k is the frequency index, R is the window-length, L is the temporal hop, and M is the DFT length. Both the overlap factor of R/L, where 1 ≤ L ≤ R, and frequency sampling of M > R leads to more STFT coefficients than signal values N, where the length of x(n) is N. We note that K i in ∗i is equal to N/L. In order to make the adjoint of the forward STFT equal to its left inverse, we start with (5) and denote s(n, m) as the inverse discrete Fourier transform (IDFT) of S(k, m). Multiplying by the symmetric window v(n) gives s(n, m) = IDFT M (S(k, m))v(n) = x(n − Lm)v 2 (n). (6) Substituting p = n − Lm, we sum over the overlapping blocks to obtain the original signal provided v(n) satisfies x( p)v 2 (Lm + p) = x( p). (7) Fig. 1. ISAR image before and after MCA. TABLE I ISAR R ADAR D ESCRIPTOR m We have used a sine window which satisfies the abovementioned relationship [12]. Consequently, we have the following tight frame relationship between and its adjoint: ∗ x = x. (8) Now, the algorithm of the Section II can be applied. TABLE II CSIR R ADAR D ESCRIPTORS IV. A PPLICATION OF MCA TO S MALL B OAT D ETECTION AND ISAR A. Application to ISAR ISAR obtains a high-resolution image of an object by coherently processing pulses that are produced as the object of interest rotates [13]. In high sea states, and for boats producing turbulent sprays due to dynamic interactions with the sea, the effects on the Doppler spectrum of the spray from the boat and sea clutter also appears in individual range-cells. These nontarget features adversely affect the recognition of the object. Consequently, the full classification potential of ISAR may not be realized in this scenario. Since the reflections from the boat are narrowband relative to the nonstationary sea clutter and spray, the separation of the wideband and narrowband components can be performed via MCA. Furthermore, the wideband components that are extracted can be used as an additional feature of a classifier as each boat may have different Doppler profiles generated from its turbulent spray. The NRaD linear frequency modulated radar is a wideband, coherent, high-range resolution radar that operates at X-band [14]. Table I summarizes the radar parameters. The radar is located on the west side of Point Loma, San Diego, CA, overlooking the Pacific Ocean, at approximately 120 ft above sea level. The experiment here used a utility boat. The raw range-Doppler data set was first rangealigned using a cross-correlation algorithm [13]. Afterward, MCA was applied to the complex 2-D range-Doppler matrix on a range-cell-by-range-cell basis with 100 iterations per range-cell. A regularization value of λ = 0.8 was chosen with STFT window-lengths of R1 = 4 and R2 = 128, and a temporal hop of L = 1 for each dictionary. Fig. 1 (in dB scale and the Doppler axis is normalized by the pulse repetition frequency (PRF)) shows the results of separating the spray/sea clutter (“wideband”) from the (“narrowband”) utility boat. The additional clarity of the narrow-band component allows for further application of motion compensation algorithms to enhance the ISAR images while the wake/spray can be used as an additional feature to reduce misclassifications. It should be noted that the MCA approach is drastically different than thresholding. Fig. 2 shows the narrow-band and wideband component of the spectrogram of the range-cell before and after separation (decomposition). We note that there are high amplitude wideband components (from the spray) that were correctly separated into the wideband stream and high amplitude narrow-band components that were correctly separated into the narrow-band stream. In a threshold method, both high amplitude components (narrowband and wideband) are selected which results in no separation between target scatterers and clutter. FARSHCHIAN: TARGET EXTRACTION AND IMAGING OF MARITIME TARGETS IN THE SEA CLUTTER SPECTRUM Fig. 2. Fig. 3. SCR gain as a function of λ for data set 2. Fig. 4. Doppler spectra of the target cell. 235 Spectrogram before and after MCA. B. CSIR Data We also applied the new extraction algorithm to several measured S-band and X-band data sets provided by The Council for Scientific and Industrial Research (CSIR) in South Africa [15] with consistent results. Unlike the ISAR results, there is no range-migration or rotational diversity in our processing interval, and we focus solely a Doppler profile of the target. Here, we mention only one experiment in detail where the target is slowly drifting. We denote SCRi and SCRo as the input and output signal-to-clutter ratio (SCR) before MCA and after the MCA narrow-band component, respectively. Similarly, MCRi and MCRo are the input and output maximum clutter ratio (MCR) (maximum clutter cell power divided by the average power of all clutter cells) before and after MCA, respectively. We denote the separation metric values between the target and maximum clutter cell before and after MCA as i = SCRi − MCRi and o = SCRo − MCRo , where the subtraction is done in the dB scale. It should be noted that SCRi is used in traditional radar processing such as CFAR detectors [1] and its small value is the reason for the difficulty of detection of small targets in the endoclutter region. The CSIR data set “TFC15-007” (summarized in Table II) was examined as an example of an endoclutter target. The target is a rigid inflatable boat with an RCS of the range 1–5 m2 . The target plus clutter cell was located at range-gate 20 (roughly corresponding to 3285 m). For display comparisons, we use a clutter range-gate number 47 (roughly 3690 m) as a representative of a clutter cell. For the MCA separation, we chose a window-length of R1 = 32 samples for the wideband dictionary and a windowlength of R2 = 2048 samples for the narrow-band stream (corresponding to 0.41 s). An overlap of 80% was used for the STFT of both dictionaries. A four second portion of the datastream was taken for analysis and data processing. It should be noted that the MCA can be applied to different data lengths. However, we experimented with a longer data length in order to analyze the variation of the MCA output over a few seconds. In Fig. 5, from the top to the bottom, the spectrogram of a portion of the pre-MCA data, the post-MCA wideband spectrogram, and the post-MCA narrow-band spectrogram are shown, respectively, for the clutter cell and the value of λ = 0.18. Beyond 100 iterations of the algorithm to separate the data into a narrow-band and wideband component, there was no significant improvement seen in the quality of the separation. In Fig. 4, from the top to the bottom, the spectrogram of a portion of the pre-MCA data, the post-MCA wideband spectrogram, and the post-MCA narrow-band spectrogram are shown for the target plus clutter cell for the value of λ = 0.18. For both Figs. 4 and 5, a window-length of R = 1024 was used to display the spectrograms of the narrow-band and wideband streams. Fig. 3 shows SCRi , SCRo , MCRi , and MCRo for this experiment. Fig. 3 shows that a significant improvement of 4.76–40.7 dB for the average SCR can be obtained when λ is between 0.10 and 0.25. Similarly, i = 0.43 dB, while the maximum value of o , which occurs at λ = 0.13, is equal to 34.73 dB. These numbers are due to the fact that MCA separates the two signals based on their Doppler bandwidth. That is, using the new algorithm, we have increased the separation of the target power relative to the maximum clutter power in the narrow-band component. The increase in o relative to i suggests that the false alarms can be reduced 236 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 14, NO. 2, FEBRUARY 2017 covariance-based adaptive filter methods [2]. It should be mentioned that for ultrawideband radar, the resonance-based dictionaries may also be suitable due to the time scaling of the returns. Comparison of the STFT approach to the resonance and other dictionaries for narrow-band and wideband radar detection, discrimination and imaging of exoclutter and endoclutter targets with varying trajectories is left as a future research topic. It should be noted that the MCA approach allows for several different dictionaries from different transform families to be combined. Given the reasons and results of this letter, we believe the image-processing-based MCA approach is a promising alternative technique to covariancebased adaptive filter methods for nonstationary radar clutter, and future work will compare these two different approaches. R EFERENCES Fig. 5. Doppler spectra of a clutter cell. by processing the narrow-band component in the detection chain rather than the mixed component. The optimal value of λ may depend on several factors, including the observation time for processing, the persistence of the target and the clutter, the PRF of the data, and the Doppler bandwidth (which is determined by environmental conditions). However, even if the optimal value of λ is not chosen, significant gains are achieved for a wide constant range of λ values. Consequently, this new approach has allowed us to extract the Doppler profile of the endoclutter small boat target with a large separation from the clutter background. This Doppler profile can be used for boat detection and Doppler profile classification. V. D ISCUSSION AND C ONCLUSION We proposed a technique to separate the clutter from the target in the endoclutter region using the MCA framework. For maritime ISAR imaging, the technique was compared with traditional range-Doppler processing for a boat, and for a boat in a single range-cell, the SCR improvement was calculated before and after the application of MCA. The key contributions of this letter can be summarized as follows: 1) a method to extract maritime targets in the endoclutter region using MCA; 2) the first—as far as we are aware—application of MCA for ISAR image formation; 3) the first—as far as we are aware—application of MCA to small boat endoclutter Doppler extraction. The advantages of STFT dictionaries relative to resonance-based dictionaries that were used in [8] and [10] were also explained in the endoclutter region. We also highlighted the fundamentally more intuitive approach of improving SCR in nonstationary sea clutter relative to the [1] M. A. Richards, J. Scheer, and W. A. Holm, Principles of Modern Radar: Basic Principles. Raleigh, NC, USA: SciTech, 2010. [2] J. Carretero-Moya, J. Gismero-Menoyo, A. Asensio-Lopez, and A. Blanco-Del-Campo, “Small-target detection in high-resolution heterogeneous sea-clutter: An empirical analysis,” IEEE Trans. Aerosp. Electron. Syst., vol. 47, no. 3, pp. 1880–1898, Jul. 2011. [3] E. Conte and M. Longo, “Characterisation of radar clutter as a spherically invariant random process,” IEE Proc. Commun., Radar Signal Process., vol. 134, no. 2, pp. 191–197, Apr. 1987. [4] M. Greco, F. Gini, A. Younsi, M. Rangaswamy, and A. Zoubir, “Nonstationary sea clutter: Impact on disturbance covariance matrix estimate and detector CFAR,” in Proc. IEEE Int. Conf. Radar, Sep. 2008, pp. 558–562. [5] D. Walker, “Doppler modelling of radar sea clutter,” IEE Proc. Radar, Sonar Navigat., vol. 148, no. 2, pp. 73–80, Apr. 2001. [6] J.-L. Starck, F. Murtagh, and J. M. Fadili, Sparse Image and Signal Processing: Wavelets, Curvelets, Morphological Diversity. Cambridge, U.K.: Cambridge Univ. Press, 2010. [7] I. W. Selesnick, K. Y. Li, S. U. Pillai, and B. Himed, “Doppler-streak attenuation via oscillatory-plus-transient decomposition of IQ data,” in Proc. IET Int. Conf. Radar Syst., Oct. 2012, pp. 1–4. [8] M. Farshchian and I. Selesnick, “Application of a sparse timefrequency technique for targets with oscillatory fluctuations,” in Proc. IEEE Int. Waveform Diversity Design Conf. (WDD), Jan. 2012, pp. 191–196. [9] I. W. Selesnick, “Sparse signal representations using the tunable Q-factor wavelet transform,” Proc. SPIE, vol. 8138, p. 81381U, Sep. 2011. [10] S. T. N. Nguyen and W. A. Al-Ashwal, “Sea clutter mitigation using resonance-based signal decomposition,” IEEE Geosci. Remote Sens. Lett., vol. 12, no. 11, pp. 2257–2261, Nov. 2015. [11] M. V. Afonso, J.-M. Bioucas-Dias, and M. A. T. Figueiredo, “Fast image recovery using variable splitting and constrained optimization,” IEEE Trans. Image Process., vol. 19, no. 9, pp. 2345–2356, Sep. 2010. [12] I. W. Selesnick. Short-Time Fourier Transform and Its Inverse, accessed on Dec. 2016. [Online]. Available: http://eeweb.poly. edu/iselesni/EL713/STFT/stft_inverse.pdf [13] C. Ozdemir, Inverse Synthetic Aperture Radar Imaging With MATLAB Algorithms, vol. 210. Hoboken, NJ, USA: Wiley, 2012. [14] J. Trischman, “Architecture and algorithms for real-time ISAR imaging of dynamic targets,” Naval Command, Control Ocean Surveillance Center RDT&T Div, San Diego, California, Tech. Rep. 19960820-0-43, Jul. 1996. [15] P. L. Herselman, C. J. Baker, and H. J. de Wind, “An analysis of Xband calibrated sea clutter and small boat reflectivity at medium-to-low grazing angles,” Int. J. Navigat. Observat., vol. 2008, Jun. 2008, Art. no. 347518, doi:10.1155/2008/347518