RFI suppression in SAR based on filtering interpretation of SSA and fast implementation
© Wang et al; licensee Springer. 2012
Received: 16 November 2011
Accepted: 9 May 2012
Published: 9 May 2012
Synthetic aperture radar (SAR) has proven to be a powerful remote sensing instrument for underground and obscured object detection. However, SAR echoes are often contaminated by radio frequency interferences (RFI) from multiple broadcasting stations. Essentially, RFI suppression problem is one-dimensional stationary time series denoising problem. This article proposed a novel RFI suppression algorithm based on singular spectral analysis (SSA) from a linear invariant systems perspective. It can be seen that SSA algorithm has an equivalence relation with finite impulse response (FIR) filter banks. Besides, this article first introduce two approximated methods which can remarkably speed up spectral decomposition--Nyström method and Column-Sampling approximation--to obtain the coefficients of above SSA-FIR-filter. Simulation results and imaging results of measured data have proved the efficiency and validity of this algorithm.
Keywordssynthetic aperture radar radio frequency interference suppression singular spectral analysis finite impulse response filter Nyström method Column-Sampling approximation.
Synthetic aperture radar (SAR) operates in the P-, L-, C-, and X-bands to image the Earth's surface along the carrier platform's flight path while the antenna is oriented perpendicular to the flight direction in the downward-slanted direction. However, the above frequency bands are not reserved exclusively for SAR applications. They are also occupied by other services like radio and television stations, as well as telecommunication purposes. When the carrier platform passes over these broadcasting stations, the receiver of SAR picks up the signals from these stations also. These interfere signals are called radio frequency interferences (RFIs) which will overlay the SAR information and become visible in the SAR image . Sometimes, they will make the SAR receiver saturated because their power goes beyond the receiver's dynamic load.
Therefore, it remains to be a hot research topic on RFI suppression for decades. The early methods are mainly based on coherent estimation and subtraction of RFIs . Their performances heavily rely on the complicated data models and parameter estimation accuracy so that they are awkward and not robust. Other commonly used methods are notch filtering . Their main drawback is degradation in the range domain of SAR imagery when multiple RFI sources are present. Then, more attractive methods are based on least mean square (LMS) adaptive filters . However, they have a tradeoff between length of the filter and numerical sensitivity of adaptation. Afterwards, some researchers proposed subspace filters for estimating RFI signals with Gram-Schmidt orthogonalization procedure or Eigenvalue Decomposition (ED) [5, 6]. The performances of these methods are robust but the large computational burden limits their application in practice.
This article has proposed a novel RFI suppression algorithm based on subspace projection by taking use of singular spectral analysis (SSA). The roles of signal and noise in the proposed algorithm are exchanged just as in the LMS mechanism. Each SAR echo received during pulse repetition time can be considered as a one-dimensional time series. Thus, RFI suppression problem is the one-dimensional stationary time series denoising problem in nature, and SSA algorithm is a prevalent algorithm when clustering subspace models are applied to time series datasets . The aim of SSA is to obtain RFI clustering and 'wideband noise' clustering. It can be considered as three steps: 'embedding' step, 'spectral decomposition' step, and 're-embedding' step. Interestingly, these steps can be achieved through an analysis-synthesis-joint finite impulse response (FIR) filters group [8–10]. Based on this equivalence relationship, a special kind of filtering mechanism for RFI suppression can be established.
However, on account of large amount, abundant information, the burden of SAR imaging system for data storage and processing is increased. For this reason, this article introduces two random sampling techniques, Nyström method [11–15] and Column-Sampling approximation [14–17], to provide powerful tools to approximate the coefficients of the proposed SSA-FIR filter banks. These techniques are recently used in the machine learning community  and analyzed in the theoretical Computer Science community [13, 16, 17]. Their reconstruction error bound analysis and application guidelines are elaborated in [14–17]. For the first time, we combined them with SAR signal processing and RFI suppression.
The remainder of this article is organized as follows. In Section 2, the signal model of RFI sources contaminating the SAR signals is expressed in formulas, followed by the implementation details of FIR filtering method that is equivalent to SSA algorithm, and then in Section 3, we give two quick approaches, the Nyström method and the Column-Sampling approximation, for the filtering method in Section 2. In Section 4, the proposed algorithm is tested by means of using numerical simulations and P-band SAR real data, and comparison and analysis of these results are given. Finally, Section 5 comes up with the conclusions.
2. Signal model and filtering interpretation of SSA
Generally, for distributed targets (e.g., in real scenario) secho(t r , t a ) resembles Additive White Gaussian Noise (AWGN), and sNoise(t r , t a ) generated by system is AWGN, too. RFI signal has comparatively narrow band, and can be considered but not limit to complex exponential signal. Previous researches [1–6] pointed that the power of RFI signals is tens of dBs greater than the power of the SAR echoes. So, secho(t r , t a ) and sNoise(t r , t a ) can be together considered as wideband background noise which is independent from the RFI source, denoted as sWB(t r , t a ) = secho(t r , t a )+sNoise(t r , t a ).
where P is a diagonal selected matrix, with the i th diagonal entry satisfies P ii = 1 if the i th row of Q = PUHS is selected whereas P ii = 0 it will be discard. Thus, we select the first r = rank(G) corresponding to the dimension of RFI subspace entries, .
where , , and z is a complex value.
where , .
Besides, once alternative embedding procedure is adopted--the trajectory matrix S has Hankel structure but not Toeplitz structure (see [8, 9] for example), the properties of H i (e jω ) and F i (e jω ) will interchange, i.e., the former corresponding to an anti-casual filter whereas the latter a casual filter.
3. Approximation of spectral decomposition
From Equations (10) (13), and (14), we learn that the entries of eigenvector u i correspond to the coefficients of the SSA FIR filter we proposed. Traditionally, eigenvectors can be obtained via spectral decomposition algorithm, SVD on S or ED on G. Nevertheless, they are prohibitive for large L. Here, we introduced two different approximate spectral decomposition methods for large dense matrices: Nyström method [11–15] and Column-Sampling approximation [14–17]. Both of them are based on random sampling techniques which only operate on a subset of the matrix and provide a powerful alternative for approximate spectral decomposition issue.
Note that W is also SPSD since G is SPSD. Through performing ED or SVD on much smaller scale matrices W and C, it can generate approximations of eigenpairs U and Λ, denoted as and respectively, which we now describe.
Here, we adopts and . Compared with (17), it is obvious that the two methods have resembling forms with each other, the latter replaces W in (14) with .
The above sampling-based approximations adopt the most basic uniform sampling approach to pick out l columns from the original matrix G. Alternatively, some researchers derived several non-uniformly sampling approaches, i.e., selecting the i th column with a weight that is proportional to either the corresponding diagonal element G ii or the squared of the column-norm [13, 17]. However, in the recent articles which have the guiding significance for the applications of sampling-based approximations on various problems, Kumor et al. [14, 15] pointed that, for large dense matrices, the uniform sampling without replacement approach, in addition to being more efficient both in time and space, produces more effective approximations. Otherwise, Kumor et al.  summarized that the low-rank approximation issues can be classified into two groups--matrix projection and spectral reconstruction. Furthermore, the former approximations tend to be more accurate than the latter approximations when adopt sampling-based approximated methods. According to Equation (4), the RFI signal reconstruction belongs to the first kind. Based on the above analysis, using uniform sampling approach is a good choice.
In the basic SSA algorithm, the most time-consuming step is the spectral decomposition of the L × L matrix G (or of the L × K trajectory matrix S). The computational cost of SVD on S is usually computed by the means of Golub and Reinsch algorithm , which requires O(L2K+LK2+K3) multiplications. Using ED on G is even more prohibitive required O(L3) multiplications. Contrarily, the Nyström method just needs to calculate the eigenpairs of the sampled submatrix of G which requires O(l3+l2L), r ≤ l, l3 required by ED on W and l2L for multiplication with C, and the Column-Sampling approximation needs O(L2l+Ll2+l3) multiplications for SVD on C. After measuring differences of floating-point numbers in arithmetic, we supply a comparison of concrete CPU time when implementing different spectral decomposition algorithms on P-band real datasets, shown in Section 4.
4. Numerical experiments
In this section, we first perform numerical simulations for a time series to illustrate the methods discussed above. We consider a linear frequency modulation (LFM) signal, namely chirp signal, which is often used as SAR transmitting pulse : secho = exp(-jπKt2), -T/2 ≤ t ≤ T/2, where the pulse width T = 32 μs, bandwidth B = |K·T| is 9.6 MHz, the center frequency at zero and sampling frequencies is 39.6 NHz. All these parameters are typical for a SAR transmitting pulse, except the sampling frequency. According to Nyquist principal 12 MHz is enough for the given bandwidth, but we adopt a high sampling frequency in order to obtain a comparatively 'long' time series which has 1,844 sample points. Three real sinusoidal signals as RFIs, whose center frequencies are 1.8, 3.2, and 3.5 MHz, respectively, have been added. To set two RFIs with similar frequencies, we can examine the RFI suppression capability of SSA-FIR filter banks when RFIs with close frequencies present, and the signal-to-noise ratio is 40 dB.
The reconstructed signals using accurate ED method are shown in Figure 2e, f, corresponding to time and frequency domains, respectively, and in Figure 2g, h, the performances of the Notch filtering method are represented. Then, we first set l = floor(L/8) = 57, in this condition, the reconstructed signals using the Column-Sampling approximation and the Nyström method are shown in Figure 2i-l. One can see the Column-Sampling approximation works better and more approximately up to the ED reconstructed performance. Contrarily, the performance of the Nyström method is kind of unsatisfactory. It still leaves about 10 dB RFIs energy and the time domain signal wave is 'chaotic'. Next, we set l = floor(L/4) = 115 and the reconstructed signal using the Nyström method are shown in Figure 2m, n. This time it works better, more RFIs energy is removed but its performance still little poorer than the Column-Sampling approximation's, even when the latter with half of the sampling columns. The reason we will give through further numerical analysis. However, the Nyström method with l = floor(L/8) = 57 sampling columns still removes more RFIs energy than the Notch filtering method does. We can see the Notch caused some frequency spectrum fracture and lost the most useful signal.
Mean deviation from orthonormality for different number of samples
Main system parameters
Pulse repetition frequency
In this article, we give analysis of that the RFI suppression problem in SAR can be considered as one-dimensional stationary time series denoising problem. Furthermore, by applying a linear invariant system theory, the RFI suppression task can be achieved expediently by a zero-phase FIR filter whose coefficients are related to the eigenvectors of the input signal covariance matrix. Owing to the outputs being in phase with the inputs, phase preserving can be available through the proposed algorithm. What is more, in view of the large amount of SAR original data and high complexity of computing eigenvalues and eigenvectors, we first introduce two random sampling methods to speed up the SSA-FIR filter banks construction process. The results of simulations and practical experiments illuminate that the proposed filtering algorithm can be provided with both efficiency and validity.
- Soumekh M, Nobles D, Wicks MC, Gerard JG: Signal processing of wide bandwidth and wide beam width P-3 SAR data. IEEE Trans Aerosp Electron Syst 2001, 4(37):1122-1141.View ArticleGoogle Scholar
- Miller T, Potter L, McCorkle J: RFI suppression for ultra wideband radar. IEEE Trans Aerosp Electron Syst 1997, 33(4):1142-1156.View ArticleGoogle Scholar
- Kimura H, Nakamura T, Papathanassiou KP: Suppression of ground radar interference in JERS-1 SAR data. IEICE Trans Commun 2004, E87-B: 3759-3765.Google Scholar
- Luo X, Ulander LMH, Askne J, Smith G, Frolind P-O: RFI suppression in ultra-wideband SAR systems using LMS filters infrequency domain. IET Electron Lett 2001, 37(4):241-243. 10.1049/el:20010153View ArticleGoogle Scholar
- Nguyen L, Ton T, Wong D, Soumekh M: Adaptive coherent suppression of multiple wide-bandwidth RFI sources in SAR. Proc SPIE 2004, 5427: 1-16.View ArticleGoogle Scholar
- Wang W, Wyatt LR: Radio frequency interference cancellation for sea-state remote sensing by high-frequency radar. IET Radar Sonar Navigat 2011, 5(4):405-415. 10.1049/iet-rsn.2010.0041View ArticleGoogle Scholar
- Golyandina N, Nekrutkin V, Zhigljavsky A: Analysis of Time Series Structure: SSA and Related Techniques. Chapman & Hall, London; 2001.View ArticleGoogle Scholar
- Harris TJ, Yuan H: Filtering and frequency interpretations of singular spectrum analysis. Physica D 2010, 239: 1958-1967. 10.1016/j.physd.2010.07.005MathSciNetView ArticleGoogle Scholar
- Hansen PC, Jensen SH: FIR filter representations of reduced-rank noise reduction. IEEE Trans Signal Process 1998, 46(6):1737-1741. 10.1109/78.678511View ArticleGoogle Scholar
- Dologlou I, Carayannis G: Physical interpretation of signal reconstruction from reduced rank matrices. IEEE Trans Signal Process 1991, 39(7):1681-1682. 10.1109/78.134407View ArticleGoogle Scholar
- Williams CKI, Seeger M: Using the Nyström method to speed up kernel machines. NIPS 2001, 13: 682-688.Google Scholar
- Zhang K, Tsang I, Kwok J: Improved Nyström low-rank approximation and error analysis. Proc ICML'08, New York, USA 2008, 273-297.Google Scholar
- Drineas P, Mahoney MW: On the Nyström method for approximating a Gram matrix for improved kernel-based learning. JMLR 2005, 6: 2153-2175.MathSciNetGoogle Scholar
- Kumor S, Mohri M, Talwalkar A: Sampling techniques for the Nyström method. Conference on Artificial Intelligence and Statistics JMLR: W&CP, Clearwater Beach, Florida, USA 2009, 5: 304-311.Google Scholar
- Kumor S, Mohri M, Talwalkar A: On sampling-based approximate spectral decomposition. Proc 26th Annual International Conference on Machine Learning, Montreal, Canada 2009, 553-560.Google Scholar
- Deshpande A, Rademacher L, Vempala S, Wang G: Matrix approximation and projective clustering via volume sampling. Theor Comput 2006, 6: 225-247.MathSciNetView ArticleGoogle Scholar
- Drineas P, Kannan R, Mahoney MW: Fast Monte Carlo algorithms for matrices II: computing a low-rank approximation to a matrix. SIAM J Comput 2006, 36(1):158-183. 10.1137/S0097539704442696MathSciNetView ArticleGoogle Scholar
- Wax M, Kailath T: Detection of signals by information theoretic criteria. IEEE Trans Acoust Speech, Signal Process 1985, 33(2):387-392. 10.1109/TASSP.1985.1164557MathSciNetView ArticleGoogle Scholar
- Badeau R, David B, Richard G: Selecting the modeling order for the ESPRIT high resolution method: an alternative approach. IEEE International Conference on Acoustics, Speech, Signal Processing, Montreal, Quebec, Canada 2004, 2: 1025-1028.Google Scholar
- Papy J-M, De LathauwerL, Huffel SV: A shift invariance-based order-selection technique for exponential data modeling. IEEE Signal Process Lett 2007, 14(7):473-476.View ArticleGoogle Scholar
- Kritchman S, Nadler B: Determining the number of components in a factor model from limited noisy data. Chemomet Intell Lab Syst 2008, 94(1):19-32. 10.1016/j.chemolab.2008.06.002View ArticleGoogle Scholar
- Johansson K: Shape fluctuation and random matrices. Commun Math Phys 2000, 209(2):437-476. 10.1007/s002200050027View ArticleGoogle Scholar
- Johnstone IM: On the distribution of the largest principal components analysis. Ann Statist 2001, 29(2):295-327. 10.1214/aos/1009210544MathSciNetView ArticleGoogle Scholar
- Baik J, Ben ArousG, Péché S: Phase transition of the largest eigenvalue for non-null complex sample covariance matrices. Ann Probab 2005, 33(5):1643-1697. 10.1214/009117905000000233MathSciNetView ArticleGoogle Scholar
- Delves LM, Mohamed JL: Computational Methods for Integral Equations. Cambridge University Press, Cambridge; 1985.View ArticleGoogle Scholar
- Golub G, Reinsch C: Singular value decomposition and least squares solutions. Numer Math 1970, 14: 403-420. 10.1007/BF02163027MathSciNetView ArticleGoogle Scholar
- Cumming IG, Wong FH: Digital Processing of Synthetic Aperture Radar Data: Algorithm and Implementation. Artech House, Inc., Boston; 2005.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.