DVB-T interference suppression in FL-GPR systems

Transcription

1 DVB-T interference suppression in FL-GPR systems F.I. Rial*, Roi Mendez-Rial, Lukasz Lawadka, Maria A. Gonzalez-Huici Fraunhofer Institute for High Frequency Physics and Radar Techniques (FHR), Fraunhoferstraße 20, Wachtberg, GER. ABSTRACT In this paper we show using some examples how radio frequency interference (RFI) generated by Digital Video Broadcasting Terrestrial (DVB-T) and Digital Audio Broadcasting (DAB) transmitters can be an important noise source for forward-looking ground penetrating radar (FLGPR) systems. Even in remote locations the average interference power sometimes exceeds ultra-wideband (UWB) signals by many db, becoming the limiting factor in the system sensitivity. The overall problem of RFI and its impact in GPR systems is briefly described and several signal processing approaches to removal of RFI are discussed. These include spectral estimation and coherent subtraction algorithms and various filter approaches which have been developed and applied by the research community in similar contexts. We evaluate the performance of these methods by simulating two different scenarios submitted to real RFI acquired with a FLGPR system developed at the Fraunhofer Institute for High Frequency Physics and Radar Techniques (FHR), (GER). The effectiveness of these algorithms in removing RFI is presented using some performance indices after suppression. Keywords: Forward-looking Ground Penetrating Radar (FLGPR), radio frequency interference (RFI), Digital Video Broadcasting Terrestrial (DVB-T). 1. INTRODUCTION Ground Penetrating Radar (GPR) systems are typically ultra-wideband (UWB) radars operating in a frequency range from hundreds of MHz to several GHz. Low operating frequencies are required to penetrate the soil, while a large bandwidth is required for high resolution imaging. As a result, the GPR signal occupies a wide spectrum shared by radio, TV, mobile communications, and other radio frequency (RF) communication systems. From the point of view of GPR, Radio Frequency Interference (RFI) produced by all those EM sources operating in the same frequency band may become the limiting factor on system sensitivity and severely degrade the image quality [1]. GPR systems can be cast into two main groups as down-looking (DL-GPR) and forward-looking (FL-GPR). Most of DL-GPR place their antennas in contact (or very close) with the surface to improve coupling and maximize power transmitted into the ground. This type of antennas ( ground-coupled ) is usually encapsulated (shielded) to minimize the power coming from/to other directions different from the ground. Thus, RFI influence on the data is diminished. Some DL-GPR systems predominantly used in traffic infrastructure surveys use also another type of antennas ( air-coupled ) designed to work at short standoff distances from the road surface (<1m). The greatest advantage of air-coupled systems is perhaps their repeatability, since antenna coupling does not change with the changes in soil properties [2]. However, for these antennas, RFI may become an important issue when working in urban environments and some ideas to mitigate RFI have been proposed [3]. The FL-GPR case is still more dramatic in terms of RFI. FL-GPR points its antennas forward and inspects the ground surface in front of the system with a longer standoff distance. With this configuration, it is capable of collecting data for a much larger area in much shorter time than its DL-GPR counterparts and offers the clear advantage of detecting targets before passing over them. Due to this configuration, FL-GPR systems are less RFI shielded and they are more prone to suffer from severe RFI interference than DL-GPRs.

2 Figure 1. General antenna design for GPR systems: a) DL-GPR Ground-coupled; b) DL-GPR Air-coupled (so called air-launched ); c) FL-GPR antennas. A number of RFI suppression algorithms for UWB systems have been described in the literature. Some of them are GPR specific, but approaches based on Synthetic Aperture Radar (SAR) systems can be generally adapted to the particular characteristics of the GPR signal, e.g., record length (from few ns to microseconds), Pulse Repetition Frequency (PRF) or ADC sampling rates. However, many of these algorithms assume that RFI sources consist only in narrowband channels, so they may not be effective when dealing with the current frequency spectrum where many interference signals are relatively wide-bandwidth with complicated modulation schemes. Digital Video Broadcasting Terrestrial (DVB-T) and Digital Audio Broadcasting (DAB) are European-based consortium standards for the broadcast transmission of digital terrestrial television and radio. In the last years, DVB-T and DAB stations have been widely deployed worldwide in several countries, becoming a potential RFI source for GPR systems. DVB-T broadcasters use COFDM modulation for transmitting video typically with bandwidths of 6-8 MHz in the frequency band from MHz. Power transmitted from DVB-T stations is usually in the order of tens of kw (sometimes over 100 kw). These characteristics make DVB-T signals a major interference for FL-GPR systems that may operate nearby. In this work we determine the best strategy to improve the data obtained from a particular GPR system when working in environments with high DVB-T interference. The system is a FL-GPR developed at the Fraunhofer Institute for High Frequency Physics and Radar Techniques (FHR), (GER). Through this paper, we review some of the signal processing methods recently proposed in the literature for RFI suppression. Some of them have proven to work efficiently in similar systems currently in use, so we adapt them to the particular characteristics of our system and RFI conditions. In addition, RFI suppression must run in real-time in order to achieve realtime SAR imaging and detection. The computational burden demanded by the different approaches is not discussed here. However, all the solutions evaluated have affordable computational costs and could be implemented for real time processing in our system. Other initially appealing approaches [4] have been discarded for being considered slow to fulfill the real time constraint. The remainder of this work is presented as follows. Section 2 briefly describes the methodology adopted: the characteristics of the data utilized for this study, RFI suppression approaches evaluated, and performance indices considered. Section 3 compares and analyzes the results of the RFI suppression methods over two simulated data sets corrupted by real RFI. Finally, Section 4 draws the conclusions.

3 2. METHODOLOGY 2.1 Recorded RFI and Simulated Scenarios The FL-GPR system developed at the FHR has been designed with the purpose of detecting underground objects in real time with standoff distances from a few meters to 40m. The system is mounted in a vehicle and it uses a single transmitter and a linear array of receivers arranged perpendicular to the vehicle moving direction. It can be configured to work with different waveforms and hardware settings. The configuration is usually chosen considering what specifications of the radar are crucial for a certain application or site. RFI utilized for this study was recorded by the FL-GPR system in one of the test sites that is routinely employed for the general calibration of the system. In this site, the DVB-T interference (and DAB to a lesser extent) was noticeable in the data, surpassing by far any other possible RFI source in the bandwidth of interest (Fig. 2). The RFI was collected in passive mode (no transmission) at several positions along the synthetic aperture. The spectrum was measured with records of length 1800 samples at 2 GS/s per channel with a Pulse Repetition Frequency (PRF) of 5 KHz. Similar system settings are used in the FL-GPR system for acquiring data with dense spatial sampling at ordinary vehicle velocities. Figure 2. RFI spectrum in the band from 100 MHz to 1 GHz collected in passive mode (no transmission) in the measurements area. We embedded the recorded RFI into two simulated scenarios. In both scenarios we consider a system transmitting a Linear Frequency Modulated (LFM) chirp waveform with a bandwidth from 100 to 900MHz. The first scenario consists in just one transmitter-receiver channel and 3 targets separated 15ns. Simulated targets are spheres with diameter 40cm. The model is kept simple and no complex interactions signal-target or targettarget are contemplated (i.e. reflections between targets, resonances, glory waves, etc.) and just signal attenuation due to geometrical spreading is considered. Antenna crosstalk is also not considered in the simulation. Nevertheless, it represents a meaningful example of an underground object close to the surface where top and bottom reflections are visible. For the analysis of the RFI suppression approaches we add the 1D model to 1000 consecutive RFI records, simulating the operation of the FL-GPR system with one channel in an environment with strong DVB-T interference. We apply RFI suppression methods before range compression (chirp de-ramping). Even though for some of these methods the difference between to apply them before or after range compression is small, others will not work as expected since RFI characteristics change after compression. The performance indices for this experiment are calculated as an average of the 1000 data records to yield statistically valid results. Fig 3(a) shows the synthetic data before adding RFI. Targets scattering response is confined between meters 40 and 50. Fig 3(b) shows the data embedded in one of the 1000 RFI records. The RFI has amplitude similar to the target response so to simulate the actual SNI (Signal-to-Noise-Interference) experienced by the FLGPR system

4 in the test site where the RFI was recorded. Below these two subplots we see their power spectra. The bottom of Fig.3 shows the data after range compression. Note that after compression only data from meters 34 to 46 is shown. Notice also how in Fig. 3(f) it is difficult to distinguish the targets in the RFI corrupted signal. Figure 3. 1D simulation of three consecutive targets (spheres): a) synthetic data; b) synthetic data with RFI; c) spectrum of the synthetic data; d) spectrum of the synthetic data with RFI; e) synthetic data after range compression; f) synthetic data with RFI after range compression. Figure 4. 3D simulation domain of the FL-GPR with three targets (spheres) buried in a flat ground. The radius and depth of the spheres is also indicated.

5 The second scenario is a 3D model where a finite-difference time-domain (FDTD) method is utilized to compute the scattering from 3 objects (again spheres) buried in a flat ground with dispersive soil. In this model, a complete FL-GPR system is simulated as 1 transmitter and 20 receivers distributed over a 2m-wide linear aperture (Fig. 4). The computational domain has dimensions of 20m x 4m x 4m and was discretized into 2cm x 2cm x 2cm cells. The model simulates the movement of the FL-GPR for 13m acquiring data every 10cm, so finally 130 A-Scans per channel are obtained. The recorded noise is then added to every trace. Figure 4 shows the simulation domain and target disposition. After RFI suppression, range compression and backprojection are applied before analyzing results. We use the backprojection or delay-and-sum (DAS) method [5]. This is a conventional algorithm utilized for FLGPR imaging. We evaluate the suppression methods using 2D horizontal cuts of the soil obtained after backprojection. These cuts are called slices and represent a radar image of the ground at a certain depth. 2.2 RFI Suppression Methods As pointed out by some authors [1, 6], the broad diversity of RFI suppression algorithms may be roughly categorized into two different classes: adaptive filtering and estimation-and-subtraction approaches. Adaptive filtering includes (among others): simple notch filters, RFI autoregressive modeling (AR), least mean square (LMS) approximations and subspace filters. On the other hand, estimation-and-subtraction tries to model the RFI as a sum of sinusoids. RFI modeling may become computationally complex in case of many interference sources and some authors have proposed iterative algorithms to overcome this limitation like the CLEAN [7] or RELAX [8] algorithms. Some approaches in the literature make use of sniffer pulses or listening beforehand schemes to improve signal processing strategies. These approaches usually involve collecting RFI signals at every other PRF of the radar with the transmitter off [9]. Of course, the effectiveness of these approaches depends significantly on how long the RFI remains coherent. Another common solution to improve the SNR and, at the same time, reduce RFI is signal stacking or averaging. The method is based on repeating the measurements from the same range profile (A-Scan) a number of times and make an average. Averaging does not take into account properties of the RFI signal, but it has shown effective RFI suppression for narrowband and wideband sources alike. Stacking comes in handy as ADCs can be usually configured to automatically average a predefined number of consecutive signals very efficiently. Spatial filtering is another interesting approach to suppress RFI. This technique creates antenna pattern nulls in the direction of RFI sources. The RFI Sources may arrive from known directions (conventional and null-steering beamforming) or they can be calculated on-the-fly (optimal beamforming) [10]. Nevertheless, this approach is not considered in this work. One of the reasons for that is because in many occasions RFI will arrive to the system from forward and beam nulling may compromise the detection sensitivity of the system. Considering the different alternatives, we finally tested the following approaches in our attempt to find the best solution to suppress RFI in our system coming from DVB-T transmitters: Adaptive Line Enhancer (ALE): An adaptive line enhancer (ALE) controlled by the normalized least mean square (NLMS) algorithm was implemented following the idea of [11], where a filter of this kind is successfully used in an airborne low-frequency SAR system. Autoregressive model (AR): This approach is commonly used for modeling narrowband signals [12]. The measured signal is modeled as an AR process where the UWB signal and system noise are considered white noise. RELAX: This method is an asymptotic maximum likelihood approach where the RFI sources are estimated in an iterative manner. The method has been recently implemented [13] for a FL-GPR system based on asynchronous stroboscopic sampling.

6 Singular Spectral Analysis (SSA): An approach based on subspace projection using singular spectral analysis was also implemented. Following the work of [14], we use the Column-Sampling approximation [15] to speed up the spectral decomposition. Notch Filter (NOTCH): A simple Notch Filter was also tested. This filter uses a CFAR detector to adapt the threshold for RFI suppression. A second version of this filter was also implemented. This version tries to exploit the idea of sniffer pulses by using only the last part of the data records (A-scans) to calculate RFI. Considering that a record length of 0.5µs should cover the expected range (<50m), thus with a length close to 1µs we can use the second part of the record to calculate the RFI. This approach seems more efficient than to lose one record every other PRF (200µs), and besides we assure that RFI remains more coherent within the same record. Average (AVG): Averaging or Stacking with different values is also considered and compared with the previous methods. It is interesting to note that this method of suppression does not take into account any particular properties of the RFI signal. We must say here that some filter parameters (filter order, data subdivision, estimated number of RFI ) have been adjusted manually to obtain the best possible results for the data under consideration. Some of these parameters (as the number of RFI) are of course site-dependent, but there are several adaptive approaches proposed in the literature to overcome this limitation as for instance the ones based on AIC (Akaike Information Criterion), ESTER (ESTimation Error) or SAMOS (Subspace-based Automatic Model Order Selection) among many others. 2.3 Performance indices We studied several performance indices to quantitatively evaluate algorithms and to compare them. For the sake of clarity we present here only three indices, two for each scenario. We use the Root Mean Square Error (RMSE) to obtain the residual energy after suppression. This index can be defined as in (1) where the ideal time-domain target response is compared with the target after RFI extraction. (1) being S r and S f the reference signal and the signal after RFI suppression respectively, whereas N represent the total number of samples. Besides residual energy it is important to evaluate the target distortion generated by the algorithms. For the 1D model we use the Spectral Angle Mapper (SAM) index to estimate this. SAM is a spectral classifier commonly used in hyperspectral imaging [16] and can be defined by (2). We find this index useful to compare the degree of correlation and distortion after RFI suppression. This index moves from 0 (perfect matching) to 1. ( ) (2) As mentioned before, the results for the 1D model are defined with the mean and variance of the performance indices for the 1000 data records generated to yield statistically valid results. For the 3D case we evaluate the target distortion of the 2D slices using the structural similarity index (SSIM), a method for measuring the similarity between two images. SSIM considers image degradation as perceived change in structural information. Further information about the SSIM index can be found in [17]. This index moves from 1 (perfect matching) to 0. (3) where µ r, µ f, σ r, σ f are the mean and variance of the reference signal and the signal after RFI suppression whereas σ rf represents the covariance. The parameters c 1 and c 2 help to stabilize the division.

7 3. RESULTS Fig.6 shows the results from the 1D model. Every subplot in this figure contains three A-Scans (after range compression). One of these A-scans corresponds to the simulated data without RFI. The remaining two represent the same A-Scan embedded in the recorded RFI before and after interference suppression. Table 1 shows quantitative results comparing the performance indices for the different methodologies. RSME and SAM are expressed in terms of their mean and variance considering all the traces evaluated (1000 A-Scans). From the plots we can directly infer how adaptive filters as ALE or AR-modeling have some problems to converge in case of multiple consecutive targets. ALE can be improved using sidelobe reduction procedures (ALE + SLR) as the one proposed in [1], where the output of the LMS filter is subtracted from the original signal and passed again through the filter. The other tested approaches seem to perform as expected: the three targets are readily visible and the RFI is reduced to a greater or lesser extent. Table 1 highlights the slight differences between methodologies in more detail. At the bottom of the Table 1 the results from data averaging are also included. Averaging is performed in groups of 10, 20, 40, 60 and 100 consecutive traces. It is interesting to note that, even with only 10 traces, there is already an improvement with respect to the other algorithms. A question that may arise from these results is whether a combined strategy of filtering + averaging may lead to better results (faster convergence) than using averaging alone. This approach is utilized by some authors [13], but in our case the obtained results showed no improvements with respect to the case of averaging alone. Coming to the results for the 3D model, Fig. 7 shows a horizontal slice of the ground at 0.15m depth used as reference to compare the performance of the different methods after data imaging. The two subplots at the top represent the synthetic data alone (left) and the same data after being embedded in the recorded RFI (right). The other subplots in this figure correspond to results after RFI suppression. Table 2 shows quantitatively the differences using the performance indices. From the plot in Fig. 7b we can appreciate how targets are not visible in the image because the DVB-T / DAB signal interference added to the data. Besides, this RFI generates image artifacts that could be erroneously cataloged as potential targets. Energy at the top of the slice is more significant than in the rest of the image. This effect is due to the fact that the RFI was mainly coming from one side of the FLGPR system when it was recorded in the test site. This example clarifies the importance of RFI suppression methods for these systems. The rest of the plots show the results using the different algorithms. Most of them allow for distinguishing the three targets but they also generate image artifacts. Note how some of these artifacts are already present in the raw data (Fig. 7a) but in that case with considerably less energy than the targets. In the 3D simulation the sidelobe reduction approach applied after ALE does not improve results in the image so the results from this method are not shown here. Care must be taken also when working with iterative algorithms as RELAX under variable RFI conditions. Either optimum suppression parameters or convergence restrictions might be sometimes relaxed to help the algorithm to find acceptable solutions in any case. Same as in the 1D simulation case, averaging obtains meaningful results. An average of 40 (Fig. 7i) already provides a radar image very similar to the original and without significant artifacts. It is interesting to point here how the RSME index is sometimes not a perfect indicator of the suppression method capabilities and the visual information provided to an eventual final user. An averaging of 40 (AVG 40) and AR modeling have similar RMSE but show clear differences evidenced in Fig 7 and also by the SSIM index. 4. DISCUSSION AND CONCLUSIONS In this paper we have shown using some examples how RFI generated by DVB-T and DAB transmitters can be an important noise source in FLGPR systems. We have discussed several approaches, used by the research community in similar contexts, and have applied them to subtract RFI from receiver radar data. We have attempted to evaluate the performance of these methods by simulating two simple scenarios submitted to real RFI and monitoring performance indices after suppression. We have shown how, in general, the different signal processing approaches improve the quality of the data. The good performance of signal averaging (stacking) is someway not surprising due to the nature of the RFI under consideration (OFDM based). Nevertheless, depending on the operational constraints and system capabilities might be more convenient to apply any of the

8 other suppression algorithms proposed. In that case subspace filters have proven to perform very efficiently according to the results. It is interesting to consider that averaging is also implicit in the radar imaging process where all measured data is generally added coherently at one focal point for all points of interest. Thus, the use of wide synthetic apertures for radar imaging might elude the need for averaging of every single trace and therefore reduce system acquisition times. 5. REFERENCES [1] Lord, R.T, Inggs, M.R., "Approaches to RF interference suppression for VHF/UHF synthetic aperture radar" Proceedings of the 1998 South African Symposium on Communications and Signal Processing (COMSIG). vol., no., pp , 7-8 Sep 1998, doi: /COMSIG [2] Saarenketo, T., Electrical properties of Road Materials and Subgrade Soils and the Use of Ground Penetrating Radar in Traffic Infrastructure Surveys. PhD Thesis. Acta Universitas Ouluensis, A471, 121pp, [3] Roberts, R., Feigin, J., Parrillo, R., "Mitigation of RF interference in air-launched 2 GHz GPR antennas" 13th International Conference on GPR, pp.1-6, June 2010 doi: /icgpr [4] Nguyen, L.H.; Tran, T.D., "Robust and adaptive extraction of RFI signals from ultra-wideband radar data" IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2012, pp , 2012, doi: /IGARSS [5] H. L., Van Trees; Optimum Array Processing, Part IV of Detection, Estimation and Modulation Theory, New York, Wiley, [6] Miller, T.; Potter, L.; McCorkle, J., "RFI suppression for ultra wideband radar," IEEE Transactions on Aerospace and Electronic Systems, vol.33, no.4, pp , [7] Jenho Tsao; Steinberg, B.D., "Reduction of sidelobe and speckle artifacts in microwave imaging: the CLEAN technique," IEEE Transactions on Antennas and Propagation, vol.36, no.4, pp , 1988 doi: / [8] J. Li, D. Zheng, Stoica, P., "Angle and waveform estimation via RELAX," IEEE Transactions on Aerospace and Electronic Systems, vol.33, no.3, pp , 1997, doi: / [9] L. H. Nguyen, T. Ton, D. Wong, M. Soumekh, Adaptive coherent suppression of multiple wide-bandwidth RFI sources in SAR. Proc. SPIE 5427, Algorithms for Synthetic Aperture Radar Imagery XI, 1, 2004, doi: / [10] Godara, L.C., "Application of antenna arrays to mobile communications. II. Beam-forming and directionof-arrival considerations," Proceedings of the IEEE, vol.85, no.8, pp , 1997; doi: / [11] Vu, V.-T.; Sjogren, T.K.; Pettersson, M.I.; Håkansson, L.; Gustavsson, A.; Ulander, L. M H, "RFI Suppression in Ultrawideband SAR Using an Adaptive Line Enhancer," Geoscience and Remote Sensing Letters, IEEE, vol.7, no.4, pp.694,698, Oct [12] M. Braunstein, J. M. Ralston, D. A. Sparrow, "Signal processing approaches to radio frequency interference (RFI) suppression", Proc. SPIE 2230, Algorithms for Synthetic Aperture Radar Imagery, 190 (1994); doi: / ; [13] O. Ojowu Jr., J. Li, 2013, RFI Suppression for Synchronous Impulse Reconstruction UWB Radar Using RELAX. International Journal of Remote Sensing Applications Volume 3 Issue 1, March [14] A Wang, Xiaoyu, A Yu, Weidong, A Qi, Xiangyang, A Liu, Yue, RFI suppression in SAR based on filtering interpretation of SSA and fast implementation EURASIP Journal on Advances in Signal Processing 2012, doi: / [15] S. Kumor, M. Mohri, A. Talwalkar, On sampling-based approximate spectral decomposition. Proc 26th Annual International Conference on Machine Learning, Montreal, Canada, (2009). [16] Carvalho Jr., O.; Guimaraes, R.; Gomes, R.; de Carvalho, A.; da Silva, N.; Martins, E., "Spectral Multiple Correlation Mapper," IEEE International Conference on Geoscience and Remote Sensing Symposium, IGARSS 2006, vol., no., pp , doi: /IGARSS [17] Zhou Wang; Bovik, A.C.; Sheikh, H.R.; Simoncelli, E.P., "Image quality assessment: from error visibility to structural similarity," IEEE Transactions on Image Processing, vol.13, no.4, pp , April 2004; doi: /TIP (Author provides a free version of a MATLAB implementation of the SSIM index available online at:

9 a) LMS (ALE) b) LMS (ALE + SLR) c) AR NOISE d) RELAX e) SSA f) NOTCH FILTER g) NOTCH FILTER (SNIFF) Table I Mean and Variance of the RMSE and SAM indices (1000 traces) h) AVG 20 Figure 6. RFI suppression results for the 1D model: a) Adaptive Line Enhancer (ALE); b) Adaptive Line Enhancer with sidelobe suppression (ALE+SLR); c) AR modeling; d) RELAX; e) Singular Spectral Analysis (SSA); f) Notch Filter; g) Notch Filter using the last half of the data to determine RF interferers (SNIFF); h) Averaging of 20 consecutive records.

10 a) RAW DATA b) DATA + RFI c) AR NOISE d) LMS (ALE) e) SSA f) RELAX g) NOTCH FILTER Table II RMSE and SSIM indices h) AVG 20 i) AVG 40 Figure 7. RFI suppression results for the 3D model considering a slice of the soil at 15cm depth after backprojection; a) Raw data (simulated); b) Data with recorded RFI; c) Data after AR; d) Data after ALE; e) Data after SSA; f) Data after RELAX; g) Data after Notch Filtering; h) Data with averaging of 20; i) Data with averaging of 40.