IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 2, MARCH 2011 331

Polarization Phase Difference Analysis for Selectionof Persistent Scatterers in SAR Interferometry

Sergey Samsonov and Kristy Tiampo

Abstract—In this letter, we propose a technique for selectingpersistent scatterers (PSs) based on their polarization phase dif-ference (PPD). We analyze a normalized PPD between HH and VVchannels averaged over a temporal set of images and select pixelsthat demonstrate predominantly even or odd bounce scatteringproperties. We compare selected scatterers to PSs selected byapplying an amplitude dispersion threshold as suggested by astandard PS interferometry (PSI) approach and show that bothmethods are complementary. However, the proposed approach canbe potentially used on a small set of synthetic aperture radar(SAR) images, which can be beneficial in the early stage of dataacquisition. We apply the proposed technique to produce a de-formation map for the San Francisco region from six quad-polRADARSAT-2 SAR images acquired during 2008–2009. The cov-erage and the precision of the produced deformation map arehigher than if it was calculated with the standard PSI techniqueapplied to the same data set.

Index Terms—InSAR, interferometry, persistent scatterers(PSs), polarization phase difference (PPD), synthetic apertureradar (SAR).

I. INTRODUCTION

D IFFERENTIAL synthetic aperture radar interferometry(DInSAR) is a technique for measuring ground defor-

mation with a high spatial resolution and accuracy over alarge area. The differential interferogram is calculated fromtwo synthetic aperture radar (SAR) images acquired by thesame or a similar satellite at two different times. The inter-ferometric processing consists of the following steps: slave tomaster coregistration and resampling, interferogram formation,removal of topographic and Earth curvature phases, filtering,and phase unwrapping [17], [24]. The calculated interfero-gram is an image of the line-of-sight (LOS) deformation thatoccurred between the two acquisitions. This image often iscontaminated with residual topographic and atmospheric noise[25]. At present, DInSAR is routinely used for mapping tec-tonic or seismic deformation [9], [26], [30], volcanic activity[5], [15], and various anthropogenic signals [25].

The quality of an interferogram is described by itscoherence, which is the magnitude of a cross-correlationcoefficient calculated for a master–slave pair of images. The

Manuscript received June 17, 2010; revised August 9, 2010; acceptedAugust 26, 2010. Date of publication October 7, 2010; date of current ver-sion February 25, 2011. This work was supported by the Natural Sciencesand Engineering Research Council of Canada and Aon Benfield/Institute forCatastrophic Loss Reduction Industrial Research Chair in Earthquake HazardAssessment.

The authors are with Department of Earth Sciences, University of WesternOntario, London, ON N6A 5B7, Canada (e-mail: ssamson@uwo.ca).

Digital Object Identifier 10.1109/LGRS.2010.2072904

coherence depends on a variety of parameters, including SARwavelength, incidence angle, perpendicular baseline, and landcover. In a densely vegetated environment, the coherencegenerally is low, which prevents the formation of a high-qualityinterferogram and phase unwrapping. For these land coverconditions, a persistent scatterer (PS) interferometry techniqueis used instead [7], [8]. The PS approach is based on a selectionof pixels that are dominated by a single strong scatterer withinthe pixel and which are consistent over a long period of time.These pixels are usually selected based on their amplitude oramplitude dispersion. Currently, PS interferometry (PSI) is anestablished technique that has proven successful for mappingground deformation of various origins on a worldwide scale[6], [12], [13].

The standard PSI approach has also a few disadvantages. Inorder to confidently select PS pixels, it is necessary to have alarge number of images. In general, the selection of PS pixelswith a high degree of accuracy is possible only when at least 30SAR images are available [8]. This is not always possible or de-sirable. Not only it is often difficult or expensive to acquire largenumbers of images but also many images inevitably span largetime periods, on the order of years to decades. Hazard estimatesfor volcanic or seismic monitoring, for example, often requirethe production of time series and deformation rates on shorttime frames, i.e., after the acquisition of only a few images. Inaddition, the selection of PS pixels is limited to those pixelsthat have a very high amplitude or a low standard deviation orsatisfy both conditions. In some regions, the network of suchPS pixels is very sparse, which limits the processing of suchinterferograms, particularly the phase unwrapping. A differentgroup of the advanced DInSAR algorithms [1], [19] can be usedto perform a time series analysis using a smaller number ofimages. However, these algorithms are generally not based onthe PS approach and produce limited results in regions of lowcoherence.

The advantages of polarimetric DInSAR for mappingground deformation were previously investigated in [22] for aground-based quad-pol SAR sensor and in [20] for dual-polTerraSAR-X. There, it was suggested that the calculation ofcoherence and amplitude dispersion for the selection of reliablepixels can be significantly improved in the case of polarimetricSAR data. In this letter, we demonstrate a technique that isable to select a large number of PSs from a small temporalset of SAR images by analyzing their copol phase difference[4], [16], [18], [28]. We designate these polarization phasedifference scatterers (PPDSs). We successfully select pixelswith predominant even or odd bounce scattering mechanismsand exclude pixels with diffusive scattering caused by inter-action with vegetation. We also show that the majority of thePS pixels selected with the standard approach have primarily

1545-598X/$26.00 © 2010 IEEE

332 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 2, MARCH 2011

even and odd bounce scattering mechanisms, suggesting thatthe two techniques are complementary. The proposed approachrequires simultaneously acquired HH and VV SAR imageswith a preserved relative phase that are presently availablefrom only a few satellites, such as RADARSAT-2, AdvancedLand Observing Satellite Phased-Array-Type L-Band SAR, andTerraSAR-X. It is anticipated that the availability and usage ofpolarimetric data will improve in the future, and the techniqueproposed in this letter will become increasingly valuable fordeformation mapping.

In this letter, we apply the proposed technique in order toproduce a deformation map for a region along the northernHayward Fault from six quad-pol RADARSAT-2 SAR imagesacquired during a nine-month time period from April 26, 2008,to January 15, 2009. In recent years, it has been determined thatmovement along this portion of the Hayward Fault is unlocked,dominated primarily by a creeping motion in the horizontaldirection, while slow-moving landslides have been identifiedin the hills along the fault scarp [2], [3], [10], [11], [14], [31].These results were produced from a selection of anywhere from30 to 49 images for the PSI analysis over time periods thatranged from 1992 to 2001 [2], [10], [14]. In addition, [27],13 stacked interferograms over the same period produced sim-ilar results for the west side of the Hayward Fault but werelimited by the lack of coherence along the hills on the easternside. Here, we employ this new PPDS technique to estimate theshort-term time-dependent deformation in this region for betterquantification of the tectonic structure and regional hazardestimates. The earlier results then can be compared with thosefrom this new technique, which are acquired over shorter timeperiods and increased spatial details.

II. METHODOLOGY

In the standard PSI analysis, PSs are selected based on theiramplitude dispersion, which is calculated as

D =σ

A(1)

where σ is the standard deviation and A is the mean amplitudecalculated for the same pixel of the coregistered set of SARimages. For a sufficiently large set of images, the pixels withD < 0.25 are said to be persistent and are selected for furtherprocessing, and pixels with D > 0.25 are excluded [8]. Thepolarization phase difference (PPD) Δφ for each pixel is cal-culated in the following way:

Δφ = φHH − φVV (2)

where φHH is the phase of a wave transmitted and receivedin horizontal polarization and φVV is the phase of a wavetransmitted and received in vertical polarization. The extremevalues of the PPD which are equal to either zero and ±πcorrespond to deterministic odd and even bounce scatterers,e.g., scatterers with a predominant reflective mechanism [21],[29]. The PPD value diverges as the contribution from thediffusive scattering increases [4], [28]. The diffusive scatteringfrom vegetation produces PPD values that are randomly dis-tributed in [−π, π]. Thus, by selecting pixels with appropriatePPD values, it is possible to exclude pixels with a diffusivescattering mechanism caused by the interaction with vegetation.

The phase information of such pixels varies in an unpredictableway from one acquisition to another and, therefore, cannot beused for mapping ground deformation.

In this letter, we use the normalized PPD for the selection ofreliable scatterers whose phase information carries consistentinformation. This can be done by calculating, for each pixel,a normalized average of absolute values of the PPD for atemporal set of SAR images and by selecting pixels with valuesclose to zero and one

χ =

∑Kk=1 |Δφk|Kπ

(3)

where K is the number of a SAR image used for processing.This quantity ranges from zero to one where χ = 0 corre-

sponds to pixels with strictly odd bounce scattering and χ = 1corresponds to pixels with a strictly even bounce scatteringmechanism. Furthermore, in order to exclude single-bouncepixels corresponding to a reflection from the water surface, aswhat occurred in the San Francisco images, it is recommendedto set a threshold value on the pixel amplitude, which alsoexcludes pixels with a large number of bounces (e.g., morethan two). For example, it is possible to select only those pixelsthat have intensity values larger than the average value for thewhole image. However, the threshold value will vary dependingon the amount of water presented in the SAR image. Furtheranalysis is required in order determine the best method for theproper selection of the appropriate threshold values for χ, andit may be that this threshold value will depend on the type ofland cover, satellite wavelength, incidence angle, and numberof images available.

III. RESULTS

For this letter, we collected six fine quad polarization (FQ7)RADARSAT-2 images acquired on April 26, June 13, July 7and 31, and December 22, 2008, and January 15, 2009, overa region along the Hayward Fault located northeast of SanFrancisco. The land cover in the area covered by this imageconsists of regions of dense vegetation, urban areas, and openwater. The PPD reliable scatterers were selected with χ lessthan 0.2 and greater than 0.8, and the average intensity wasset at twice the mean intensity for the whole image. Theseresults are shown in Fig. 1. In total, 252 000 single-bouncepixels and 69 000 double-bounce pixels were selected, whichcorresponds to about 1.34% and 0.37% of the total number ofavailable pixels. For comparison purposes, the number of PSpixels selected based on the amplitude dispersion (less than 0.2)was equal to 176 000 or 0.94%.

Standard interferometric processing was performed on theimages, and the topographic phase was removed using a 10-mNational Elevation Dataset digital elevation model provided bythe U.S. Geological Survey. In total, we created 15 differentialinterferograms with a maximum perpendicular baseline close to400 m and time spans ranging from 24 days to approximately9 months. Each interferogram was unwrapped, and all 15 inter-ferograms were used to solve for the mean deformation ratesand the residual topographic error for each pixel independently.Similar processing was performed for a set of images with thePS pixels selected based on the amplitude dispersion.

SAMSONOV AND TIAMPO: POLARIZATIONS PHASE DIFFERENCE ANALYSIS 333

Fig. 1. Selection of reliable scatterers based on PPD and amplitude dispersion. (a) Single bounce. (b) Double bounce. (c) Single and double bounce. (d) Amplitudedispersion (less than 0.2). These images were acquired by right-looking SAR from descending orbit.

Fig. 2. Differential wrapped interferograms (full color range corresponds to [−π, π], in YYYYMMDD format and B⊥): (a) 20080426–20080613, −156 m;(b) 20080707–20080731, 68 m; (c) 20081222–20090115, 18 m; (d) 20080426–20090115, −133 m. These images were acquired by right-looking SAR fromdescending orbit.

The examples of the wrapped differential interferogramsare presented in Fig. 2 (in YYYYMMDD format) with thecorresponding values of the perpendicular baseline: Fig. 2(a)20080426–20080613, −156 m; Fig. 2(b) 20080707–20080731,68 m; Fig. 2(c) 20081222–20090115, 18 m; and Fig. 2(d)

20080426–20090115, −133 m. The interferograms spanningthe last time period 20081222–20090115 and 20080426–20090115 [(c)–(d)] show a clear signal around the fault area.However, more images acquired recently have been required toclassify the observed signal as the true ground deformation or

334 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 2, MARCH 2011

Fig. 3. Linear deformation rate calculated for reliable scatterers selected based on the following: (a) PPD and (b) amplitude dispersion.

Fig. 4. Distribution of pixels according to their amplitude dispersion andPPDs. Only pixels with amplitude five times greater than average amplitudefor the whole image were selected.

to exclude water vapor or soil moisture related artifacts. Cor-respondingly, the mean deformation rates presented in Fig. 3may not be the representative of long-term averages, in case ofa nonlinear or short-term deformation processes.

For these six SAR images, acquired over a time period of lessthan 9 months, we were able to calculate the mean deformationrates with an accuracy close to 0.49 cm/year. The same quantitycalculated for the PS pixels selected based on the amplitudedispersion is close to 0.54 cm/year. The mean deformationrates and standard deviation were estimated by applying linearregression [23] to the calculated time series, as described in[25]. This larger error in case of the standard PS analysis isprobably due to the presence of unwrapping errors in individualinterferograms resulting from the sparseness of the PS network.

In order to demonstrate that PPD can be used to describethe pixel reliability, we plotted the distribution of the pixelsaccording to their amplitude dispersion and PPDs (see Fig. 4).

Only pixels with a high amplitude were considered since it isassumed that, for these pixels, the amplitude dispersion basedon the six images is sufficiently accurate. According to thisfigure, the majority of the pixels with PPD values close to zero(odd bounce) and to one (even bounce scattering mechanism)also have a small amplitude dispersion. This suggests that bothtechniques (PS and PPDS) are complementary and can be usedfor cross validation, but the technique proposed in this letter canselect reliable pixels even for a small set of available images.

IV. CONCLUSION

In this letter, we have presented a methodology for theselection of reliable scatterers based on their PPD. Those pixelswith a normalized average of absolute values of the PPDclose to zero and one are considered reliable because they aredominated by odd and even bounce scattering mechanisms.Diffusive scattering from vegetation produces pixels with thePPD randomly distributed in the [−π, π] interval and withnormalized average PPD values far from zero or one. Suchpixels are excluded from the interferometric processing.

The analysis of the distribution of pixels according to theiramplitude dispersion and PPDs has suggested that the majorityof pixels that would be selected by a standard PS approachalso have normalized PPDS values close to zero or one and,therefore, will be selected by the proposed technique also.However, as anticipated, the proposed technique has resultedin greater coverage. In addition, it expected that the method ismore accurate, even for a smaller set of SAR images, althoughthis will be evaluated further when at least 30 RADARSAT-2images suitable for interferometric studies are available andwhen the amplitude dispersion can be considered a reliableparameter for the selection of the PS.

The proposed technique was used to compute the meandeformation rates for a region along the Hayward Fault

SAMSONOV AND TIAMPO: POLARIZATIONS PHASE DIFFERENCE ANALYSIS 335

northeast of San Francisco, which is partially covered by densevegetation. We were able to achieve subcentimeter precisionof the measured deformation rates, and the precision of eachindividual interferogram is high. The isolated signals near theHayward Fault are visible, which possibly corresponds to thoseidentified by [11] in a study of local hillslope motions thatspanned 2 years and 17 images. In addition, the average hor-izontal deformation rates along each side of the fault comparefavorably with those obtained by others [10], [14] despite thefact that their long-term averages are calculated for many moreimages (from 30 to 49) from a period exceeding ten years.Smaller scale differences may be the result of nonlinear orintermittent variation in the creep rate across the fault that areidentifiable only over the shorter time periods in this letter,confirming that one of the primary strengths of this method isthe ability to quantify motion over time periods on the orderof one year with the accuracy of longer term PS-InSAR tech-niques. Future work will concentrate on determining the limitsof spatial and temporal accuracy as a function of the number ofavailable images. In addition, analysis for this particular regionwill be repeated when more SAR images are available, in orderto better quantify these short-term variations.

ACKNOWLEDGMENT

Some images were plotted with the Generic Mapping Toolssoftware.

REFERENCES

[1] P. Berardino, G. Fornaro, and R. Lanari, “A new algorithm for surfacedeformation monitoring based on small baseline differential SAR inter-ferograms,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, pp. 2375–2383, Nov. 2002.

[2] R. Burgmann, G. Hilley, A. Ferretti, and F. Novali, “Resolving verticaltectonics in the San Francisco Bay Area from permanent scatterer InSARand GPS analysis,” Geology, vol. 34, no. 3, pp. 221–224, Mar. 2006.

[3] R. Burgmann, D. Schmidt, R. Nadeau, M. d’Alessio, E. Fielding,D. Manaker, T. McEvilly, and M. Murray, “Earthquake potential alongthe Northern Hayward Fault, California,” Science, vol. 289, no. 5482,pp. 1178–1182, Aug. 2000.

[4] D. L. Evans, T. G. Farr, J. J. van Zyl, and H. Zebker, “Radar polarime-try: Analysis tools and applications,” IEEE Trans. Geosci. Remote Sens.,vol. 26, no. 6, pp. 748–789, Nov. 1988.

[5] J. Fernandez, R. Romero, K. Carrasco, D. Tiampo, G. Rodriguez-Velasco,A. Aparicio, V. Arana, and F. Gonzalez-Matesanz, “Detection of displace-ments on Tenerife Island, Canaries, using radar interferometry,” Geophys.J. Int., vol. 160, no. 6, pp. 33–45, 2005.

[6] A. Ferretti, F. Novali, R. Burgmann, G. Hilley, and C. Prati, “InSARpermanent scatterer analysis reveals ups and downs in San Francisco BayArea,” EOS Trans. AGU, vol. 85, no. 34, p. 317, Aug. 2004.

[7] A. Ferretti, C. Prati, and F. Rocca, “Nonlinear subsidence rate estima-tion using permanent scatterers in differential SAR interferometry,” IEEETrans. Geosci. Remote Sens., vol. 38, no. 5, pp. 2202–2212, Sep. 2000.

[8] A. Ferretti, C. Prati, and F. Rocca, “Permanent scatterers in SAR inter-ferometry,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 1, pp. 8–20,Jan. 2001.

[9] Y. Fialko, “Interseismic strain accumulation and the earthquake potentialon the southern San Andreas Fault system,” Nature, vol. 441, no. 7096,pp. 968–971, Jun. 2006.

[10] G. Funning, R. Burgmann, A. Ferretti, F. Novali, and A. Fumagalli,“Creep on the Rodgers Creek Fault, northern San Francisco Bay Areafrom a 10 year PS-InSAR dataset,” Geophys. Res. Lett., vol. 34,p. L19 306, Oct. 2007.

[11] G. Hilley, R. Burgmann, A. Ferretti, F. Novali, and F. Rocca, “Dynamicsof slow-moving landslides from permanent scatterer analysis,” Science,vol. 304, no. 5679, pp. 1952–1955, Jun. 2004.

[12] A. Hooper, “A multi-temporal InSAR method incorporating both persis-tent scatterer and small baseline approaches,” Geophys. Res. Lett., vol. 35,no. 16, p. 16 302, Aug. 2008.

[13] A. Hooper, P. Segall, and H. Zebker, “Persistent scatterer InSARfor crustal deformation analysis, with application to Volcan Alcedo,Galapagos,” J. Geophys. Res., vol. 112, p. B07 407, 2007.

[14] R. Jolivet, R. Burgmann, and N. Houlie, “Geodetic exploration of theelastic properties across and within the northern San Andreas Fault zone,”Earth Planet. Sci. Lett., vol. 288, no. 1/2, pp. 126–131, Oct. 2009.

[15] O. Kwoun, Z. Lu, C. Neal, and C. Wicks, “Quiescent deformation of theAniakchak Caldera, Alaska, mapped by InSAR,” Geology, vol. 34, no. 1,pp. 5–8, Jan. 2006.

[16] J.-S. Lee and E. Pottier, Polarimetric Radar Imaging From Basics toApplications. Boca Raton, FL: CRC Press, 2009.

[17] D. Massonnet and K. Feigl, “Radar interferometry and its application tochanges in the Earth’s surface,” Rev. Geophys., vol. 36, no. 4, pp. 441–500, 1998.

[18] D. Massonnet and J.-C. Souyris, Imaging With Synthetic Aperture Radar.Lausanne, Switzerland: EPFL Press, 2008.

[19] O. Mora, J. Mallorqui, and A. Broquetas, “Linear and nonlinear terraindeformation maps from a reduced set of interferometric SAR images,”IEEE Trans. Geosci. Remote Sens., vol. 41, no. 10, pp. 2243–2253,Oct. 2003.

[20] V. D. Navarro-Sanchez, J. M. Lopez-Sanchez, and F. Vicente-Guijalba, “Acontribution of polarimetry to satellite differential SAR interferometry:Increasing the number of pixel candidates,” IEEE Geosci. Remote Sens.Lett., vol. 7, no. 2, pp. 276–280, Apr. 2010.

[21] D. Perissin and A. Ferretti, “Urban-target recognition by means of re-peated spaceborne SAR images,” IEEE Trans. Geosci. Remote Sens.,vol. 45, no. 12, pp. 4043–4058, Dec. 2007.

[22] L. Pipia, X. Fabregas, A. Aguasca, C. Lopez-Martinez, S. Duque,J. Mallorqui, and J. Marturia, “Polarimetric differential SAR interferome-try: First results with ground-based measurements,” IEEE Geosci. RemoteSens. Lett., vol. 6, no. 1, pp. 167–171, Jan. 2009.

[23] W. Press, S. Teukolsky, W. Vetterling, and B. Flannery, NumericalRecipes in C: The Art of Scientific Computing, 3rd ed. Cambridge, U.K.:Cambridge Univ. Press, 2007, p. 994.

[24] P. Rosen, P. Hensley, I. Joughin, F. Li, S. Madsen, E. Rodriguez, andR. Goldstein, “Synthetic aperture radar interferometry,” Proc. IEEE,vol. 88, no. 3, pp. 333–382, Mar. 2000.

[25] S. Samsonov, “Topographic correction for ALOS PALSAR interferome-try,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 7, pp. 3020–3027,Jul. 2010.

[26] S. Samsonov, K. Tiampo, and J. Rundle, “Application of DInSAR-GPSoptimization for derivation of three-dimensional surface motion of south-ern California region along the San Andreas Fault,” Comput. Geosci.,vol. 34, no. 5, pp. 503–514, May 2008.

[27] D. Schmidt, R. Bürgmann, R. Nadeau, and M. d’Alessio, “Distribution ofaseismic slip rate on the Hayward Fault inferred from seismic and geode-tic data,” J. Geophys. Res., vol. 110, no. B8, pp. B08 406-1–B08 406-15,Aug. 2005.

[28] F. T. Ulaby, D. Held, M. C. Dobson, K. McDonald, and T. Senior, “Relat-ing polarization phase difference of SAR signals to scene properties,” IEEETrans. Geosci. Remote Sens., vol. GRS-25, no. 1, pp. 83–92, Jan. 1987.

[29] J. van Zyl, “Unsupervised classification of scattering behavior using radarpolarimetry data,” IEEE Trans. Geosci. Remote Sens., vol. 27, no. 1,pp. 36–45, Jan. 1989.

[30] T. Wright, E. Fielding, and B. Parson, “Triggered slip: Observations of the17 August 1999 Izmit (Turkey) earthquake using radar interferometry,”Geophys. Res. Lett., vol. 28, no. 6, pp. 1079–1082, 2001.

[31] M. Wyss, “Locked and creeping patches along the Hayward Fault,California,” Geophys. Res. Lett., vol. 28, no. 18, pp. 3537–3540, 2001.