Accuracy assessment of SAR interferometry using the ERS-1

Accuracy assessment of SAR interferometry using the ER§- 1 D. Carrasco, J. Alonso, A. Broquetas Antcniias, Microwaves and Radar Group Dcpartnicnt of Signal Tlicory and Communications Universidad Polilccnica de Ciitaluiia Gran Capitrin shi Campus Norte UPC - Modulo D3. 08071 Barcclona (Spain) Tcl: +33 3 4016849. Fax: +33 3 3017232, e-mail: [email protected] Abstract -- A SAR raw data simulator and a SAR processor are uscd as a tool for pcrformance assessment of SAR interferometry algorithms like the multi-baseline aiid multifrequency. Traditional algorithms have also been tested with real ERS-1 data and validated uith a high accuracy refereiicc DEM. INTRODUCTION Obtaining digital clc~~ation models (DEM) \\ it11 interferometric techniques has bccome oiic of the major potential applications of SAR satellites The pcrformance assessment of SAR Interferometry (InSAR) IS oftcn difficult by its inherent compleiitj , the multiple sourccs of error (decorrelation, la! over, elc ) aiid the lack of suflicientl) accurate referciice DEM ox er suitable rcgions For this reasons, a raw data interfcromctric SAR simulator has been developed 111, vliich has alloued a detailed stud) of thc InSAR algoritlims pcrformancc Given a reference DEM and a radar rcflcctivit) map, the simulator obtains for a gn en sensor flight the hologram or ra\\-data Tlic ra\\ -data IS then proccssed \tilth a phase prcscrving algorithm Finall!. thc interferomctric proccss IS pcrfornicd The simulator has bceii used to dclclop and improve the pliase-un\\rrapping procedures, and to study the multibaseline and multibasclinc approaches Traditional single baseline techniques have also been tcstcd 11 ith real ERS-1 data The restills haw bccn \ alidatcd through comparison inth a high accuracy rcfcrcncc DEM This allons the dctcction of unc\pcctcd crrors as I\ 111 bc sliou II THE SAR SIMULATOR The SAR r:iw data simulator IS divided in tno blocks. The first one, pro-jccts the ground scatterers o\cr the slant-range plane and tlie sccond one generates the hologram or raw data. Tlie first block starting point arc thc satcllntc orbital paramclers and a DEM and a dielectric charactcrizallon of the area under tcst From thc DEM. tlic Earth surfr-ice is modcllcd 111 frlccts Each facct IS asslgned a reflection coefficient dcpcnding on its distaiicc to thc satellite, its roughness (n hich IS modellcd through the bcamu idth of the radiation pattern) aiid 11s diclcctric propcrtics Spcckle IS 0-7803-2.567-2195 $4 00 0 1995 IEEE 78 1 simulatcdl by adding coherently inine facets (independent scattercrs) within the same resolution cell. Each facet is obtained from the DEM adding a gaussian height which randomii.es the reflectivity of tlhe point scatterers and generates speckle. Finally, every resolution cell is projected over thc slant range plane so that foreshortening, layover aiid shadowing phcnomcna appear naturally. The sccond block of thc simulator generates the raw data from the slant-range rcflcctiaity map. The hologram can be generated by con\rolving the reflcctivity map with the SAR impulse response (gcneratcd by a rcflectivity delta). The calculation is implemcntcd in the transformed Fourier domain. The main problem is that rhe SAR impulse response is range li'ariant, but this can be solved by preprocessing the rcflcctivity map uhich can thcn be multiplied with the nonvariant SAR impulse rcsponse anid then two-dimensional rc\wse Fourier transformed to obtain the final hologram. Qncc ne havc the r:iu data. we can feed it into a phasepresening SAR processor and obtain the image. The combination SAR proccssor-simulator is a very powerful tool which \':e lim~e uscd to test new SAR interferometry algorithms. INTERFEROMETRY WITH SIMULATED DATA We have uorkcd U ith the multibaseline and multifrequency approachcs [2] Both nictliods lie on tlie same principle: gcneratc two interferograms it11 different fringe frequency, U hich I\ 111 lead to an unambiguous phase unwrapping Since the inteafcromctric phase IS proportional to the frequency and the basclinc separation, bj changing any of them we can modify the fringe frequency. In thc multifrequency approach, tito iiiterfcrograms are generated at tno different frecuencies (3 SAR images are nccdcd). The multibaseline algorithm requires only three imagcs and it may be applied to real ERS-1 dilta Tlic SAR simulator has alloivcd us to test these procedures. As an csamplc, the rcsults of a simulation over an area in Cataloni,a \villi the multifrcqucncy unwrapping algorithm are shoivii. In Fig. I. one of tlie intcrfcrograms obtaincd from the proccsscd SAR raw data is shown. The final DEM is prescntcd in Fig.2. Tlie :werage absolute error is 9.1 ni and lhc rnis crror is 15.6 111. Most errors are concentrated in the hilly arcx INTERFEROMETRY WITH REAL ERS-1 DATA We have carried out the interferometric processing over a small area in Tarragona (Spain) very close to Ebro river. Our first aim was to validate the resulting DEM with a high accuracy reference DEM since a comparison uzith a topographic map could hide systematic errors and would not allow a real validation of the interferometric process. The images were acquired on 12'" and 15" septcmbcr 1991 during the commissioning phase (3 daj repeat cycle) and the baseline perpendicular component was 160 m. First of all, registration of the two images was performed in two stages. As a first step, a coarse registration was done through cross-correlation of the amplitude of two windows in each image. The amplitude \\.as quantized to three levels in order to achieve a higher peak of the cross-correlation. Secondly, a fine registration \\,as performed in the four corners of the image interpolating and shifting one image seeking for a coherence maximum. The shift applied to the center of the image was interpolatcd from the \.slues of thc four corners. Both images were filtered in the range direction since the spectrums of the two images are shifted and only the common band leads to the generation of fringes [3]. The spectral shift is calculated by spectrum cross-correlation of the two images. Then each image is filtered Ica\.ing only the common frequency band. Notice that since the selected area is small and there is a significative average slope. \\.e can carry out this process o\.er the whole iniage without slopcdcpcndant segmentation. After the filtering. some spatial rcsolution is lost, but the coherence is higher. Range filtering lowers bandwidth a 15.7 '% and spatial resolution in the slant range direction lowers from 8.58 ni to 9.93 in. Now the complex interferogram ma! be constnicted b) niultiplicating one image by the complex conjugale of tlic other image The phase nAl show the interfcronietric fringes From the orbit data and the Earth coordinates of the image, we calculate for each pixel of the interferogram the distance to the satellite assuming a spherical Earth. This will allow us to substract the phase due not to the height variations but to the flat contribution of the ground. The phase is then aicraged with a pyramidal filter (9 pixels in azimuth by 5 pisels in range). Resolution is now 17.9 m in azimuth and 19.8 in in range. The filtered phase is shown in Fig. 3. It is very smooth and the phase unwrapping was not critical since the filtering reduced notably the noise level. The phase univrapping method is based in the residue theory [4]. Residues are joined in pairs through ghost lines which are not to be crossed when integrating the phase. Once the heights are known, we can project them over the ground. The resulting DEM was compared to a high accuracy reference DEM which has 2 m rms height error and a 15 ni by 15 m grid. We selected three local maxima in the InSAR DEM and looked for their equivalents in the reference DEM. With this illformation, the high accuracy DEM was stretched and rotated so that it sized the SAR DEM. Then, the original heights in the SAR DEM were restored by adding a constant in order to achieve the save average height in both DEM. The absolute value of the difference between both DEM was computed and the result is shown in Fig. 4. Notice that the error is minimum in the image center and grows linearly in the azimuth direction. We haw a linear component added to the original DEM. This is due to a spectral shift of the two SAR images in the azimuth direction which can be caused by a different Doppler ccntroid in the two images [2]. Tlus problem vas only dctected by checking the obtained DEM \i ith the reference DEM This fact stresses the need of ground control points in the image of known height in order to correct SI stcmatic errors, as the one shown before, when no rcfercnce DEM IS mailable Finally. the spectral shift in the azimuth was eliminated and the final DEM is shown in Fig. 5. The error between this emp-1 dot - Fig. 1. Interferometric frlnges from simulated SAR rau data Fig 2 DEM obtaincd by the multifrequency method. 782 corrcctcd DEM and de rcfcrencc DEM is shown in Fig. 6. 70 % of the points in the obtained DEM show an crror under 10 ni and the 95 96 is under 20 m. The rnis error is 10.36 m and the nicdium crror is 8.27 ni. CONCLUSIONS A SAR simulator-processor has bccn presentcd as a tool for SAR interferometry algorithms tcst. Multifrequency results have been prescntcd. Finally. it is sho\\~n hou a rcfcrence DEM can hclp in detecting systematic crrors. Ground control points should be uscd when it is not available. ACKNOWLEDGMENTS This work has bccn supported by tlic Coni. Intcrministcrial dc Cicncia y Tccnologia CICYT TIC 92-0615. REFERENCES [l] G. Franiccschctti, D. Riccio, "SARAS: a synthetic aperture radar (SAR) raw signal simulator", IEEE Trans. on Gcoscicncc and Rcmole Sensing, vol 30, no 1, pp. 110-122, Jan. 1992. [2] J. Alonso, "Interferonietria SAR mediante diversidad en frccucncia y cn lineas de base", Proyecto Final de Carrera, ETS dc Ingenicria de Tclecomunicacion de Barcelona, 1995. [3] F. Gatclli et al.. "The wavenumber shift in SAR interfcromctry", IEEE Trans. on Geoscience and Remote Sensing, vol. 32, no 4. pp. 855-864, July 1994. [1] R. Goldstein, H. Zebker, C. Werner, "Satellite radar interferonictry: two-dimensional phase unwrapping", Radio Scicncc, vol. 23, no 4. pp. 713-720, July-August 1988. The authors wish to thank the ESA ERS-1 Fringe Working Group for providing the SAR images and the Institut CartogrBfic de Catalunya for supplying the reference DEM. 'tor-2 dot' - Fig. 3. Intcrfcromctric fringcs filtered and ground corrected. Fig. 5. Final DEM obtained. ICrLlrint - 'tar-2e.dot' - Fig 1 Absolute crror bct\\ccn tlic lnSAR DEM and the rcfcrcncc DEM rcfcrcncc DEM Fig 6 ,4bsolutc crror betuecn thc final InSAR DEM and the 783