AUTOMATIC REGISTRATION FOR REPEAT-TRACK INSAR DATA PROCESSING Mingsheng LIAO, Li ZHANG, Zxn ZHANG, Jiangqing ZHANG Whan Technical University of Srveying and Mapping, Natinal Lab. for Information Eng. in Srveying, Mapping and Remote Sensing Loy Road 9, Whan 430079, P. R. China liao@wtsm.ed.cn Working Grop IC- KEY WORDS: InSAR, Registration, SAR processing, Atomatic algorithm, Remote Sensing. ABSTRACT Interferometric Synthetic Apertre Radar (InSAR) allows prodction of high resoltion DEM and detection of small earth motions sing mltiple pass SAR data sets obtained by remote sensing satellite. Bt the digital processing to extract the DEM is qite complicated. The whole procedre has not yet reached sfficient efficiency and robstness to warrant atomated DEM prodction as commonly prodced by stereo vision with optical images. The atomatic algorithm for precision registration is one of the bottlenecks for improving both the accracy and the efficiency of mlti-sorce data analysis, inclding the repeat-track D-InSAR processing. In this paper, an atomatic approach with mlti-step image matching algorithm is presented. All procedres cold be atomatically implemented. The primary experiment reslt is promising in the fast precision registration for the repeat-track InSAR data and reveals the potential of the presented atomatic strategy. INTRODUCTION InSAR presents a completely new approach for topographic mapping. Bt the whole process has not yet reached sfficient robstness to warrant atomated DEM prodction as commonly prodced by stereo vision with optical image (Trove 998). Moreover, the compting time shold be kept as short as possible for an operative data processing system. One of the critical factors in generating the interferograms is the efficient and precise registration of the single look complex data. To obtain a high qality interference fringe pattern, it is necessary for the images to be registered to sb-pixel accracy (Gabriel 988, Zebker 986). Usally image registration deals with geometrical transformation and resampling of the pixel vales. The identical points on both images can be identified as the control points. If sfficient control point pairs are determined, the transformation fnction cold be chosen. The new pixel vales are calclated according to the geometric transformation and the resample formlae (bi-linear or bi-cbic interpolation). The determination of sitable control point pairs has become the critical point. The traditional method for registration sally tilizes the tie points which are manally determined. However, the manal procedre constittes a tedios and time-consming task and may be impossible for SLC SAR data. Varios researchers have investigated the atomatic matching approach (Zhang 998). Bt it is mch more difficlt in SAR image matching than in traditional optical image since the speckle and the blr textre featre. Lin et al presented the registration algorithm based on the optimization of a target with the average flctation fnction of the phase difference image (Lin 99). Gabriel ea tl presented an approach based on the coherence estimation, which was adopted by many investigators (Gabriel 988). A two-dimensional fast Forier transform was performed on the interference fringe generated in the matching window, then the relative qality was assessed with the signal-noise ratio of the fringe spectrm. Zebker et al also sed small patches as tie points on ERS images and performed the D convoltions to find maximm power spectra. Offsets in range and azimth can be determined based on the difference of these patches (Zebker 994). Bt in fact, the power distribtion will differ greatly from the ideal case with a single peak. Mltiple peaks may appear, reslting in difficlty in correlation determination (Wang 990). In these algorithms, the generation of the fringe and the FFT are the very time consming task. Additionally the entire matching procedre is repeated as the search area was scanned pixel by pixel. In this paper, a mlti-step matching approach based on intensity or power detected images is presented. In the primary matching, the distinct points are selected and their conjgation points are determined atomatically based on the This research was fnded by National Natre Science Fondation of China (Contract No. 697800) and R&D Fondation for Srveying and Mapping Technology (Contract No.99008). 86 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000.
normalized correlation coefficient algorithm, which tolerates any linear radiometric relationship between the images. Then the global matching procedre based on the bridge mode and probability relaxation techniqe is sed to ensre the reliability of the matching reslt. The atomatic strategy has been sed for the registration between SPOT and TM images, which are acqired from different sensors (Zhang 998). Finally, the Least Sqare (LS) matching algorithm is sed to improve the precision (less than /0 pix.). The freqency independent coherence estimator (Garnieri 997) is sed as the matching criteria of LS algorithm. The experiment is carried ot sccessflly with SIR-C / L&C band data. METHODOLOGY. Primary Image The image matching is based on the measrement of similarity of image strctres or gray level distribtion between two images. At this scenario, only the intensity image of SLC data is sed in image matching procedre. Firstly the featre points are detected with Forstner interest operator in the master image (Wang 990), then the initial matching based on the similarity is condcted to find the candidates of the corresponding point in the slave image. The primary matching procedre employed normalized correlation coefficients between 0 and, which is defined as C( p,q) ρ = () C ( p,q)c ( p,q) gg g g Here C gg (p, q) and C g g (p, q) are respectively the variance of master image g(x, y) and slave image g ( x, y), C(p,q) is the covariance of g(x, y) and g (x, y). The closer to the vale of ρ, the more similar images are to each other. The advantage of sing the normalized correlation coefficient is that it tolerates any linear radiometric relationship between the two images. The comptational complexity is not obvios. Bt the single point matching does not take the compatibility of matching reslts with its neighbors into consideration. It is difficlt to improve the reliability for only the finite information contained in a matching window. In order to overcome the drawback, an algorithm of global matching with relaxation has been developed and frther sed to the second step of or registration procedre, which ensre the reliability of the matching reslt (Zhang 990).. Global Image with Probability Relaxation As well known, image matching is sed to define the corresponding point pair, that is the point j in slave image corresponding to the point i in the master image. Assme that there are an object set O={O, O,, O n } and a class set C={C, C,, C m }, the point i as an object, the points j in slave image are classif ied into class. If image matching is going to solve the problem of O i C j, the global consistency shold be considered. Therefore the probability of O i C j and the compatible coefficient C(i,j; h,k) of O i C j O h C k mst be defined according to the probability relaxation algorithm. Since the area based matching algorithms se the similarity of gray level distribtion as their measrement of image matching, the correlation coefficient ρ ij may be sed as the measrement of P ij of O i C j. Based on the principle of Figre. Probability relaxation with bridge model International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000. 87
bridge mode method as shown in Fig., the correlation coefficient ρ( ih, jk ) between the image segment ih in the master image and the image segment jk in the slave image can be sed as a measrement of the compatible coefficient C(i,j; h,k) of O i C j O h C k, i.e.: C(i,j;h,k) ρ( ih, jk ) () After P ij and C(i,j; h,k) are defined, the relaxation iterative procedre wold be indced, Q(i,j) = n(h) h = ( m(k) k = C(i,j;h,k) P(i,j) ) P (r) (i,j) = P (r ) (i,j) (+ B Q(i,j)) (r) P (i,j) = P (r) (i,j)/ m(j) P (r) (i,j) j (3) Here n(h) is the nmber of neighbor points (8- or 4-neighborhood), m(k) and m(j) is the nmber of the candidates. The scalar B is a constant with which the convergence rate of the iteration procedre cold be adjsted. The vale P(r) determines whether the iteration is stop. Finally the reliable point pair can be confirmed if P (r) > T, here T is the threshold..3 LS Procedre The relaxation algorithm discssed above can improve the reliability of the determination of the conjgate point. Bt it is difficlt to refine the matching accracy. It is necessary to take the frther procedres. As well-known, LS correlation method has sccessflly been developed, which cold achieve very high accracy of correlation by means of minimization of the root mean sqare vale of the gray vale differences of the image pair. Here the principle of minimization of grey vale differences between the two correlation windows is sed instead of the conventional approach based on maximization of the correlation coefficient (Rosenholm 986). We describe the general principle in one-dimensional case. Assme that there are two grey vale fnctions g (x) and g (x) with an image pair. In an ideal case, these two fnctions shold be the same except for a displacement x 0 between them. Their noises are n (x) and n (x) respectively. Then the observation fnctions are shown as (sbscript i representing the corresponding vales at pixel i ): g g (x i ) = g (x i ) + n(x i ) (x i ) = g (x i ) + n (x i ) = g (x i x 0 ) + n(x i ) g(x i ) = g (x i ) g (x i ) = g (x i x0 ) g (x i ) + n(x i ) n(x i ) (4) (5) Assme x 0 to be small, we linearize g (x i -x 0 ) in the st order term and write eqation(5) again as following: g( x i ) + v(xi ) = g& (x i ) x 0 (6) where v(xi ) = n (x i ) n (x ), which is the linearized error eqation. The observed vale is the gray vale difference g(x i ). The nknown vale is x 0. Using the method of Least Sqares, the following eqation is obtained and vale x 0 is ths solved: v (x ) = ( n (x ) n (x )) i i i = min (7) Practically, since each pixel pair can constitte an error eqation, there are M M observed vales within a M M matching window, which are redndant for the soltion of x 0. Becase the linearization exists in Eqation (6), the soltion has to be iterated. The principle can be easily generated into D image matching case. Herewith the initial vales for the iteration can be obtained from the reslt of the relaxation matching as described in Section.. Theoretically the accracy of determination for conjgate points in this method may reach the order of /50 to /00 pixel. Considering other factors, the accracy of /0 pixel for InSAR data registration will be satisfactory. 88 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000.
.4 The Criteria for LS Iterative Procedre Since the LS matching is an iterative procedre, the criteria for the matching shold be selected to decide whether the iterative is stopped. Herewith the correlation coefficient of the SLC images is sed as the criteria, which is sally adopted as the coherence measrement of fringe. The correlation coefficient is calclated according to γ = E[ E[ * ] ] E[ ] (8) Its estimation holds γ = ( n, * ( n, e jϕ( n, n m ( n, ( n, n m n m (9) where, are the SLC images to be registered, N M is the size of the estimation window. Bt it shold be noted that even if the ergodicity hypothesis holds, the accrate interferometric phase shold be estimated and compensated in expression (9). The complexity is highly increased. Therefore the freqency independent estimator is introdced. So the absolte vale of the complex coherence can be calclated as γ M ρ = 0 ρ > 0.5 ρ 0.5 (0) where ρ = ( n, ( n, n m n m 4 (, ) ( n, n m n m 4 () The estimator described in expression (0) and () is mch more efficient in comptational costs than the sal estimator, despite a redced statistical confidence. If the coherence is not very low and the estimation window is large, the accracy will be good enogh as the matching criteria (Garnieri 997). 3 EXPERIMENTS The mlti-step matching method is applied on the registration between an SLC image pair of SIC-C/C&L-band. All the procedres were implemented atomatically. The images cover the Hawaii area as shown partly in Fig.. The portion with image size 500 500 is sed as the test site. The image pair of L-band is shown in Figre. More than 00 point pairs have been determined within a short period of time with abot 0.5 to pixel RMS by global relaxation algorithm. The cross crsors marked in Fig. indicate the conjgate point pairs. Then the point pairs were inpt as the initial vales for the Least Sqare. After abot 5 LS iterations, the RMS of the global polynomial transformation is less than /0 pixel. The Table and Table show the reslts from L-band image pair and C-band image pair respectively. The window size for matching and coherence estimation is 63 63. So some points were cancelled becase they were very close the image edge. The geometric registration was performed by means of nd order polynomials. The accracy after LS matching is less than /0 pixel and available for InSAR data processing. Finally the fringes and coherence images were mlti-look processed with looks in range and looks in azimth. The fringes and coherence images are shown in Figre 3 and Figre 4 respectively. Abot 7 mintes was reqired for one image pair with image size 500 500 on Indigo- workstation. The program code will be optimized frther. International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000. 89
Figre. The L-band image pair Figre 3. (a) Fringe from the L-band image pair (b) Coherence image from the L-band image pair 90 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000.
(a) Fringe from the C-band image pair Figre 4. (b) Coherence image from the C-band image pair algorithm Only relaxation matching Least Sqare Nmber of Points RMS Error Maximm residal Range Azimth Total RMS Range Azimth Total residal 55 0.60 0.3 0.0.980 0.447.985 5 0.065 0.06 0.063 0.768 0.40 0.78 Table. L-band image pair: the comparison of the accracy with different method algorithm Only relaxation matching Least Sqare Nmber of Points RMS Error Maximm residal Range Azimth Total RMS Range Azimth Total residal 05 0.096 0.9 0.4 0.97 0.639 0.984 98 0.035 0.04 0.055 0.9 0.37 0.37 Table. C-band image pair: the comparison of the accracy with different method 4 CONCLUSIONS An efficiency registration approach for InSAR data processing based on mlti-step matching is presented. The experiment reslt indicates its promising possibility in InSAR data processing. It appears to be a good compromise International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000. 9
between accracy and comptation cost. The procedres were implemented atomatically. Frther investigation will focs on the evalation of the accracy and reliability of the DEM qality generated in the next steps. ACKNOWLEDGMENTS The athors wold like to thank Mr. Jie YANG, Wen YANG and Pan ZHU et al for the technical spport. REFERENCES Gabriel, A. K., R. M. Goldstein, 988. Crossed orbit interferometry: Theory and experimental reslts from SIR-B. Int. J. Remote Sensing, 9(5), pp. 857-87. Garnieri, M. A. and Cladio Prati, 997. SAR Interferometry: A Qick and Dirty Coherence Estimator for Data Browing. IEEE Trans. On Geos.&RS., 35( 3), pp.660-669. Lin, Q., J. F. Vesecky, 99. Registration of Interferometric SAR Images. Proceedings of IGARSS'9, pp. 579-58. Rosenholm, D., 986. Accracy Improvement of Digital for Evalation of Digital Terrain Model, IAPRS, Vol. 6, Comm. III, Rovaniemi, pp.573-587. Trove, E., Nicolas J., and Maitre H., 998. Improving Phase Unwapping Techniqes by the Use of Local Freqency Estimates. IEEE Trans. On Geos.&RS., 36(6), pp.963-97. Wang, Z., 990, Principles of Photogrammetry (with Remote Sensing), Pblishing Hose of Srveying and Mapping, Beijing Zebker, H.S., Goldstein R. M., 986. Topographic Mapping from Interferometric Synthetic Apertre Radar Observations. J. Geophys. Res., Vol.9(B5), pp.4993-4999. Zebker, H. A., Werner C. L., Rosen P. A., Hensley S., 994. Accracy of Topographic Maps Derived from ERS- Interferometric Radar. IEEE Trans. on Geos.&RS., Vol.3, pp.83-835. Zhang, Z. et al, 990. Image by Bridge Mode Based on Dynamic Programming. Int. Arch. Of ISPRS, Vol.8, Part 3/, Whan, China, pp.06-0. Zhang, Z., Zhang J., Liao M., Zhang L., 998. Atomatic registration between SPOT and TM images. Proceedings of SPIE, Vol.3545, pp.358-36 9 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B. Amsterdam 000.