5014

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014

The Epipolarity Constraint in Stereo-Radargrammetric DEM Generation Karlheinz Gutjahr, Roland Perko, Hannes Raggam, and Mathias Schardt

Abstract—For stereometric processing of optical image pairs, the concept of epipolar geometry is widely used. It helps to reduce the complexity of image matching, which can be seen to be the most crucial step within a workflow to generate digital elevation models. In this paper, it is shown that this concept is also applicable to the cocircular geometry of synthetic aperture radar (SAR) image pairs. First, it is proven that, for any feasible SAR acquisition, the deviation from true epipolar geometry is within subpixel range and therefore acceptably small. Based on this, we propose a method to create “epipolar” geometry for arbitrary stereo configurations of any SAR sensor through appropriate geometric image transformations. Consequently, the semiglobal matching (SGM) algorithm can be applied, which is restricted to epipolar geometry and is thus known to be highly efficient. This innovative approach, integrating both epipolar transformation and SGM, has been applied to a TerraSAR-X stereo data set. Its benefit has been demonstrated in a comparative assessment with respect to results, which have been previously achieved on the same test data using state-of-the-art stereometric methods. Index Terms—Digital elevation model (DEM), epipolar geometry, stereo-radargrammetry, synthetic aperture radar (SAR), 3-D mapping.

I. I NTRODUCTION

R

ECENT very high resolution (VHR) spaceborne synthetic aperture radar (SAR) systems like TerraSAR-X or COSMO-SkyMed allow the acquisition of scenes under different off-nadir look angles with a very high level of detail. The usability, potential, and limitations of such multiple off-nadir look angle data sets to extract digital elevation models (DEMs) in a stereo-radargrammetric workflow have already been investigated, e.g., in [1]–[4] or [5]. Within such workflow, image matching is clearly found to be the most challenging step, with its aim to automatically identify corresponding pixels, objects, or features in one (or more) stereo image pairs. This is based on selected matching criteria, which need to be analyzed over an appropriate search area in order to find the best match. The quality of such area-based matching approaches strongly depends on image similarity and uses to increase if the search

Manuscript received April 15, 2013; revised August 28, 2013; accepted September 24, 2013. Date of publication November 11, 2013; date of current version February 27, 2014. These research activities of DIGITAL Institute for Information and Communication Technologies, Joanneum Research, are embedded into a project running within the Austrian Space Applications Programme and are funded by the Austrian Research Promotion Agency (FFG). The authors are with DIGITAL Institute for Information and Communication Technologies, Joanneum Research, 8010 Graz, Austria (e-mail: karlheinz. [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2013.2286409

areas can be kept small. In order to achieve a certain level of similarity, one stereo image is usually warped to the second one using, e.g., an affine transformation. This image registration accounts for offset, rotation, and scale differences between both stereo images and reduces the remaining pixel differences of corresponding image features—so called disparities—in one image direction (usually the line direction). Still, we have to account for disparities in both image directions and the search area for the matching process is a 2-D window. For perspective image data, their epipolar characteristics (see Section II) can be utilized in this concern in order to reduce the matching search area from a 2-D region to a 1-D line through appropriate epipolar rectification [6]. For linear pushbroom images as well as for imaging models based on rational polynomial coefficients (RPCs), the generation of approximated epipolar data sets was already published in recent work [7], [8]. For SAR images, however, only approximate solutions [5] or indirect solutions based on RPCs [8] are described. The authors of [5] do not show the accuracy of their epipolar resampling, whereas empirical values for two TerraSAR-X scenes (without any further specification) and their mixtures with optical scenes are presented in [8]. Neither a detailed investigation of the feasibility of SAR epipolarity, rigorous description analog to optical sensors, nor such development does yet exist. In this paper, a generic approach is presented, which is applicable to generate epipolar images from arbitrary SAR image pairs and thus provides the base to subsequently make use of the semiglobal matching (SGM) method.

II. BACKGROUND ON E PIPOLARITY The epipolar geometry is well defined in the literature. It has been described for perspective images, e.g., in [9]. In the following, a recapitulation of epipolar geometry is given with respect to optical (i.e., colinear) image pairs (see Fig. 1). Each center of projection maps onto a distinct point into the other camera’s image plane. These two image points are denoted by eL and eR and are called epipoles or epipolar points. Both epipoles and both centers of projection lie on a straight 3-D line. The projection of X onto each of the image planes is denoted by XL and XR . If we move X along the left line of sight to Xi , the projection on the left image plane remains, of course, XL . However, the projections of these points onto the right image plane lie on a line (indicated as a red line in Fig. 1). This line is called epipolar line and can also be defined as the

0196-2892 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

GUTJAHR et al.: EPIPOLARITY CONSTRAINT IN STEREO-RADARGRAMMETRIC DEM GENERATION

5015

TABLE I D ETAILED PARAMETERS FOR THE T ERRA SAR-X S POTLIGHT I MAGES OF THE T EST S ITE “B URGAU ”

Fig. 1. Example of epipolar geometry for perspective imagery. Two cameras, with their respective centers of projection points OL and OR , observe a point X. The projection of X onto each of the image planes is denoted by XL and XR . Points eL and eR are the epipoles. Source: WIKIPedia.

intersection of the epipolar plane, which contains the points OL and OR , and X, with the respective image plane. Summarizing, the epipolar geometry is described by two facts [10]. First, the corresponding point of a reference point in the left stereo partner can be found on a straight line in the right stereo image. Second, all these epipolar lines intersect in one point, i.e., the epipole. For optical sensors, rectification algorithms exist which move the epipole at infinity in such a way that all epipolar lines are parallel and, moreover, parallel to the image column direction (see, e.g., [11] or [12]). A major motivation of the epipolar rectification is the reduction of the matching problem from 2-D to 1-D. Aside from the decreased search space (and therefore decreased computing time and decreased probability of mismatches), the main advantage is the potential use of matching strategies originally developed for stereo vision applications [13]. III. E PIPOLAR G EOMETRY FOR SAR I MAGES With respect to SAR stereo data sets, it is first discussed whether sufficiently accurate epipolar geometry is feasible at all. Therefore, a test data set as described hereinafter was used. Then, we extended an epipolar rectification approach, which originally was developed for spaceborne linear pushbroom images, in order to cope with any spaceborne SAR imagery. A. Test Data A set of three TerraSAR-X products, which have been acquired in summer 2009 over an Austrian rural test area at different look angles from ascending orbit, has been used for the testing of developed algorithms and validation of achieved results. These images had been ordered as multilook-groundrange-detected products in dual polarization (HH, VV) with science orbit accuracy. The same image triplet has also been used in previous investigations (see [1] and [14]) devoted to SAR stereo mapping. Acquisition parameters of this image triplet are given in Table I. The images are further denoted by ASC1, ASC2, and ASC3. This test area mainly covers agricultural and forest areas and shows flat to slightly hilly terrain, with the ellipsoidal heights ranging from 270 to 445 m above sea level.

To enable quantitative evaluations, a LiDAR digital surface model (DSM) was used as reference. The LiDAR data were also acquired with a density of 4 points per square meter in 2009 and covers about 67% of the stereo overlap (see Fig. 6). The accuracy of the LiDAR DSM (orthometric heights) was validated by using 21 cadaster points resulting in a height difference of about 0.20 ± 0.14 m. The major part of this difference is due to the gridding of the original LiDAR point cloud. Still, the accuracy is better than the 3-D geolocation accuracy of the used SAR stereo data sets which was found to be ±1.07 m minimum [1]. B. Discussion of SAR Epipolarity The mathematical basics of the cocircular SAR geometry are given through the well-known range and Doppler equations [15] which describe the relation between image coordinates (x, y) or equivalent time and range coordinates (t, r) and the 3-D coordinates P = P (ϕ, λ, h) of a target point λwave fDC   ˙ − S)( ˙ P − S)  =0 |P − S| − (P 2  = 0. r − |P − S|

(1) (2)

The target point is a function of its geographic latitude ϕ, longitude λ, and height h ⎡ ⎤ (N + h) cos ϕ cos λ ⎦. P = ⎣  (N + h) cos ϕ sin (3)  λ (1 − e2 )N + h sin ϕ N and e2 are parameters of the underlying ellipsoid.  P˙ is the first derivative of the target point with time. The 3-D coordinates of the sensor position are a function of the  = S(x)  along-track coordinate S as well as the velocity vector   ˙ ˙ S = S(x). λwave denotes the wavelength of the SAR sensor, and fDC denotes the zero Doppler center frequency. Based on these formulas, we will now discuss two properties of epipolarity which have to be considered [10]: 1) Straightness approximation: The epipolar curve has to be approximated by a straight line. 2) Local conjugacy: All epipolar lines have to be parallel. C. Straightness Approximation For (colinear) perspective images, the fundamental matrix describes the relation between corresponding image coordinates in the left and in the right scene. Such a fundamental matrix cannot be found for SAR geometries, as we have to deal with cocircular range circles instead of colinear lines

5016

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014

Fig. 2. Deviation of the epipolar curve of the left image center point in the right stereo partner from low-order polynomials. Only the height range of the Earth is considered.

of sight. Thus, we used a numerical simulation to study the characteristics of an epipolar curve in SAR images. Without loss of generality, we assumed SAR scenes with zero Doppler center frequency and an Earth-fixed Earth-centered coordinate ˙ in (1) and set up the simplified Doppler system eliminating P and range equation for the left and the right image  ˙ (x ) P (ϕ, λ, h) − S i (xi ) = 0 S (4) i i



i (xi )

= 0. (5) ri (yi ) − P (ϕ, λ, h) − S We chose the scene ASC1 as our left and the scene ASC3 as our right stereo partner. The range circle, which is defined by the scene center of the left scene, was transformed to the right stereo partner for a height interval of −100 to 9000 m above sea level, thus spanning the whole range of terrain heights being feasible on Earth. This is done in two steps, first intersecting a cocentric Earth ellipsoid at predefined height (image-tomap transformation) and then transforming the resulting 3-D coordinates into pixel coordinates of the right scene (map-toimage transformation [16]). Both transformations are nonlinear, and the solution can only be found iteratively by using Taylor series expansions of (4) and (5). These transformations result in a generic curve in the right stereo image, the epipolar curve, whose straightness was to be further investigated. A linear as well as a quadratic polynomial was fitted through this epipolar curve. The remaining residuals are in the range of −2.7 to +1.3 pixels for the linear polynomial, while the fitting error of a quadratic polynomial would be even below ±0.02 pixels (see Fig. 2). A limitation to a height interval of 4000 m—which may be considered to be sufficiently realistic for the coverage of a single SAR stereo pair—reduces the error of a linear fit below approximately ±0.5 pixels. D. Local Conjugacy Second, we investigated the parallelism of the fitted epipolar lines. Therefore, in addition to the scene center point, eight points equally distributed over the left scene were projected to

Fig. 3. Fitted epipolar lines of nine equally distributed points. Only a height range of 0 to 4000 m is considered.

the right stereo partner using the algorithm as described earlier. Again, we limited the fit to a height interval of 0 to 4000 m. Fig. 3 shows the resulting epipolar curves. As the parallelism or convergence of these lines is hardly visible, we fitted a 2-D polynomial to the gradient angle of all epipolar lines. The residual error is about ±2.1e − 6 rad which means that the deviation of a systematic convergence of all epipolar lines is about maximum ±0.03 pixels per 15 000 pixels. Thus, it can be stated that the residual errors of the straightness approximation as well as of the local conjugacy are distinctly less than one pixel for arbitrary SAR image pairs acquired anywhere on the globe. IV. E PIPOLAR R ECTIFICATION OF SAR I MAGES Epipolar rectification now describes the process to transform both stereo images in such a way that the epipoles move to infinity. This has the effect that epipolar lines get parallel and coincide with the column direction, and thus, corresponding points have the same line coordinates. If we assume strictly parallel orbits, zero Doppler processed SAR images already would fulfill the epipolarity constraint, and a simple rotation of both images would achieve epipolar rectified images. However, the platform of spaceborne SAR systems has to follow Kepler’s laws, and images, in addition, may be focused using an arbitrary Doppler center frequency (e.g., Advanced Land Observing Satellite Phased Array LBand SAR). Therefore, we propose a generalized epipolar rectification for SAR images, which assures the achievement of “epipolarity” for arbitrary SAR image pairs at the level of 1–2 pixels within the height range possible on the Earth surface. The quotations marks should indicate that rigorous epipolarity as for (colinear) perspective images in general is not really feasible for spaceborne SAR images. The development is based on an appropriate adaptation of an approach, which was introduced by Wang et al. [17] for linear pushbroom images, to SAR images and comprises two basic processing steps as follows. First, both stereo images are projected onto a common reference plane, which is defined, e.g., via the mean height of the mapped area. For the projection, a standard SAR geocoding procedure is utilized [16], [18]. This step already ensures that

GUTJAHR et al.: EPIPOLARITY CONSTRAINT IN STEREO-RADARGRAMMETRIC DEM GENERATION

5017

TABLE II R ESIDUAL E RROR OF THE A FFINE R EGISTRATION OF THE O RIGINAL AND THE “E PIPOLAR ” S TEREO I MAGE PAIR ASC1 AND ASC3

Fig. 4. Epipolar curve of the marked left point in the right stereo partner in the original ground range image geometry.

Fig. 5. Epipolar curve of the marked left point in the right stereo partner after applying the proposed epipolar rectification.

the epipolar curves can be assumed as parallel lines with deviations below the SAR resolution. In a second step, a Helmert transformation with a fixed scale is applied to both projected images. The rotation angle of this transformation can be calculated as described in [17], and thus, the transformation ensures that the direction of the epipolar lines corresponds to the column direction. The shifts of this transformation ensure zero parallaxes in the line direction and control the parallaxes in the column direction. For the calculation of the shift parameters, we transform the center point of the warped left scene to the warped right scene using a mean height of the area of interest (AOI). Depending on the accuracy of the used SAR image metadata (e.g., preliminary versus science orbits), we recommend to measure one or more tie points (TPs) to calculate the shift parameters. This way, misalignments due to residual errors of the SAR imaging geometry can be eliminated. To avoid double resampling, both processing steps are combined, thus establishing the transformation from a pixel in the epipolar image to the corresponding position in the original image. This can be done on a pixel-by-pixel basis independently from the resampling which is done in a final step. Fig. 4 shows a subset of 300 × 200 pixels of the TerraSAR-X image pair ASC1 and ASC3. A crossroad is marked in the left stereo image, and the corresponding epipolar curve is drawn in the right image. It is evident that the ground sampling distance (GSD) of both stereo images differs, particularly in the range direction, and that the epipolar curve does not coincide with the column direction. For the same crossroad feature, the effect of the proposed epipolar rectification is shown in Fig. 5. Aside from this visual control, a numeric proof of the benefit of the epipolar concept was carried out as follows. Ninety-six TPs were automatically generated for the image pair ASC1 and ASC3, and an affine transformation was fitted into this TP set. Furthermore, these

TPs were also transferred to the geometry of the epipolar image pair. The remaining residual errors of the affine fit in the original stereo case and the remaining offset between the TPs without any additional fitting in the “epipolar” stereo case are summarized in Table II. In both cases, the main topography-induced displacements are accumulated in the column direction (i.e., y-residuals). While, in the traditional affine registration, nonnegligible topographic displacement also remains in the line direction, expressed by x-residuals in the range of about −9 to +5 pixels, the epipolar rectification does not show significant pixel displacements in the line/x-direction anymore. As these residuals give indication on the size of the search area to be specified for matching, it becomes obvious that this can be reduced to a 1-D line in the case of epipolar SAR image pairs. V. DSM G ENERATION R ESULTS To demonstrate the benefit of the proposed epipolar SAR rectification, DSMs were derived from the three TerraSAR-X stereo image pairs based upon three different processing schemes: 1) Affine WTA as in [1], including affine image registration as well as the prediction of the matching location based upon the sensor models and a coarse DEM and a winnertakes-all (WTA) matching decision. 2) “Epipolar” WTA, which is based upon the epipolar rectification of both SAR stereo images as described in the previous section as well as the prediction method and matching decision of 1 but with reduced search window size. 3) “Epipolar” SGM, which is based upon the epipolar rectification of both SAR stereo images as described in the previous section and which then can utilize the SGM method as well as the prediction method of 1. Aside from the individual stereo pairs, also the triplet constellation was used to generate a multi-stereo DSM. The applicability of such image constellation as well as benefits has been demonstrated and discussed in [1], [14], and [19]. A qualitative assessment of the DSMs, which have been generated using these different input scenarios and applying the three approaches listed earlier, was made in comparison to the available LiDAR reference DSM. The key performance indicators of this benchmark are the completeness of the derived DSMs and the accuracy and precision of the retrieved height information with respect to the reference DSM.

5018

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014

A. Processing Details One major benefit of the epipolar rectification is the reduction of the matching problem from 2-D to 1-D. Aside from the decreased search space, the main advantage is the potential use of matching strategies originally developed for stereo vision applications. A good overview and benchmark of such algorithms is given on the Stereo Vision Research page of the Middlebury College as described in [13]. We chose the SGM approach [20] which in the field of remote sensing until now was only used for VHR optical airborne data [21] or VHR optical spaceborne data [22]. We furthermore incorporated dedicated improvements, like the combination of different normalized cross-correlation kernel sizes as proposed by Meric et al. [23] and the prediction of image disparities using the physical SAR sensor model and a coarse DEM, as proposed by Raggam et al. [1]. In order to preserve small details as well as to reduce the effect of speckle noise, the mean of three normalized cross correlation functions with kernel sizes of 3 × 3, 7 × 7, and 15 × 15 pixels was used for the computation of the matching costs. The benefit of the usage of several cross correlation kernels was discussed previously in [23]. Due to the size of the search window which is on the order of about ±110 pixels (see Table II), we applied a hierarchical matching approach. This offers the possibility to reject unreliable matching results by applying filter and interpolation techniques before stepping to a lower pyramid level. Internal quality checks are based on the actual matching cost and the so-called back matching distance which is found via forth and backward matching. In all three processing lines listed earlier, the same parameters were used for the matching purposes (see Table III). Thus, any differences in the output DSMs can be assumed to be due to the following. 1) The preparation, i.e., the registration of the input data, which further has an impact on the size and dimensionality of the matching search window. 2) The matching approach being used to find the best match. The areal matcher in the first two processing schemes uses the WTA decision on a pixel-by-pixel base, whereas the semiglobal matcher tries to find a “semiglobal” best solution by penalizing strong disparity discontinuities. All DSMs were calculated with a grid spacing of 2 m using a “standard” forward intersection for SAR images based on linearized range/Doppler equations. Thus, the output of the “epipolar” matching approaches was first transformed back to the original stereo image geometry using the inverse address calculation as used in the “epipolar” rectification step. The resulting multi-stereo DSMs of all three approaches are shown in Fig. 7. B. Assessment of Completeness First, the DSMs resulting from the three processing schemes were evaluated with respect to the completeness of the produced DSMs. The matching step as well as the subsequent forward intersection to determine the ground coordinates al-

TABLE III S UMMARY OF M AIN M ATCHING PARAMETER

TABLE IV C OMPLETENESS OF THE S TEREOMETRIC D ERIVED DSMs C ALCULATED AS P ERCENTAGE OF VALID DSM P IXELS

lows to a certain extent the automatic detection of blunder. Pixels belonging to blunder are then marked as invalid pixels and are shown in blue color in Fig. 7. Table IV summarizes the percentage of DSM pixels which passed the automatic blunder test with respect to the total number of pixels of the final DSM. Best values are indicated as bold numbers. With respect to completeness, the affine WTA and the “epipolar” SGM approach show very similar results. The “epipolar” WTA approach shows significant less completeness for those two stereo pairs which include scenes with different GSDs (pairs ASC1-ASC2 and ASC1-ASC3). As expected, the usage of all stereo configurations achieves the best completeness at already a level of about 99% of reliable surface information.

GUTJAHR et al.: EPIPOLARITY CONSTRAINT IN STEREO-RADARGRAMMETRIC DEM GENERATION

5019

Fig. 6. Reference and input data for the used test site. In the LiDAR DSM, the AOI for the quantitative analysis is shown in red, and zoom of Figs. 7 and 8 is shown as dashed yellow box.

Fig. 7.

Surface models derived from combining the three individual stereo pairs (unreliable matching areas rejected and shown in blue color).

C. Assessment of Accuracy and Precision Second, an assessment of the accuracy and precision of the three DSM output products was made, using the reference LiDAR DSM, which was shown to be more accurate than the DSMs derived from SAR stereo processing. It has to be emphasized that effects of SAR layover and shadow are present along the borders of forest stands and clearcuts. For such areas, a reliable surface reconstruction is hardly feasible. Moreover, one has to cope with unpredictable SAR penetration effects over forest areas, as investigated in [14]. Therefore, extended forest areas were excluded in order to reliably assess the DSM accuracy achieved from the various processing schemes. For the qualitative analysis, an AOI, which covers mainly rural areas and only insular tree stands, was defined. This is shown in a superposition to the LiDAR DSM in Fig. 6.

The SAR stereo DSMs were calculated in an ellipsoidal height system and thus had to be shifted by the local geoidal height. We used the globally available Earth gravitational model EGM96 [24] given as undulation to the WGS84 datum. The standard deviation of the difference of the EGM96 to the Austrian orthometric height system is about 0.02 to 0.06 m [25] which is negligible in our context. Fig. 8 assembles the differences of the multi-stereo DSMs with respect to the reference LiDAR DSM. The difference DEM achieved from the classical stereo approach shows extended bright and dark areas, indicating larger elevation differences, which are due to the propagation of initial matching areas occurring at the top level of the hierarchical matching procedure. Within the specified AOI, a decrease of noise can be observed from the affine WTA toward the “epipolar” SGM

5020

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014

Fig. 8. Difference height models with respect to LiDAR reference (unreliable matching areas rejected and shown in blue color). TABLE V ACCURACY OF THE S TEREOMETRIC D ERIVED DSMs E XPRESSED AS M EAN AND STD DEV OF THE D IFFERENCES TO THE R EFERENCE LIDAR DSM W ITHIN THE S PECIFIED AOI

approach. The “epipolar” SGM result shows the smoothest behavior, which is in total agreement with the expectations to the SGM. Statistical values of the height differences between the stereo-derived DSMs and the LiDAR DSM were calculated for the specified AOI. Among these, the standard deviation is considered to give indication on the precision, and the mean value is considered to give indication on the accuracy of the stereoderived DSMs. Table V summarizes both values for the three processing schemes as well as the various stereo combinations that have been used. The best values achieved for the individual image combinations are indicated as bold numbers. Concerning precision, the standard deviation values achieved from all three approaches do not show significant differences with slightly better results for the “epipolar” SGM approach which is again in total agreement with the expectation to the SGM. The accuracies of the “epipolar” WTA DSMs show a systematic behavior, expressed by mean height errors of about

−1.5 m. Furthermore, the reduction of the matching search window from 2-D to 1-D does not increase the accuracy. This behavior is not as expected and may be explained by the fact that the epipolar rectification introduces resampling artifacts to both stereo images. On the other hand, “epipolar” SGM significantly outperforms affine WTA and “epipolar” WTA in terms of accuracy. A possible explanation to that is the better outlier detection inherent to the SGM. Affine WTA and “epipolar” SGM results show similar accuracy numbers except the ASC2–ASC3 stereo configuration. A weak geometric stereo disposition, in general, is inherent to this image pair due to the small stereo intersection angle. Aside from the careful interpretation of the magnitude of the numbers due to the issues described earlier, the accuracy of the “epipolar” SGM DSMs is found to be about 0.5 m better than that for the affine WTA. This would indicate an accuracy gain of the new approach of about 41%. VI. D ISCUSSION AND S UMMARY In this paper, it was demonstrated that the epipolarity constraint is also applicable to SAR geometry, thus enabling to generate epipolar rectified SAR stereo image pairs. This novel approach limits stereo matching from 2-D to 1-D, and highly sophisticated matching algorithms can be further employed. Aside from the theoretical explanations, the epipolar rectification method was integrated into the stereo-radargrammetric DEM generation workflow and further applied to a TerraSAR-X image triplet. The resulting DSMs as well as traditional derived DSMs were analyzed in terms of completeness and accuracy. The “epipolar” WTA approach does not perform better than the affine WTA approach, but both completeness and accuracy confirm the visual impression that the “epipolar” SGM DSM is more consistent, and thus, the accuracy numbers are more reliable.

GUTJAHR et al.: EPIPOLARITY CONSTRAINT IN STEREO-RADARGRAMMETRIC DEM GENERATION

Future algorithmic improvements will be devoted to the development of customer-tailored cost functions for the SGM method in combination with appropriate SAR despeckling procedures.

[22]

ACKNOWLEDGMENT The authors would like to thank the German Aerospace Center (DLR) and Infoterra Germany GmbH for providing the TerraSAR-X image data and the anonymous reviewers for their valuable comments and suggestions.

[23] [24]

R EFERENCES [1] H. Raggam, K. Gutjahr, R. Perko, and M. Schardt, “Assessment of the stereo-radargrammetric mapping potential of TerraSAR-X multibeam spotlight data,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 2, pp. 971– 977, Feb. 2010. [2] C. G. de Oliveira, W. R. Paradella, and A. de Queiroz da Silva, “Assessment of radargrammetric DSMs from TerraSAR-X stripmap images in a mountainous relief area of the amazon region,” ISPRS J. Photogramm. Remote Sens., vol. 66, no. 1, pp. 67–72, Jan. 2011. [3] T. Toutin, “Impact of Radarsat-2 SAR ultrafine-mode parameters on stereo-radargrammetric DEMs,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 10, pp. 3816–3823, Oct. 2010. [4] P. Capaldo, M. Crespi, F. Fratarcangeli, A. Nascetti, and F. Pieralice, “High-resolution SAR radargrammetry: A first application with COSMOSkyMed spotlight imagery,” IEEE Geosci. Remote Sens. Lett., vol. 8, no. 6, pp. 1100–1104, Nov. 2011. [5] Y. Belgued, E. Hervet, C. Lemaréchal, L. Rognant, P. Marthon, and F. Adragna, “An accurate radargrammetric chain for DEM generation,” in Proc. SAR Workshop, CEOS, vol. 450, ser. ESA Special Publication, R. A. Harris and L. Ouwehand, Eds., Mar. 2000, pp. 167–172. [6] W. Cho, T. Schenk, and M. Madani, “Resampling digital imagery to epipolar geometry,” in Proc. IAPRS, 1992, vol. 29, no. B3, pp. 404–408. [7] T. Kim, “A study on the epipolarity of linear pushbroom images,” Photogramm. Eng. Remote Sens., vol. 66, no. 8, pp. 961–699, Aug. 2000. [8] H. Pan, G. Zhang, and T. Chen, “A general method of generating satellite epipolar images based on RPC model,” in Proc. IEEE Int. Geosci. Remote Sens. Symp., Vancouver, BC, Canada, Jul. 2011, pp. 3015–3018. [9] Z. Zhang, “Determining the epipolar geometry and its uncertainty: A review,” Int. J. Comput. Vis., vol. 27, no. 2, pp. 161–195, Mar. 1998. [10] A. F. Habib, M. F. Morgan, S. Jeong, and K.-O. Kim, “Epipolar geometry of line cameras moving with constant velocity and attitude,” Electron. Telecommun. Res. Inst. J., vol. 27, no. 2, pp. 172–180, Apr. 2005. [11] A. Fusiello and L. Irsara, “Quasi-Euclidean uncalibrated epipolar rectification,” in Proc. 19th Int. Conf. Pattern Recognit., Tampa, FL, USA, Dec. 8–11, 2008, pp. 1–4. [12] A. Fusiello, E. Trucco, and A. Verri, “A compact algorithm for rectification of stereo pairs,” Mach. Vis. Appl., vol. 12, no. 1, pp. 16–22, Jul. 2000. [13] D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense twoframe stereo correspondence algorithms,” Int. J. Comput. Vis., vol. 47, no. 1–3, pp. 7–42, Apr. 2002. [14] R. Perko, H. Raggam, J. Deutscher, K. Gutjahr, and M. Schardt, “Forest assessment using high resolution SAR data in X-band,” Remote Sens., vol. 3, no. 4, pp. 792–815, Apr. 2011. [15] J. C. Curlander, “Location of spaceborne SAR imagery,” IEEE Trans. Geosci. Remote Sens., vol. GE-20, no. 3, pp. 359–364, Jul. 1982. [16] H. Raggam, “An efficient object space algorithm for spaceborne SAR image geocoding,” in Proc. 16th Int. Congr. ISPRS, Jul. 1–10, 1988, vol. 27, pp. 393–400. [17] M. Wang, F. Hu, and J. Li, “Epipolar resampling of linear pushbroom satellite imagery by a new epipolarity model,” ISPRS J. Photogramm. Remote Sens., vol. 66, no. 3, pp. 347–355, May 2011. [18] H. Raggam, D. Strobl, and W. Hummelbrunner, “Product quality enhancement and quality evaluation,” in SAR Geocoding: Data and Systems. Kalsruhe, Germany: Wichmann, 1993, ch. 8, pp. 187–206. [19] H. Raggam, “Accuracy analysis and surface mapping using SPOT-5 stereo data,” in Proc. ISPRS, 2004, pp. 1–7. [20] H. Hirschmüller, “Accurate and efficient stereo processing by semi-global matching and mutual information,” in Proc. IEEE Comput. Soc. Conf. CVPR, San Diego, CA, USA, Jun. 20–26, 2005, vol. 2, pp. 807–814. [21] H. Hirschmüller, F. Scholten, and G. Hirzinger, “Stereo vision based reconstruction of huge urban areas from an airborne pushbroom camera

[25]

5021

(HRSC),” in Pattern Recognition, vol. 3663, Lecture Notes in Computer Science, W. G. Kropatsch, R. Sablatnig, and A. Hanbury, Eds. Berlin, Germany: Springer-Verlag, 2005, pp. 58–66. T. Krauß, P. Reinartz, and U. Stilla, “Extracting orthogonal building objects in urban areas from high resolution stereo satellite image pairs,” in Proc. PIA, U. Stilla, H. Mayer, M. Schmitt, B. Jutzi, and F. Rottensteiner, Eds., Munich, Germany, Sep. 19–21, 2007, vol. 36, no. 3/W49B, pp. 77–82. S. Meric, F. Fayard, and E. Pottier, “A multiwindow approach for radargrammetric improvements,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 10, pp. 3803–3810, Oct. 2011. F. G. Lemoine, S. C. Kenyon, J. K. Factor, R. G. Trimmer, N. K. Pavlis, D. S. Chinn, C. M. Cox, S. M. Klosko, S. B. Luthcke, M. H. Torrence, Y. M. Wang, R. G. Williamson, E. C. Pavlis, R. H. Rapp, and T. R. Olson, “The development of the joint NASA GSFC and the National Imagery Mapping Agency (NIMA) geopotential model EGM96,” NASA GSFC, Greenbelt, MD, USA, Tech. Rep., 1998. C. Briese, N. Höggerl, E. Imrek, J. Otter, and D. Ruess, “Transformation von gnss-höhen in österreichische gebrauchshöhen mittels einer transformationsfläche (höhen-grid),” Österreichische Zeitschrift Vermessung Geoinf., vol. 4, pp. 267–281, 2011.

Karlheinz Gutjahr received the M.Sc. and Ph.D. degrees in geodesy from the Graz University of Technology, Graz, Austria, in 1997 and 2002, respectively. He is currently a Senior Scientist with DIGITAL Institute for Information and Communication Technologies, Joanneum Research, Graz. He is interested in remote sensing with emphasis on geodetic engineering, geometric processing of remote sensing data including (differential) synthetic aperture radar (SAR) interferometry, stereo modeling, and digital elevation model (DEM) generation, as well as the analysis of multisensor remote sensing data. He was involved into algorithm and software development in the frame of the German processing and archiving facility for the shuttle radar topographic mission X-band and the environmental satellite sensors and is currently contributing to the Infoterra TerraSAR-X value-added processors. He has contributed comprehensive algorithmic development for the Remote Sensing Software Package Graz, particularly concerning SAR interferometry and polarimetry, sensor modeling, numerical methods, and adjustment procedures.

Roland Perko received the M.Sc. degree in mathematics and the Ph.D. degree in computer science from the Graz University of Technology, Graz, Austria, in 2001 and 2004, respectively. From 2006 to 2008, he was a Postdoctoral Researcher with the University of Ljubljana, Ljubljana, Slovenia, where he researched on methodologies for integrating visual context with object detection. He is currently a Senior Researcher with DIGITAL Institute for Information and Communication Technologies, Joanneum Research, Graz. His main research interests are cognitive computer vision and geometrical image processing with focus on remote sensing data.

5022

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 52, NO. 8, AUGUST 2014

Hannes Raggam received the M.Sc. degree in geodesy and the Ph.D. degree devoted to stereoradargrammetry from the Graz University of Technology, Graz, Austria. He has been with DIGITAL Institute for Information and Communication Technologies, Joanneum Research, Graz, since 1984, where he is leading a working group on geometric processing issues of remote sensing imagery. His working spectrum includes geometric sensor modeling, stereo data processing, coregistration, and 2-D and 3-D mapping, as well as the analysis of multisensor remote sensing data. He has contributed comprehensive algorithmic and software developments dealing with airborne as well as spaceborne and optical as well as synthetic aperture radar (SAR) imagery, with a focus on the implementation of customized processing chains with respect to automated data value adding. He was involved in the realization of such developments for data processing centers like European Space Agency (Envisat), German Aerospace Center (DLR) (European Remote Sensing Satellites), or Infoterra Germany (TerraSAR-X). He acted as Principal Investigator in several remote sensing missions, like Japanese Earth Resources Satellite, ERS tandem, Envisat, Spaceborne Imaging Radar-C-band, Advanced Earth Observing Satellite, Advanced Spaceborne Thermal Emission and Reflection Radiometer, SRTM, ALOS, Radarsat-1 and 2, and TerraSAR-X.

Mathias Schardt received the M.S. and Ph.D. degrees in forest science from the Albert Ludwigs University of Freiburg, Freiburg, Germany. He finished the habilitation in the field of landscape planning, remote sensing, and geographic information system at the Technical University of Berlin (TU Berlin), Berlin, Germany, in 1997. He has more than 23 years of experience in remote sensing, GIS, and digital image processing. From 1986 to 1990, he was a Junior Scientist with the German Aerospace Research Establishment (DLR), Oberpfaffenhofen, Germany. From 1990 to 1995, he was a Senior Scientist with the Institute of Landscape Planning and Landscape Management, TU Berlin. Since 1995, he has been the Head of Remote Sensing and Geoinformation with DIGITAL Institute for Information and Communication Technologies, Joanneum Research, Graz, Austria, and since 2002, he has been as well a Professor for remote sensing and photogrammetry with the Graz University of Technology, Graz. Dr. Schardt has been a member of the European Space Science Committee since 2009. He is currently the Coordinator of the International Union of Forest Research Organizations Working Group on Remote Sensing and GIS and also the Austrian representative for European Association of Remote Sensing Laboratories and International Society for Photogrammetry and Remote Sensing WG-VII.