Imagin[e,g] Europe I. Manakos and C. Kalaitzidis (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-494-8-371
371
Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X Hannes Raggam1, Roland Perko and Karlheinz Gutjahr Institute of Digital Image Processing, Joanneum Research, Graz, Austria Abstract. — The first German space mission TerraSAR-X was launched on June 15th, 2007. With respect to topographic mapping, this mission has initiated a new generation of high-resolution spaceborne SAR sensors, as it provides SAR images at very high resolution down to 1 meter in the Spotlight mode on the one hand, and variable off-nadir looking angles on the other hand. Thus, high resolution SAR stereo image pairs can be acquired at optimized imaging conditions, stimulating the traditional stereo mapping approach to be used for 3D mapping, either in parallel with or as an alternative to SAR interferometry as the evolving mapping technique of the last decade. In this paper, the stereoradargrammetric potential of TerraSAR-X is investigated based on multiple Spotlight data sets for selected Austrian test sites. First, a-priori estimates of 3D mapping accuracy being feasible from TerraSAR-X stereo pairs are deduced based on control point measurements. Therefore, various stereo dispositions were analyzed. Second, surface models were generated by means of stereo mapping procedures and qualitatively validated by visual comparison to appropriate reference data. Keywords. TerraSAR-X, 3D mapping, stereo-radargrammetry, digital surface model, height model.
Introduction The German TerraSAR-X satellite operates a versatile, new generation highresolution SAR sensor in X-band. SAR image products can be acquired in various imaging modes, like the ScanSAR, the Stripmap or the Spotlight mode, the latter providing SAR images at very high resolution in the range of about 1 meter [1]. Besides, images can be acquired at different off-nadir viewing angles, or imaging beams. In mid-European areas, for instance, three coverages acquired at different incidence angles are typically feasible in the full-performance incidence angle range of the high-resolution Spotlight mode. Hence, pairs of TerraSAR-X images being acquired at different viewing angles may be used in a stereo processing chain in order to extract 3D information or to generate 3D surface models. Thereby, the processing may benefit from optimized geometric conditions, like high resolution as well as multiple and therefore selectable viewing angles. 3D mapping accuracies as never achieved before from spaceborne SAR stereo images may therefore be anticipated. Prior to TerraSAR-X, the Canadian Radarsat as well as to the European Envisat-ASAR sensor were capable to acquire images at 1
Corresponding Author; E-mail:
[email protected].
372 H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X
varying look angles in a long-term context, however at distinctly worse pixel resolution. Related work was published e.g. by Toutin [10] or Schubert et al. [9]. Also at the Institute of Digital Image Processing of Joanneum Research (JR-DIB) stereo mapping experiments were carried out for Radarsat image pairs [5]. At JR-DIB investigations were launched to exploit the potential of TerraSAR-X for 3D mapping and information extraction. For this purpose, “multi-beam” TerraSARX data sets comprising three Spotlight products acquired at different off-nadir viewing angles from either ascending or descending orbit have been collected for various Austrian test sites, either in the MGD or the SSC format. The objectives of these investigations were 1. 2. 3. 4.
to analyze the 3D mapping accuracy being feasible from TerraSAR-X stereo or multiple image data sets, to generate digital surface models, to qualitatively and quantitatively validate the surface models, and to investigate, whether forest stands can be mapped from TerraSAR-X stereo or multiple image data sets.
First results and experiences gained with respect to geo-location accuracy assessment and stereo-data processing have yet been published in [8]. The present paper focuses on results and experience gained with respect to the first two items, where a dedicated focus is put onto the utilization of multiple image data sets. The issues 3 and 4 listed above require the availability of appropriately accurate reference data sets, like Lidar ground and surface models. Such reference data still have to be acquired or are still being produced for the selected test sites. In the present context the following test sites and TerraSAR-X test data are used: Test site “Burgau”: This rural area shows flat to slightly hilly terrain, the ellipsoidal heights ranging from 270 to 440 meters a.s.l., and represents agricultural as well as various forest types. Here, three single-look slant-range complex (SSC) Spotlight products acquired at different incidence angles from ascending orbit (science orbit accuracy) have been used with ~1.5m GSD. Test site “Baernbach”: This test site represents a mountainous area, with terrain heights ranging from 515 to 1100 meters a.s.l. in the kernel investigation area. It covers rural regions with extended forest coverage, which on the other hand was severely damaged January 2008 by a thunderstorm. For this test site, six MGD Spotlight products were acquired, three each from ascending and descending orbit with a nominal GSD of 0.75m. One objective was, to reduce the problems induced by SAR foreshortening, layover and shadow, as they typically are present in mountainous terrain, by a joint utilization of ascending and descending products.
1. Assessment of 2D Geo-location and 3D Stereo Mapping Accuracy In order to make an a-priori assessment of the 2D geo-location accuracy as well as of the 3D stereo mapping accuracy, ground control points (GCPs) were manually measured in the stereo images. For both test sites, existing ortho-photo mosaics with a pixel resolution of 1 meter and available DEMs were used to extract the planimetric ground coordinates and the heights of the GCPs, respectively. Based on these reference
H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X 373
data, natural targets, like road intersections, water bodies, and the like, had to be used throughout. Thus the GCP measurement accuracy suffers from the following facts: 1.
2.
Reference data like ortho-photo maps cannot be considered to be absolutely accurate. In fact it can be assumed that they are less accurate than the TerraSAR-X initial geometry which is reported to yield a pixel location accuracy of less than 1m (see [2]). Natural targets frequently are difficult to be selected and to be located with sufficient precision in the SAR image and the reference ortho-photo mosaic. This is shown for an exemplary GCP in Figure 1, representing a road intersection. Here, the location uncertainty of the GCP in the SAR image with respect to the location specified in the reference ortho-image can be estimated to be easily in an order of 1-2 pixels or even more.
Figure 1. Road intersection selected for GCP measurement (SAR image left; optical image right).
Point residuals between measured image coordinates and those resulting from backward point transformation were calculated in azimuth and range for the individual images. RMS and mean values of these residuals are given in Table 1. As can be seen from the mean values, a distinct bias is inherent to these residuals, which may be due to a displacement of the reference data (see “Burgau” range accuracy) but also to inaccurate sensor models (see “Baernbach” azimuth and range accuracy). According to the accuracy statements made by DLR, such effects should not be present and therefore would need to be further investigated. Table 1. 2D geo-location accuracy.
Baernbach
Burgau
Test site
RMS initial [pxl]
Mean initial [pxl]
RMS adjusted [pxl]
Image / Look angle / Date
GCPs
Az
Rg
Az
Rg
Az
Rg
ASC1 / 22.17° / 16-06-2008
22
0.94
1.38
-0.17
-0.86
0.75
0.82
ASC2 / 37.16° / 21-06-2008
22
0.85
1.79
-0.16
-1.56
0.68
0.68
ASC3 / 48.48° / 15-06-2008
22
1.08
3.03
-0.17
-0.86
0.85
0.64
ASC1 / 30.54° / 09-10-2008
34
1.51
2.20
-0.56
1.27
0.98
1.28
ASC2 / 43.54° / 03-10-2008
46
2.15
2.08
-0.08
0.78
1.01
1.09
ASC3 / 53.23° / 08-08-2008
46
1.71
1.90
-0.77
-1.08
0.78
0.80
DSC1 / 28.91° / 06-10-2008
26
1.62
6.58
-1.30
6.20
0.93
0.83
DSC2 / 42.30° / 12-10-2008
32
0.87
5.64
0.34
5.49
0.67
1.04
DSC3 / 52.30° / 07-10-2008
36
1.15
5.12
0.57
4.97
0.97
0.93
374 H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X
For our purpose, least squares parameter adjustment was applied in order to optimize the sensor models and make them consistent with the reference data. Thereby, azimuth and range geometry of the sensor models were adjusted. Then “adjusted” RMS point residuals are achieved. These indicate a geo-location accuracy - or rather geolocation correspondence between reference data and TerraSAR-X data - in the order of 1 pixel or even less. In order to determine a-priori estimates for the 3D mapping accuracy being feasible from TerraSAR-X stereo pairs, the stereo image points measurements were first used to calculate 3D ground coordinates of the GCPs. These are achieved by a socalled least squares spatial point intersection. 3D point residuals can then be determined in comparison to the manual ground coordinate measurements. In this concern, individual stereo pairs as well as image triplets were analyzed. The RMS values of the 3D point residuals are summarized in Table 2. From the RMS values given in the Table, the following conclusions can be drawn: 1. In general, planimetric as well as height accuracy in the order of 2 meters and less may be achieved if appropriate stereo or multi-beam data sets are used; 2. The accuracy significantly improves with increasing stereo intersection angles. It is therefore best for those stereo pairs, which are constituted by images 1 and 3, and obviously for the image triplet data composite; 3. Stereo pairs constituted by images 1 and 3, i.e. ASC1/ASC3 or DSC1/DSC3, and image triplets yield adequate accuracies as they span the same intersection angle range Thus, the 3D stereo mapping accuracy assessment indicates that pretty accurate surface models – or 3D information in general – could be extracted from Spotlight TerraSAR-X stereo images or multi-beam data sets, promising potential accuracies in the order of 2 meters in planimetry and height. Table 2. 3D geo-location accuracy.
Baernbach
Burgau
Test site
Model
Stereo angle
RMS [m]
GCPs East
North
Height Length
Rejected pixels (%)
ASC1 – ASC2
15.0°
22
1.78
1.65
1.33
2.77
34.6
ASC2 – ASC3
11.3°
22
3.08
1.92
2.98
4.70
19.7
ASC1 – ASC3
26.3°
22
1.42
1.71
1.30
2.57
38.2
ASC triplet
15.0° + 11.3°
22
1.31
1.61
1.33
2.47
9.6
ASC1 – ASC2
13.0°
34
1.98
0.70
1.38
2.34
47.5
ASC2 – ASC3
9.7°
46
2.37
0.62
2.79
3.71
30.3
ASC1 – ASC3
22.7°
34
1.13
0.47
1.03
1.61
49.4
ASC triplet
13.0° + 9.7°
44
1.33
0.51
1.30
1.93
16.2
DSC1 – DSC2
13.4°
25
1.50
0.48
0.99
1.86
56.1
DSC2 – DSC3
10.0°
32
2.01
0.64
2.23
3.07
31.7
DSC1 – DSC3
23.4°
26
1.18
0.60
0.83
1.56
60.9
DSC triplet
13.4° + 10.0°
31
1.41
0.46
1.39
2.03
20.3
H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X 375
2. Surface Mapping Procedures 2.1. Automatic Stereo Processing Workflow A fully automatic workflow can be applied to TerraSAR-X stereo pairs in order to generate a DSM. Typically, the following processing steps need to be executed: x
Data pre-processing to reduce the impact of speckle.
x
Coarse registration of one image (the search image) with respect to the other (the reference image).
x
Image matching in order to find correspondences of reference image pixels in the search image. A very common matching criterion is the cross correlation coefficient, which was also used in this context in an extended manner. The success of image matching depends to the major degree on the similarity of the image pair to be matched. The underlying algorithm is based on a hierarchical processing using a multi-resolution image pyramid (cf. [3, 4]) and comprises the following processing steps: (1) prediction based on sensor models and coarse elevation model (e.g. employing the SRTM model); (2) refinement using normalized cross-correlation followed by sub-pixel interpolation; (3) rejection of outliers; (4) interpolation of undefined regions and (5) propagation of the disparities to the next pyramid level.
x
Spatial point intersection, i.e. least squares approach to find the intersection point of SAR range circles as defined via the corresponding image pixels delivered by image matching. To some extent, unreliable matching results can be figured out and rejected, as they usually yield range circles which are significantly displaced with respect to each other. This procedure results in a “cloud” of 3D points, irregularly distributed on ground.
x
DSM re-gridding, i.e. interpolation of a regular raster of height values from these 3D points. Remaining gaps may be filled using an appropriate interpolation mechanism.
2.2. Utilization of Multiple Matching Results Appropriate procedures have been developed at JR-DIB, which can utilize multiple image matching results as they may be generated in case of multi-beam coverage of an area [6, 7]. The benefit of this approach is given by the fact that image pairs with best similarity can be matched, assuring optimum matching performance, while the multiple matches can then be merged in order to assure high geometric robustness for 3D mapping. These procedures have been applied in an adequate manner to the image triplets which have been treated in this investigation. 2.3. Merging Results from Ascending and Descending Orbit Over mountainous terrain the information content of the SAR images is degraded by the effects of SAR foreshortening and layover at foreslopes facing the sensor or SAR shadow at backslopes being tilted away from the sensor. For such affected areas image
376 H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X
matching becomes frequently unreliable or even impossible, yielding either erroneous or empty areas in an extracted surface model. Imagery from ascending as well as descending orbits may then be used to partly overcome these problems. An appropriate fusion of DSMs, which are generated from images acquired from ascending and from descending orbits, respectively, may be applied, aiming at a replacement of unreliable or empty areas of one of these two DSMs by information of the other DSM. It is to be noted, that in case of slopes being steeper than the SAR off-nadir viewing angles even this approach may not help to remove the problematic areas. For instance vertical structures like forest borders or building facades are in layover with respect to one illumination direction, while they are in SAR shadow with respect to the other. Therefore the extraction of reliable 3D information in a certain vicinity of such structures is not possible.
3. Surface Mapping Results The surface mapping procedure as sketched above was applied to individual stereo pairs as well as to the image triplets of the two selected test sites, i.e. to all the dispositions included in Table 2. Appropriate residual thresholds were used during spatial point intersection in order to identify and eliminate unreliable matching results. The percentage of points which are consequently rejected is given in Table 2. It is obvious, that x less points are rejected for the dispositions of the flat test site “Burgau” in comparison to the dispositions of the mountainous test site “Baernbach”, x the rejection rates correlate with the size of the stereo intersection angles, i.e. less points are rejected in case of smaller angles and vice versa, and x the rejection rates can be significantly reduced when dealing with image triplets and joint utilization of multiple matching results. For the test site “Burgau”, the DSMs resulting from the individual stereo pairs as well as from the image triplet are shown in Figure 2. The stereo-derived DSMs still include the gaps which are caused by rejection of unreliable matching results, and which in their amount more or less correspond to the rejection rates of Table 2. These gaps have been filled by appropriate interpolation in the DSM derived from the image triplet. There, the amount of the gaps is drastically reduced (9.6% versus 19.7% or more) and the comparably small gaps of this DSM can be filled by interpolation with sufficient confidence. For visual validation, height differences were calculated with respect to a reference (ground) elevation model. Ideally, these should be equal or close to zero in the open areas and include reasonable heights of vegetation, forests and buildings on the other hand. A colour-coded presentation of the height (difference) model resulting from the DSM extracted from the image triplet is shown in Figure 3 in comparison to an orthophoto mosaic. As indicated by the areas shown in reddish colour, forest stands are mapped quite reasonably with respect to shape and height values. However, borders are degraded due to the layover/shadow effects discussed previously.
H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X 377
ASC1-ASC2 stereo
ASC2-ASC3 stereo
ASC1-ASC3 stereo
Image triplet, interpolated
Figure 2. Surface models derived from individual TerraSAR-X stereo pairs (unreliable matching areas rejected) as well as from image triplet for test site “Burgau”.
Figure 3. Color coded height model (left) and ortho photo mosaic (above).
378 H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X
In an adequate manner, DSMs were generated for the test site “Baernbach”, including stereo scenarios and image triplet of ascending as well as descending data sets. Figure 4 illustrates the DSMs derived from any of the ascending stereo pairs as well as the DSMs generated from the ascending and the descending image triplet, respectively, all of them including gaps due to rejection of unreliable matching results. Finally, the Figure shows a DSM generated by a fusion of the “ascending” and the “descending” DSM, for which filling of the gaps has been applied. First of all these illustrations clearly show, that foreslopes represent critical areas for image matching in the case of mountainous terrain. Gaps resulting from unreliable matching results are mainly located in foreslope regions. The amount of the gaps increases with the steepness of the off-nadir viewing angles on the one hand, as foreshortening and layover effects are more severe in case of steep look angles, and with the size of the stereo intersection angle on the other hand. In the DSMs derived from the image triplets the amount of gaps is distinctly reduced (see also rejection rates of Table 2). Nevertheless, the slopes facing the SAR sensor are still severely lacking of height information. Thus, merging of DSMs derived from ascending and descending data sets helps to get rid of these problems to a major extent as shown in the right bottom illustration of Figure 4.
ASC1-ASC2 stereo
Ascending image triplet
ASC2-ASC3 stereo
Descending image triplet
ASC1-ASC3 stereo
ASC & DSC merged & filled
Figure 4. Surface models derived from individual TerraSAR-X ascending stereo pairs, from image triplets, as well as from the “ascending/descending merge” for test site “Baernbach”.
H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X 379
Again, height differences with respect to a reference (ground) elevation model were calculated for visual validation. This is shown in Figure 5 along with the corresponding ortho-photo map. Visual comparison reveals adequate experiences like for the “Burgau” test site: forest areas, which are shown in reddish colour are in general reasonably mapped. In addition to the problematic forest border areas, however, inconsistencies are detectable in certain topography related areas, like mountain ridges. These could be due to image matching and the filtering procedures being included or to inappropriate quality of the reference ground model, and requires more detailed analysis, preferably involving higher quality reference data sets like Lidar DEMs.
Figure 5. Color coded height model (left) and ortho photo mosaic (above).
4. Conclusion and Outlook Using multi-beam image data sets for two selected Austrian test sites, the 3D mapping potential of TerraSAR-X was investigated. The a-priori accuracy assessment has shown that a fairly high accuracy of 2 meters and less in general may be achieved if stereo measurements are performed with appropriately high accuracy. A major difficulty in the context of accuracy assessment arises from the lack of sufficiently accurate reference data, as commonly used 2D ortho-photo maps or 3D reference DEMs may not fulfil respective quality requirements. In interaction with the topography, ortho-photos for instance may have displacements which are larger than the estimated mapping accuracy of TerraSAR-X. Surface models were further generated from TerraSAR-X stereo images as well as from multi-beam data sets involving image triplets. Optimum results in general are achieved from image triplet usage, where multiple matching results are jointly utilized. The validation of the DSMs turned out to be difficult, as the selected test sites are covered by forest to a high degree, and information on the true crown surface was not available. The quality assessment was therefore restricted to visual analysis of height (difference) models. In the open non-vegetated areas, where image matching is possible with high confidence, these height models show a well acceptable correspondence between the DSMs being generated and the reference ground model. The differences can be quantified to be well within the 3*RMS-criterion, i.e. in an order of 5 meters or less. In order to quantitatively analyze and validate the DSM products, in particular with respect to vegetated/forested areas, higher quality reference data like Lidar DEMs would be required. Such kind of reference data are being collected at present.
380 H. Raggam et al. / Investigation of the Stereo-Radargrammetric Mapping Potential of TerraSAR-X
References [1] DLR, 2008. TerraSAR Ground Segment – Basic Product Specification Document, Doc. TX-GS-DD3302, issue 1.5, 103 pages [2] Fritz T., H. Breit, and M. Eineder, 2008. TerraSAR-X Products – Tips and Tricks. 3rd TerraSAR-X workshop, Oberpfaffenhofen, 25th – 26th November, 2008. Published at http://sss.terrasar-x.dlr.de/ [3] Hirschmüller H., Stereo processing by semiglobal matching and mutual information, PAMI, 2008. [4] Paar G. and Pölzleitner W., Robust disparity estimation in terrain modeling for spacecraft navigation, ICPR, 1992. [5] Raggam H. and K.H. Gutjahr, 1998. Evaluation of Radarsat Stereo Data. Proceedings of the European Conference on Synthetic Aperture Radar, EUSAR’98, 25-27 May 1998, Friedrichshafen, Germany, pp. 103-106. [6] Raggam H., 2005. Surface Mapping using Image Triplets – Case Studies and Benefit Assessment in Comparison to Stereo Image Processing. Photogrammetric Engineering and Remote Sensing, Vol. 72, Number 5, May 2006, pp. 551-563. [7] Raggam H., and K. Gutjahr, 2007. Comparative Assessment of Surface Mapping Procedures Using Images Acquired by the Digital High-Resolution UltracamD Camera. 27th Earsel Symposium, “Geoinformation in Europe”, Bolzano, June 4th – 7th, 2007. [8] Raggam H., K.H. Gutjahr, R. Perko, M. Schardt, 2008. Assessment of the Potential of TerraSAR-X with Respect to Mapping Applications Using Radargrammetric and Interferometric Techniques. 3rd TerraSAR-X workshop, Oberpfaffenhofen, 25th – 26th November, 2008. Published at http://sss.terrasarx.dlr.de/ [9] Schubert A., D. Small, E. Meier, and D. Nüesch, 2004. Extraction of Surface Topography from SAR Images Using Combined Interferometry and Stereogrammetry. Proceedings of the European Conference on Synthetic Aperture Radar, EUSAR'04, 24-27 May 2004, Friedrichshafen, Germany. [10] T. Toutin, 2000. Evaluation of Radargrammetric DEM from RADARSAT Images in High Relief Áreas, IEEE Transactions on Geoscience and Remote Sensing, vol. 38, no. 2, pp: 782-789, 2000.