towards airborne single pass decimeter resolution sar interferometry ... · u. stilla et al....

Towards Airborne Single Pass Decimeter

Resolution SAR Interferometry over UrbanAreas

Michael Schmitt1, Christophe Magnard2, Thorsten Brehm3, and Uwe Stilla1

1 Photogrammetry & Remote Sensing, Technische Universitaet Muenchen,Munich, Germany

[email protected], [email protected] Remote Sensing Laboratories, University of Zurich, Zurich, Switzerland

[email protected] Fraunhofer Institute for High-Frequency Physics and Radar Techniques,

Wachtberg, [email protected]

Abstract. Airborne cross-track Synthetic Aperture Radar interferome-ters have the capability of deriving three-dimensional topographic infor-mation with just a single pass over the area of interest. In order to get ahighly accurate height estimation, either a large interferometric baselineor a high radar frequency has to be used. The utilization of a millimeterwave SAR allows precise height estimation even for short baselines. Com-bined with a spatial resolution in the decimeter range, this enables themapping of urban areas from airborne platforms. The side-looking SARimaging geometry, however, leads to disturbing effects like layover andshadowing, which is even intensified by the shallow looking angle causedby the relatively low altitudes of airborne SAR systems. To solve this de-ficiency, enhanced InSAR processing strategies relying on multi-aspectand multi-baseline data, respectively, are shown to be necessary.

Keywords: InSAR, urban areas, very high resolution, DSM.

1 Introduction

Airborne single pass SAR interferometry (InSAR) allows deriving the topog-raphy of extended areas by combining two simultaneously acquired complexSAR images to an interferogram and then exploiting the phase of this inter-ferogram. With the advent of very high resolution SARs that are capable ofresolutions in the decimeter range, the analysis of urban areas by methodsbased on interferometric SAR data have become an increasingly important re-search topic. Several papers for instance have been published dealing with thechallenge of building extraction from InSAR datasets based on image analysis(Stilla et al., 2003; Thiele et al., 2007). Concurrently, more and more papers arepublished on the use of tomographic approaches exploiting multiple baselines to

U. Stilla et al. (Eds.): PIA 2011, LNCS 6952, pp. 197–208, 2011.c© Springer-Verlag Berlin Heidelberg 2011

198 M. Schmitt et al.

overcome the problems arising from the layover effect that has become more im-portant since resolutions have reached sub-meter qualities (Baselice et al., 2009;Reale et al., 2011). Last but not least, the use of InSAR images recorded fromdifferent aspect angles has been investigated to fill in the missing informa-tion in the shadow areas caused by the side-looking SAR imaging geometry(Thiele et al., 2010; Soergel et al., 2003).

This paper deals with the possibilities and challenges of utilizing airbornedecimeter resolution InSAR data for the analysis of densely built-up urban areasand is organized as follows: In Section 2, the MEMPHIS sensor is introduced,before in Section 3 forward and backward geocoding procedures are described. InSection 4 the test dataset considered in this paper is shown. Section 5 discussessome experimental results and proposes solutions to the occurring challenges.Section 6 finally draws a conclusion and gives an outlook.

2 MEMPHIS SAR System

MEMPHIS is an acronym for Millimeterwave Experimental MultifrequencyPolarimetric High Resolution Interferometric System and was developed bythe Research Institute for High- Frequency Physics and Radar Techniques ofthe FGAN, now Fraunhofer FHR (Schimpf et al., 2002). In the airborne side-looking configuration, it can be operated at 35 GHz (Ka band) as well as94 GHz (W band) with a bandwidth of either 200 MHz in low resolution mode or800 MHz in high resolution mode. With these modes, slant range resolutions of75 cm and 19 cm, respectively, can be realized. Interferometric single pass mea-surements, even in a multi-baseline configuration, are possible by installing fourreceiving antennas with different baselines. The smallest and the largest availablebaselines in 35 GHz mode are 5.5 cm and 27.5 cm and lead to height ambiguitiesof about 275 m and 55 m, respectively. The processing of the interferometricraw data is carried out by the Remote Sensing Laboratories of the Universityof Zurich and is described in (Magnard et al., 2007) and (Magnard et al., 2010).Since MEMPHIS is still an experimental system, it is commonly mounted on aC-160 Transall (see Fig. 1) and flown at low altitudes of about 300 to 1000 mabove ground level. In combination with a mean viewing angle of about 60◦, thisleads to a typical swath size of 600 m in range direction and up to 3000 m inazimuth direction.

3 Interferometric Geocoding Concepts

3.1 Interferometric Pre-processing

Before the interferometric data can be geocoded, i.e. three-dimensional informa-tion in a geodetically defined reference frame or map projection can be retrieved,the single look complex (SLC) SAR images have to be pre-processed. This con-ventionally includes the following steps:

Airborne Decimeter Resolution SAR Interferometry over Urban Areas 199

Fig. 1. MEMPHIS mounted on a C-160 Transall

– Image coregistration:Two SLC images have to be aligned to sub-pixel accuracy, e.g. by cross-correlation of their amplitude values. For MEMPHIS data, the misalignmentalready falls below one pixel due to the small distance between the receivinghorns and their inclination.

– Interferogram generation:The actual interferogram is created by calculating the product between thecomplex value of each pixel of one image by the complex conjugate of thecorresponding pixel of the second image.

– Multilooking and coherence estimation:To mitigate phase noise, a number of neighboring pixels have to be averaged,which is called complex multilooking for complex pixels. For decimeter res-olution data of densely built-up urban areas only adaptive algorithms forcomplex multilooking and simultaneous coherence estimation in the vein of(Vasile et al., 2004) or (Deledalle et al., 2011) are suitable.

– Interferogram flattening:In order to work with just the topographically induced phase, the phase con-tribution caused by the imaging geometry has to be removed. Therefor, thesimulated phase of the ideally flat earth is subtracted from the interferogram.

– Phase unwrapping:Since interferograms are complex images, calculating the phase gives an anglemodulo 2π. Thus, the interferometric phase has to be unwrapped if the heightambiguity is smaller than the overall height variations within the scene. Forsmaller baselines or relatively flat urban scenes without skyscrapers, thiscrucial step in SAR interferometry can be avoided.

– Absolute phase retrieval:From the filtered and unwrapped topographic phase, finally the absolute in-terferometric phase has to be retrieved. First, the phase image is unflattened,


i.e. the flat-earth phase contribution is re-added. Afterwards, at least a singleground control point has to be used to determine the so-called phase constantrefining the interferometric calibration.

3.2 Interferometric Geocoding

From the absolute interferometric phase, topographic height information can bereconstructed. Conventionally, this is done in slant range geometry by differentphase-to-height conversion methods (Small et al., 1996) before transforming theresulting height grid to a geodetic reference frame.

Forward Geocoding. Since the work of our group focuses on the utilizationof multi-aspect InSAR data, we have decided to use the straight-forward geom-etry of the interferometric range-Doppler equations (IRDE) that directly relateRADAR measurements and three-dimensional object points. For MEMPHIS alinear flight track as well as zero-Doppler processed data are assumed so we cansimply write:

R = ‖P − SM‖ (1)

V M (P − SM ) = 0 (2)

φ = −2π

λ(‖P − SM‖ − ‖P − SSl‖) (3)

R is the slant range distance between the object point P and the master sensorposition SM . V M denotes the master sensor velocity, while SSl symbolizes theslave sensor position. Finally, φ denotes the absolute interferometric phase andλ the RADAR wavelength.

Equations (1)-(3) form a non-linear equation system that has to be solved forthe unknown object coordinates P = [X, Y, Z]T for each interferogram pixel.The result is an irregular point cloud that can be fused with point clouds fromother aspects, before the data is resampled to the final surface grid.

Backward Geocoding. As proposed in (Schmitt & Stilla, 2011), the Dopplerequation (2) can easily be inverted to

t =(P − S0,M )V M

‖V M‖2 , (4)

so that starting from a known object point P the azimuth time t and with (1)and (3) also the slant range distance R and the ideal absolute interferometricphase φ can be calculated straight-forwardly. S0,M here denotes the mastersensor position for the first azimuth bin.


Then the observed interferometric phase corresponding to the (t, R) SARimage coordinates is measured, so one can calculate the difference of ideal (i.e.simulated) phase and measured phase using

Δφ = ‖φmeasured − φsimulated‖ . (5)

If this is carried out for all hypotheses of a height search interval positioned ina pre-defined surface grid, the geocoded height can be determined by searchingfor the minimal difference between simulated and measured phase for all heighthypotheses. The main advantage of backward geocoding compared to conven-tional forward geocoding is that the result is not an irregular point cloud, but a2.5D height model in already gridded form.

4 The Test Dataset

Our study area is the main campus of Technische Universitaet Muenchen (TUM)located in the Maxvorstadt neighborhood of Munich, Germany (target coor-dinates: 48◦08′56′′N, 011◦34′02′′E). The scene mainly contains dense buildingblocks, but also some larger buildings surrounded by patches of concrete orlawn, respectively, as well as many urban trees. An optical image of the testarea is shown in Fig. 2.

Fig. 2. Optical image of the test scene c©Google Earth

The test dataset originally comprised of four data takes in multi-aspect con-figuration with non-orthogonally crossing flight trajectories (see Table 1). Due tothe fact that only low-precision navigation data was available, problems duringSAR focussing occurred so that the fourth aspect had to be discarded.


Table 1. Test Dataset (Munich)

Radar band Ka band (35 GHz)

Wavelength 8.55 mm

Pixel spacing (slant range) 5 cm (azimuth) / 19 cm (range)

Effective baseline used 27.5 cm

Heading angles 11◦/70◦/169◦/(250◦)Off-nadir angle 60◦

Aircraft altitude 2460 ft AGL

Although for each aspect 4 SAR images were acquired simultaneously, onlythe two acquisitions with the longest baseline were used for this preliminarystudy. Since no phase unwrapping was required for the test scene, a utilizationof multi-baseline techniques is at the moment investigated for the solution ofthe layover problem. The magnitude image and the phase image of aspect 3 asderived by the filter proposed in (Deledalle et al., 2011) is shown in Fig. 3.

Fig. 3. Left: MEMPHIS magnitude image of the test scene. Flight direction (headingangle: 169◦) is from top to bottom, range direction from left to right. Right: Flat earthcorrected interferometric phase of the same aspect.

5 Interferometry Results for the Munich Test Scene

In order to assess the potential of airborne millimeter wave SAR interferometryover urban areas, conventional interferometric processing utilizing both forwardas well as backward geocoding have been carried out for the test dataset intro-duced in section 4. Note that all heights are given above the WGS84 ellipsoid.

5.1 Forward Geocoding Results

Fig. 4 exemplarily shows the height reconstruction achieved by classic forwardgeocoding for aspect 1 (heading angle 11◦). The point cloud was rasterized by


Fig. 4. Height image generated by forward geocoding of aspect 1. Note the black imageparts containing no information. The heading angle of the flight was 11◦.

simply tiling the dataset with respect to a grid resolution of 0.2 m, which corre-sponds to the range resolution of the MEMPHIS data and is thus also consideredfor the backward geocoding procedure. If multiple points fell into one tile, theaverage height was chosen for this tile. Additionally, the resulting height imagewas filtered with a maximum homogeneity neighbor filter (Garnica et al., 2000).

It has to be emphasized that the black parts of the height image are notcaused by RADAR shadow. Since the shadowing effect does appear in SARimages, there is seemingly an interferometric phase also in shadow parts of thedata - although these phase observations just contain noise and no topographicinformation. However, the parts of the surface grid that dont contain any heightvalues simply were not imaged by the sensor.

In comparison, the consequence of shadowing is shown in Fig. 6, depictingimage details of the results for aspect 3 (heading angle 169◦). It can be noticed,how phase noise caused by RADAR shadow results in height noise.

5.2 Backward Geocoding Results

In order to show the applicability of backward geocoding for interferometricdata of urban scenes, the method mentioned in section 3.2 was also applied tothe data of aspect 1. As can be seen in Fig. 5, now every part of the scene iscovered with height values, although a daubing effect appears where actually noinformation was available (cf. Fig. 4).

5.3 Comparison of Forward and Backward Geocoding

According to section 5.1 and 5.2 it is obvious that both geocoding methodsbasically lead to the same result. The main difference is that the forward createddata seems noisier, which can especially be seen in Fig. 6.


Fig. 5. Height image generated by backward geocoding of aspect 1. Not that the heightsearch interval was centred on a mean terrain height of 570 m.

Recalling section 5.1, Fig. 6 also enables a comparison of the effects of shadow-ing and surface parts that were not mapped by the sensor: To the northern sideof the “Alte Pinakothek” building, artificial height noise appears, correspondingto a patch of low coherence caused by RADAR shadow, while in the forwardgeocoding data several smaller areas don’t contain any height information at all.

It has also to be noted that many obvious errors appear in both height maps.This can be caused by several phenomena:

– For this study, we have considered every interferogram pixel during geocod-ing. A masking of low-coherent phase observations was not applied.

– Although a first sidelobe reduction has been carried out using a Kaiser Win-dow (Thomas et al., 2000), some strong sidelobes have not yet been removedfrom the data. They can easily be perceived in the magnitude and phase im-ages shown in Fig. 3.

– Additional errors may have been induced due to the fact that no high-precision inertial navigation data was available for SAR focussing.

New data has been acquired during a campaign in June 2011 making use ofa high-end inertial navigation system. It is therefore expected that the overallquality of the data can be notably improved in the future.

5.4 Benefit of Multi-aspect Data

The benefit of a multi-aspect InSAR configuration can be assessed by combiningthe point clouds resulting from the aspects 1, 2, and 3 (see Fig. 7).

For Fig. 7, all surface points which are based on a low-coherent phase (γ <0.5), mainly indicating shadow areas in single pass InSAR data, were masked


Fig. 6. Height grid detail for the “Alte Pinakothek” generated from aspect 3 (headingangle: 169◦). Left: Forward geocoding results. Right: Backward geocoding results. Top:Reconstructed heights. Bottom: Geocoded coherence values.

Fig. 7. Binary maps of reliable surface points. From top left to bottom right right:Aspect 1 (heading angle: 11◦), aspect 2 (heading angle: 70◦), aspect 3 (heading angle:169◦); combination of all three aspects.


in the backward reconstruction grids of all three available aspects. The combi-nation, however, shows that the overall coverage can be significantly enhancedby utilizing multi-aspect InSAR acquisitions. This is especially noticeable in thelarge shadow area located to the northern side of the “Alte Pinakothek” build-ing, where now almost all parts of the scene are covered by interferometric phaseobservations with appropriate coherence.

5.5 Additional Remarks

It has to be stated, however, that, although the overall structure of the scenecan be perceived from the interferometric reconstruction, there are quite someproblematic parts, especially, where the building blocks stand very close to eachother (cf. the upper left corners of Figs. 4 and 5, respectively).

This is caused by an intermixture of layover and shadowing as sketched inFig. 8.

Fig. 8. Layover and shadowing for densely built-up inner city areas. Note that mostparts of the scene are affected by at least one of both effects.

As can be seen in Fig. 8, most parts of a dense urban scene are affected byat least one of the two important SAR imaging effects, while even mixtures oflayover and shadowing can occur (e.g. [D’C’] or [E’F’] in Fig. 8). Thus, not only“front porches” appear in the resulting height data, but also even more seriousdisturbances, since the noise phase caused by RADAR shadow and the phaseramps (Thiele et al., 2007) caused by layover mix.

In order to give a coarse approximation, the ground area affected by layoverand shadowing can be given by

lLayover = cot (θ) · h (6)

andlShadow = tan (θ) · h, (7)

respectively. With an average off-nadir angle θ = 60◦ and a typical buildingheight h = 15m, layover spans about 8.7 m in front of the building, while shad-owing affects theoretically about 26 m of ground. If you now consider a typical


road width of also 15 m, roads are not visible at all, while a large fraction is af-fected by both shadowing and layover. Furthermore, as can also be seen in Fig. 8,the bottom part of the second building’s front facade is not imaged either. Sum-marized this means: With the typical low-altitude configuration of MEMPHIS,shadowing prevents a full 3D reconstruction of buildings for densely built-upinner-city areas. The reconstruction of street heights then again can only besolved by utilizing an orthogonal aspect with a sensor looking direction parallelto the road. Still, problems will occur in the courtyards of building blocks wherethe ground is shadowed from all directions. Thinking about future processingstrategies, ideally the shadow pixels are masked and not considered for furtherprocessing, while the layover pixels have to be further analyzed in order to findthe top height contribution of the respective resolution cell.

6 Conclusion and Outlook

From the experimental results shown in this paper, two main conclusions areobvious:

1. The problem of RADAR shadowing can partly be tackled by utilizing InSARdata from multiple aspects. This way, completing height reconstruction byfilling in the missing heights of one aspect by data from different aspectsbecomes possible.

2. Nevertheless, the layover effect still causes artificial heights (so-called “frontporches”) which need to be coped with separately. Ongoing research focuseson the exploitation of MEMPHIS’ multi-baseline capabilities.

However, the most severe problems of urban mapping by means of (interferomet-ric) SAR remote sensing occur in high-density areas, where frequently layoverand shadowing mix. Additionally, due to low flying heights and relatively flatlooking angles, some parts of the city scene will not be imaged at all.

It is, however, expected that by combining multi-aspect and multi-baselinedata, both effects will be mitigated in a way such that digital surface model(DSM) generation of urban areas by airborne single pass SAR interferometry inthe decimeter range will become feasible.

References

Baselice, F., Budillon, A., Ferraioli, G., Pascazio, V.: Layover Solution in SAR Imaging:A Statistical Approach. IEEE Geoscience and Remote Sensing Letters 6, 577–581(2009)

Deledalle, C.-A., Denis, L., Tupin, F.: NL-InSAR: Nonlocal Interferogram Estimation.IEEE Transactions of Geoscience and Remote Sensing 49, 1441–1452 (2011)

Garnica, C., Boochs, F., Twardochlib, M.: A New Approach to Edge-PreservingSmoothing for Edge Extraction and Image Segmentation. International Archivesof Photogrammetry, Remote Sensing and Spatial Information Sciences 33(B3),320–325 (2000)


Magnard, C., Meier, E., Ruegg, M., Brehm, T., Essen, H.: High Resolution MillimeterWave SAR Interferometry. In: Proceedings of IEEE International Geoscience andRemote Sensing Symposium 2007, pp. 5061–5064 (2007)

Magnard, C., Meier, E., Small, D., Essen, H., Brehm, T.: Processing of MEMPHISMillimeter Wave Multi-Baseline InSAR Data. In: Proceedings of IEEE InternationalGeoscience and Remote Sensing Symposium 2010, pp. 4302–4305 (2010)

Reale, D., Fornaro, G., Pauciullo, A., Zhu, X., Bamler, R.: Tomographic Imaging andMonitoring of Buildings with Very High Resolution SAR Data. IEEE Geoscienceand Remote Sensing Letters 8, 661–665 (2011)

Schimpf, H., Essen, H., Boehmsdorff, S., Brehm, T.: MEMPHIS A Fully PolarimetricExperimental Radar. In: Proceedings of IEEE International Geoscience and RemoteSensing Symposium 2002, pp. 1714–1716 (2002)

Schmitt, M., Stilla, U.: Fusion of Airborne Multi-Aspect InSAR Data by SimultaneousBackward Geocoding. In: Proceedings of Joint Urban Remote Sensing Event 2011,pp. 53–56 (2011)

Small, D., Pasquali, P., Fuglistaler, S.: A Comparison of Phase to Height ConversionMethods for SAR Interferometry. In: Proceedings of IEEE International Geoscienceand Remote Sensing Symposium 1996, pp. 342–344 (1996)

Soergel, U., Thoennessen, U., Stilla, U.: Iterative Building Reconstruction in Multi-Aspect InSAR Data. International Archives of Photogrammetry, Remote Sensingand Spatial Information Sciences 34 (3/W13), 186–192 (2003)

Stilla, U., Soergel, U., Thoennessen, U.: Potential and Limits of InSAR Data for Build-ing Reconstruction in Built-Up Areas. ISPRS Journal of Photogrammetry and Re-mote Sensing 58, 113–123 (2003)

Thiele, A., Cadario, E., Schulz, K., Thoennessen, U., Soergel, U.: InSAR Phase Profilesat Building Locations. International Archives of Photogrammetry, Remote Sensingand Spatial Information Sciences 36 (3/W49A), 203–208 (2007)

Thiele, A., Wegner, J.D., Soergel, U.: Building Reconstruction from Multi-Aspect In-SAR Data. In: Soergel, U. (ed.) Radar Remote Sensing of Urban Areas, pp. 187–214.Springer Science+Business and Media B.V., Heidelberg (2010)

Thomas, G., Flores, B.C., Sok-Son, J.: SAR Sidelobe Apodization Using the KaiserWindow. In: Proceedings of IEEE International Conference on Image Processing2000, pp. 709–712 (2000)

Vasile, G., Trouve, E., Ciuc, M., Buzuloiu, V.: General Adaptive-Neighborhood Tech-nique for Improving Synthetic Aperture Radar Interferometric Coherence Estima-tion. Journal of the Optical Society of America A: Optics, Image Science, and Vi-sion 21, 1455–1464 (2004)

towards airborne single pass decimeter resolution sar interferometry ... · u. stilla et al....

Documents