Removal of Residual Errors From SAR-Derived Digital Elevation
Models For Improved Topographic Mapping of Low-Relief Areas

K. Clint Slatton (1), Melba M. Crawford (1), James C. Gibeaut (2), and Roberto Gutierrez (2)
(1): Center for Space Research, University of Texas at Austin
3925 W. Braker Ln., Suite 200, Austin, TX 78759-5321
(2): Bureau of Economic Geology, University of Texas at Austin

This work was supported by the Texas Regional Change Program through the Texas Space Grant Consortium and the Johnson Space Center, a National Aeronautics and Space Administration grant under the Topography and Surface Change Program (Grant NAG5-2954), and the Texas Advanced Technology Program.
Abstract -- Interferometric synthetic aperture radar data can be used to precisely map topography, but low-relief areas are problematic because errors in the data can be large compared to the topographic variations. The NASA/JPL TOPSAR system acquired data over a low-relief test site on the Texas coast in 1996. Due to unusually high turbulence during the acquisition and the mild topography, residual height errors were visible in the Digital Elevation Model (DEM). The characteristics of the error signal are described and a method is outlined for removing the residual error and mosaicking the affected TOPSAR frames.

INTRODUCTION

1 Land surfaces with mild topography, such as river floodplains and coastal zones, are typically very prone to flooding due to precipitation and storm-surge events. Topography-based flood models have been developed that predict the extent and severity of flooding in such areas under a variety of circumstances, and Digital Elevation Models (DEMs) are needed as inputs to these topography-based models [1]. Because errors in the DEMs propagate directly into the predictions of flood extent, it is important to maximize the accuracy and precision of the DEMs that are used in these models.

Standardized and georeferenced DEMs are produced by the United States Geological Survey (USGS) and are widely available for most of the United States. These DEMs are often sufficiently precise for areas with significant topography because any height errors will typically be small relative to the actual surface height variations. The specification for the Level 1 USGS DEM is ¾15 m vertical root mean square (rms) error. Figure 1 shows a Level 1 DEM for the test site. The horizontal data spacing is 30 m and the vertical data spacing is 1 ft [2]. However, greater precision is needed when mapping low-relief areas because these errors in the DEMs may be of similar magnitude to the surface height variations, and so have a great effect on the accuracy of flood models that use the DEMs. DEMs used for flood modeling must also have fine horizontal resolution so that small hydrologically-important features, such as stream beds, are accurately mapped. Improved resolution can potentially be achieved by generating DEMs using radar data.

Figure 1. USGS derived DEMFigure 2. SAR derived DEM

In June 1996, the NASA/JPL AIRSAR system collected interferometric synthetic aperture radar (INSAR) data over low-relief regions on the coast of Texas, USA. To collect INSAR data, the AIRSAR system operates in its topographic (TOPSAR) mode. Figure 2 shows two mosaicked TOPSAR frames over Mission Bay, Texas. The ground-range-projected TOPSAR DEMs have data spacings of 10 m horizontally and 0.1 m vertically. The TOPSAR data are within sensor specifications in terms of rms height error [3]. However, small systematic height errors are still visible in the original DEMs because the area has such low relief. This paper describes some of the errors observed in the data and outlines the procedures used to minimize those errors and mosaic the DEMs.

INSAR BACKGROUND

DEMs, like those produced by the USGS, have traditionally been derived from stereo processing of aerial photography or optical spaceborne data. In recent years, DEMs have also been derived from SAR data using interferometric processing. The data for INSAR DEMs may be acquired day or night and in most weather conditions, but the primary advantage of INSAR methods is that the elevation of each pixel is determined independently. In stereo-optical DEMs, individual pixels are binned into discrete elevations to create noise-free closed-contour topographic maps. The primary disadvantages of INSAR DEMs are their sensitivity to sensor motion and their noise characteristics.

DEMs can be generated from INSAR data by combining two complex (phase and magnitude) SAR images acquired from similar vantage points [4]. Once the two images are co-registered, a differential phase can be calculated for each pixel. Using a known position of at least one pixel and unwrapping the modulo 2p phase, a map of absolute phase differences is generated. Geometric relationships can then be used to create a height map (DEM) relative to the radar position. The height map can be referenced to a geocentric coordinate system by collecting Global Positioning System (GPS) data onboard the sensor platform.

Most of the work to date in generating INSAR DEMs has focused on data collected from spaceborne systems using multiple observations (repeat-pass). In particular, the European Remote Sensing satellites (ERS-1 and -2) have been used extensively for this purpose [5]. However, any changes that occur in surface or atmospheric conditions in the imaged area between observations will introduce errors into the subsequent DEM. The shortest time interval between observations suitable for INSAR processing is about one day for the tandem ERS system [6]. Significant changes in the backscattering properties of the surface or refractive properties of the atmosphere due to precipitation or humidity changes can occur on this time scale, thus reducing the number of suitable image pairs [7].

INSAR systems with more than one antenna, such as TOPSAR, can make dual observations simultaneously so that decorrelation of the scene through time is not a factor. This is especially important for vegetated, humid regions, such as the Texas coast, which can decorrelate rapidly. TOPSAR data are also available at higher spatial resolution than currently-available spaceborne data, (e.g. 25 m for ERS-1). This improves the mapping of small-scale features. The primary disadvantage of single-pass airborne systems is that the platform motion is perturbed more frequently and in a less deterministic manner than spaceborne platforms. Standard processing of TOPSAR data does include motion compensation, but if the motion is severe or high-frequency, residual errors on the order of ± 1.5 m may be observed in the DEMs. If the actual topographic variations are on the order of ¾10 times this magnitude, the error signal may be visible in the DEMs.

TEST SITE

TOPSAR flightlines were acquired along coastal stream beds in the San Antonio-Nueces watershed on the Texas coast. This watershed is located on a low-lying coastal plain. Flightlines were oriented approximately normal to the shoreline to observe the topography along the streams that carry most of the water runoff to the bays. The topographic variation in the 20 km nearest to the shore is only about 13 m. The TOPSAR data analyzed for this paper are from a flightline over Mission Bay. Hurricane models implemented for similar areas along the Texas coast predict storm surge penetrations of up to 15 km inland for a category 1 hurricane (74-95 mph winds), with flooding distributions that are highly dependent upon small topographic variations such as stream beds [1].

CHARACTERIZING THE DATA

1st order errors in the TOPSAR DEMs are manifest as planar tilting in range. This tilting is the result of uncompensated path delays in the radar system. When mapped into heights, those time delays can produce linear slopes in the DEMs. The DEMs can also exhibit higher order errors due to aircraft motion. Errors due to aircraft motion were observed in one of the TOPSAR frames over the Mission Bay test site.

Two adjacent 10 km x 10 km TOPSAR frames were acquired from a single flightline. A periodic signal superimposed on the topography was apparent in one of the frames. This "ripple" was primarily a function of azimuth, but also exhibited a weak inverse dependence on range. The approximate peak-to-peak amplitude was 3 m, and there were 8 complete periods in the frame. A printout of the aircraft motion file was obtained and the ripple signal appeared to be exactly correlated with the roll motion of the aircraft, which exhibited an 8 Hz frequency and peak-to-peak amplitude of 1°. Neither yaw nor pitch motion exhibited any significant correlation with the ripple signal.

TOPSAR DEMs have demonstrated relative rms height errors of 1-2 m in relatively flat areas [3]. The DEMs acquired over this test site exhibited rms error levels well within those reported levels. The residual errors due to aircraft motion were visible because nearby storms produced excessive turbulence during the acquisition and the total topographic variation in the test site is only about 10 times the magnitude of the residual signal.

ERROR REMOVAL AND EVALUATION

To produce a mosaicked DEM strip from individual TOPSAR frames, the relative errors must be corrected. After an internally consistent DEM strip is produced, it can be georeferenced using GPS data collected on the ground. The following procedures were followed.

1.0 Correct relative errors
1.1 filter out the motion signal
1.2 image-to-image registration
1.2.1 1st order correction of elevations
1.2.2 standard 2-dimensional registration
1.3 smooth noise over low-backscatter targets

2.0 Georeference the DEM strip
2.1 image-to-GPS registration
2.1.1 1st order correction of elevations
2.1.2 standard 2-dimensional registration

It is necessary to correct the relative errors before georeferencing so that overlapping portions of adjacent DEMs will only differ to a 1st order. A stop-band Infinite Impulse Response (IIR) filter was used to remove the ripple signal. The filter removed the 8 Hz ripple while preserving small-scale topographic features.

Elevations of features in the overlap between the two DEMs were used to add the best planar correction (in a least squares sense) to the slave DEM to obtain agreement with the master DEM's elevations to a 1st order. Those same control points were then used to do a 2-dimensional image-to-image registration to mosaic the two DEMs. JPL is currently developing the capability to output continuous strips of TOPSAR data, which will eliminate the need for mosaicking frames on a single flightline. Some open water areas in the far range of the DEMs exhibited very low signal to noise ratios (SNR), which were manifest as regions with very high-frequency, large-magnitude noise. These areas were assigned a constant elevation equal to elevation of the surrounding bank.

Georeferencing the DEMs was accomplished via 3-dimensional registration to GPS tie points after the DEMs were made internally consistent and mosaicked. The DEMs are georeferenced during the operational processing at JPL by giving the latitude/longitude of the scene center, but more accurate in situ georeferencing is needed for the DEMs in low-relief areas.

Static GPS points were collected for georeferencing, but more static GPS points will be collected to validate these results. Kinematic GPS transects have also been collected along several roads in the imagery, but the solutions have not yet been analyzed. Figure 3 shows transects extracted from the co-registered TOPSAR and USGS DEMs. The transects show that the superior resolution of TOPSAR allows it to capture topographic variations that are not resolved in the USGS DEM. The TOPSAR data also exhibit greater variability due to noise and non-surface features such as trees.

CONCLUSIONS

The higher-order errors observed in these data do not appear to be significant in most TOPSAR DEMs. The errors were visible in these data because of the extreme low-relief of the region and the proximity of storms during the acquisition.

Future work will include improvements to the filtering of the motion signals and validation of the results with more GPS surveys. However, these preliminary results do indicate that systematic errors can be minimized and precise DEMs can be generated for low-relief areas using TOPSAR data.

REFERENCES

[1] Texas A&M University, College of Architecture, Storm Atlas: Brazoria, Galveston, and Harris Counties, Texas A&M University, pg. 3-1, September 1993.

[2] United States Geologic Survey (USGS), Digital Elevation Models: Data Users Guide 5, USGS, pg. 14, 1993.

[3] Madsen, S. N., J. M. Martin, and H. A. Zebker, "Analysis and Evaluation of the NASA/JPL TOPSAR Across-Track Interferometric SAR System", IEEE Trans. Geosci. Remote Sensing, vol. 33, no. 2, pg. 383-391, March, 1995.

[4] Zebker, H. A. and R. M. Goldstein, "Topographic Mapping from Interferometric Synthetic Aperture Radar Observations", Journal of Geophysical Research, vol. 91, no. B5, pg. 4993-4999, April 10, 1986.

[5] Massonnet, D. and K. L. Feigl, "Discrimination of Geophysical Phenomena In Satellite Radar Interferograms", Geophysical Research Letters, vol. 22, no. 12, pg. 1537-1540, June 15, 1995.

[6] Schwäbisch, M., M. Matschke, W. Knöpfle, and A. Roth, "Quality Assessment of INSAR-Derived DEMs Generated With ERS Tandem Data", Proceedings of IGARSS'96, pg. 802-804, 1996.

[7] Kenyi, L. W. and Hannes Raggam, "Atmospheric Induced Errors In Interferometric DEM Generation", Proceedings of IGARSS'96, pg. 353-355, 1996.