| GISdevelopment.net ---> Technology---> Remote Sensing |
Evaluation of Digital Elevation Models created from different satellite images
Dr. K.S.Siva Subramanian, Amitabh Singh, Manda Sudhakar
RMSI Private Limited, A - 7 Sector 16, NOIDA 201 301, India
Tel +91 120 251 1102, 251 2101, Fax +91 120 251 1109, 251 0963
Dr. K.S.Siva Subramanian, Amitabh Singh, Manda Sudhakar
RMSI Private Limited, A - 7 Sector 16, NOIDA 201 301, India
Tel +91 120 251 1102, 251 2101, Fax +91 120 251 1109, 251 0963
Abstract
The generation of Digital Elevation Models (DEMs) from satellite images scores over conventional methods of DEM generation using topographic maps and aerial photographs. This global availability of satellite images allows for quicker data processing for an equivalent area. This paper discusses the evaluation of DEM, generated using stereo pairs of different satellites. The various satellite datasets taken for study include Radarsat standard beam mode data, IRS 1-C Pan, ASTER VNIR bands, and Spot Pan data. PCI Geometica OrthoEngine v8.2 software has been used for DEM generation, and Erdas Imagine v8.4 has been used for the DEM evaluation. Each of the source image pairs used for DEM creation have their own merits and demerits. The Radar images are generally processed in two different ways viz. interferometric method and stereoscopic method. An analysis of the literature survey shows that the interferometric method yields better results in a relatively low relief area but not so in moderate relief areas. Conversely the stereoscopic method gives better results for moderate relief areas falling in flat topographic terrains.
In addition the beam mode selection too plays a key role in the output DEM quality. A larger beam mode separation for the stereo pair (say S2-S5) results into a good parallax. The automated pixel matching method adopted by the software, however fails when the parallax is too high. A low beam mode separation (say S2-S3), on the other hand results in errant values in low relief areas. The problem is further compounded by the presence of speckles.
Alternatively hand stereo pairs obtained by satellites, operating in optical mode yield more consistent results over different kinds of terrains. The only method of processing these images is the stereoscopic method of measuring parallax by pixel matching.
Studies reveal that of the three different pairs processed, the ASTER images yield the best results. This is possible because the temporal difference between the stereo pair is only a few seconds, whereas pairs acquired through other satellites require a larger time cycle. Results also show that raster match is good in case of ASTER images due to short temporal differences. Further the unique fixed b/h ratio yields more consistent and better results over various terrains.
Introduction
Creating Digital Elevation Model (DEM) by digitizing contour lines from topographic maps or through stereoscopic semi automated methods from aerial photographs are proven methods. However, DEM generation from satellite stereo image pairs of optical and microwave sensors, is still not a common practice. The DEM generated from satellite stereo pairs have some significant advantages over other sources, viz:
The present work has been undertaken:
Inputs details
* ASTER satellite generates an along track stereo pair where, one image is captured in Nadir view and the other is captured 4 seconds later by a backward looking camera by viewing the same geographic area as seen by the Nadir view camera
** These satellites use across track viewing to generate a stereo pair, i.e., the same geographic area is viewed from two different tracks by tilting the cameras
Software used
GCPs and stereo GCPs are collected from reference maps in order to reference the images to ground. Further, TPs (Tie points) are collected to improve matching between the two stereo pairs. The following table gives the number of GCPs and Tie Points used to create a DEM:
Once registered to the same ground area, any along track (in case of SPOT, IRS and Radarsat images) or across track (in case of ASTER images) positional differences are assumed to be due to parallax, which results due to relief. Measured parallax differences on a pixel-by-pixel basis are converted to absolute elevations using trigonometric functions and the orbital data (orbital position, altitude, attitude and the scene center). The computation relies on the inherent parallax between stereo images. An automated image correlation algorithm [1] is used to derive elevations from the parallax, by a set of well located GCPs and tie points (TPs).
The image matching system operates on a reference and a search window. For each position in the search window, a match value is computed from gray level values in the reference window. The match value is computed with the mean normalized cross-correlation coefficient and the sum of mean normalized absolute difference [2]. The correlation window size varies from low resolution (8 pixels) to 32 pixels at the full resolution. Elevation points are extracted at every pixel for the complete stereo pair. The 3-D intersection is performed using the above computed geometric model [3] to convert the pixel coordinates in both images determined in the image matching of the stereo pair to the three dimensional data [4].
The output elevations are not computed for the pixels where the image matching fails to find the corresponding pixel in the reference image, resulting into some failure areas. In case of small and scattered failures the software does interpolate and compute most probable values for them. The DEM thus generated is in raw state and does not contain geo-referencing information. So, the DEM needs to be georeferenced.
Workflow
Accuracy assessment
The accuracy assessment has been done by using the best reference DEMs (USGS, Toposheets etc.) available. The analysis of error in the output DEM can be done by the following two methods:
The accuracy of the DEM has been expressed as linear error (LE) and circular error (CE). Linear error represents the vertical accuracy while circular error is for horizontal positional accuracy. Linear error gives an estimate of the overall accuracy of the DEM at the said confidence level (it is assumed that X, Y are correct and a comparison in Z is performed). An LE90 of 25 implies that 90% of the pixel within the DEM varies from the reference DEM by 25m or less. Circular Error (CE) is used to measure the combined errors in X and Y of the final DEM. This is a circular radius in meters, which will include 90% of the positional errors of the final DEM versus the reference DEM. (in other words, if 100 shots have been fired at a target, how large a circle should be drawn in order to accommodate 90 shots around bull’s eye). The radius of that circle is the CE 90 for the final DEM.
A model has been created in Erdas imagine software to compute the pixel level elevation differences both for negative and positive values as positive number.
Results & discussion
Several combinations of GCPs and TPs have been attempted during the process of DEM generation but only the best results obtained are being discussed here.
The RADARSAT derived DEM
The minimum elevation in the output Radarsat DEM (Fig. 1) was –80m while the maximum was 3580m. This elevation range shows that this area is a rugged terrain with high relief. The output Radarsat DEM was validated with respect to the USGS 30m DEM. The vertical accuracy LE80 (failed area constituting 15% was not used for error computation) was 65m with positional accuracy CE90 at 25m and standard deviation in overall error of 45m. The failure areas are restricted either in high escarpment areas or in the mountainous terrain. The reason for such failures appears to be principally due to layover and shadow effects of such terrains.
Accuracy of the Radarsat DEM can be further improved by using an additional image for the same area. For example in high relief areas the Layover and shadow effect is more pronounced in the used stereo pairs (S1 and S6), as the difference in Local Incidence angle is very high and which, probably accounts for DEM failure in these areas. To get a good DEM in all such areas, along with S1 image, we may use another image of S2/S3 beam mode that have close incidence angles in order to get a better DEM.
Fig. 1: A) S1 beam mode, Ascending , B) S6 beam mode, Ascending, C) DEM at half resolution (25m) and D) DEM (Thematic at 50m interval).
The following graph shows horizontal distance in meters (X-axis) for a selected cross section and Y-axis shows the elevations derived from Radarsat image and USGS DEM elevations at same points. It may be observed that both of the lines maintain the same pattern with the elevations on moderate slopes exhibiting very close match, whereas on hilltops the difference increases to about 80m higher than actual.
The Radar images are processed in two different ways viz. interferometric method and stereoscopic method. Interferometric method yields better results in a relatively low relief area but not so in moderate relief areas, whereas, the stereoscopic method gives better results for moderate relief areas and failing in flat terrains [3].
Alternatively, stereo pairs obtained by satellites, operating in optical mode yield more consistent results over different kinds of terrains. The only method of processing these images is the stereoscopic method of measuring parallax by pixel matching.
The SPOT derived DEM
The DEM extracted from SPOT stereo images (Fig: 2) contain the elevation range of 1m to 1722m. The DEM created at full resolution (i.e., 10m) has resulted into large areas of failure and hence the same has been regenerated at half resolution (20m), which has shown significant improvement in the output with less than 5% failure area. The resulting DEM was validated with USGS DEM of 30m resolution. At 32m was the value at LE90, and CE90 was at 19.5m having a standard deviation also of 32m.
Fig. 2: A) SPOT PAN Left Image (Clip), B) SPOT PAN Right Image (Clip), C) Digital Elevation Model at 20m resolution and D) DEM (Thematic at 50m interval).
As can be observed in the graphic presentation of results (above) that the SPOT derived DEM shows variable results in different slope/terrain conditions. Consistently low values can be seen in the diagram when compared to USGS DEM (30m resolution).
The IRS 1C derived DEM
The Dehradun area is a mixed terrain of moderate relief of the outer Himalayas and Gangetic plains. The IRS stereo pair (Fig. 3) used for this terrain yielded a minimum elevation of 418m to a maximum of 1475m. The DEM thus obtained, was validated with DEM extracted from 1:50,000 scale, Survey of India topographic maps. The LE 90 was 35m and 18m positional accuracy (CE90). The standard deviation in error this case was 34.5m. In this case also the failure area is less than 5% of the complete area.
Fig. 3: A) IRS-IC PAN Left Image (Clip), B) IRS-IC PAN Right Image (Clip), C) Digital Elevation Model at 25m resolution and D) DEM (Thematic at 50m interval)
The following distance vs elevation graph for IRS Pan derived DEM and topographic map derived DEM shows significantly close values for both of the sources whereas secondary undulations in the terrain are depicted better in the IRS derived DEM.
The ASTER derived DEM
The elevation range of the DEM obtained from the ASTER stereo images (Fig. 4) was 1000m to 3696m. The ASTER DEM was validated with the DEM extracted from contours obtained from 1:100,000 scale, Russian topographic maps. The vertical accuracy (LE90) was 30m with 20m positional accuracy (CE90). The standard deviation in the measured error w.r.t. reference DEM was 28m. The failure pixels were significantly less (less than 5%), which probably is because of the near synchronous acquisition of stereo images (difference of 4 seconds only) by the satellite.
Fig. 4: A) ASTER VNIR (nadir), Clip, B) ASTER VNIR (backward looking) Clip, C) Digital Elevation Model at 15m resolution and D) DEM (Thematic at 50m interval)
The following plot of distance (X-axis) vs elevation (Y-axis) also exhibits a good agreement of the ASTER DEM with topographic map derived DEM with minor variation in positional accuracy.
The following points have been important, while creating good DEM:
*In the case of Radarsat LE 80 has been considered
Conclusion
Good quality DEMs can be created from different satellite image stereo pairs. The number and distribution of GCPs and TPs play a crucial role in the final accuracy of the DEM.
Layover and shadow effects in used Radar images were pronounced, rendering it a not so viable option for high relief areas. However, judicious selection of beam mode combination and by using more than one pair would give better results.
Study reveals that, of the three different pairs of optical sensor data processed, the ASTER images yield the best results. The elevation range and relief covered by ASTER image in this sample study was also significantly large (2700m approx.). The better results appear to be due to the negligibly small temporal difference between the stereo pair (only four seconds), whereas pairs acquired through other satellites in optical mode have comparatively large temporal variation. Further, the unique fixed b/h ratio yields more consistent and better results over various terrains. The net cost of DEM derived from ASTER images also is very less as compared to IRS and SPOT images, which cost nearly three times more than ASTER.
Since the output DEM reflects a pixel level true picture of the terrain, the DEMs thus created can be used for various application areas of GIS like Telecom, Watershed delineation, 3-D analysis etc.
References
Thierry Toutin, “3D Topographic Mapping with ASTER Stereo Data in Rugged Topography”, Personal communication to the authors, 2001.
PCI Geomatics, “OrthoEngine SE 3D, v8.2” in Reference Manual, Version 8.2 (October19, 2001).
The generation of Digital Elevation Models (DEMs) from satellite images scores over conventional methods of DEM generation using topographic maps and aerial photographs. This global availability of satellite images allows for quicker data processing for an equivalent area. This paper discusses the evaluation of DEM, generated using stereo pairs of different satellites. The various satellite datasets taken for study include Radarsat standard beam mode data, IRS 1-C Pan, ASTER VNIR bands, and Spot Pan data. PCI Geometica OrthoEngine v8.2 software has been used for DEM generation, and Erdas Imagine v8.4 has been used for the DEM evaluation. Each of the source image pairs used for DEM creation have their own merits and demerits. The Radar images are generally processed in two different ways viz. interferometric method and stereoscopic method. An analysis of the literature survey shows that the interferometric method yields better results in a relatively low relief area but not so in moderate relief areas. Conversely the stereoscopic method gives better results for moderate relief areas falling in flat topographic terrains.
In addition the beam mode selection too plays a key role in the output DEM quality. A larger beam mode separation for the stereo pair (say S2-S5) results into a good parallax. The automated pixel matching method adopted by the software, however fails when the parallax is too high. A low beam mode separation (say S2-S3), on the other hand results in errant values in low relief areas. The problem is further compounded by the presence of speckles.
Alternatively hand stereo pairs obtained by satellites, operating in optical mode yield more consistent results over different kinds of terrains. The only method of processing these images is the stereoscopic method of measuring parallax by pixel matching.
Studies reveal that of the three different pairs processed, the ASTER images yield the best results. This is possible because the temporal difference between the stereo pair is only a few seconds, whereas pairs acquired through other satellites require a larger time cycle. Results also show that raster match is good in case of ASTER images due to short temporal differences. Further the unique fixed b/h ratio yields more consistent and better results over various terrains.
Introduction
Creating Digital Elevation Model (DEM) by digitizing contour lines from topographic maps or through stereoscopic semi automated methods from aerial photographs are proven methods. However, DEM generation from satellite stereo image pairs of optical and microwave sensors, is still not a common practice. The DEM generated from satellite stereo pairs have some significant advantages over other sources, viz:
- World wide availability of satellite data without any restriction (often available as archived data) as against restricted and non availability of topographical maps and aerial photographs
- Large area coverage per scene
- Moderately high resolution
- Faster processing through sophisticated software and little manual effort
- Low processing cost
- All weather and day/night image acquisition capabilities (in case of microwave sensors)
The present work has been undertaken:
- In order to evaluate the possibility of using satellite DEMs as an alternative source to conventional methods
- To assess the accuracies of such DEM’s using different input GCP (Ground Control Points) sources, like different scale paper maps, USGS DEM etc. The different satellite images used for this study include:
- IRS 1C
- SPOT
- ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) VNIR bands
- Radarsat 1, standard beam mode data (SGF)
Inputs details
* ASTER satellite generates an along track stereo pair where, one image is captured in Nadir view and the other is captured 4 seconds later by a backward looking camera by viewing the same geographic area as seen by the Nadir view camera
** These satellites use across track viewing to generate a stereo pair, i.e., the same geographic area is viewed from two different tracks by tilting the cameras
Software used
- PCI Geometica OrthoEngine v8.2 and
- Erdas Imagine v8.4
GCPs and stereo GCPs are collected from reference maps in order to reference the images to ground. Further, TPs (Tie points) are collected to improve matching between the two stereo pairs. The following table gives the number of GCPs and Tie Points used to create a DEM:
| Satellite | Number of GCPs | Number of Stereo GCPs | Number of TPs |
| Radarsat | 53 | 35 | 65 |
| ASTER | 25 | 20 | 30 |
| IRS-IC | 12 | 10 | 20 |
| SPOT | 35 | 22 | 38 |
Once registered to the same ground area, any along track (in case of SPOT, IRS and Radarsat images) or across track (in case of ASTER images) positional differences are assumed to be due to parallax, which results due to relief. Measured parallax differences on a pixel-by-pixel basis are converted to absolute elevations using trigonometric functions and the orbital data (orbital position, altitude, attitude and the scene center). The computation relies on the inherent parallax between stereo images. An automated image correlation algorithm [1] is used to derive elevations from the parallax, by a set of well located GCPs and tie points (TPs).
The image matching system operates on a reference and a search window. For each position in the search window, a match value is computed from gray level values in the reference window. The match value is computed with the mean normalized cross-correlation coefficient and the sum of mean normalized absolute difference [2]. The correlation window size varies from low resolution (8 pixels) to 32 pixels at the full resolution. Elevation points are extracted at every pixel for the complete stereo pair. The 3-D intersection is performed using the above computed geometric model [3] to convert the pixel coordinates in both images determined in the image matching of the stereo pair to the three dimensional data [4].
The output elevations are not computed for the pixels where the image matching fails to find the corresponding pixel in the reference image, resulting into some failure areas. In case of small and scattered failures the software does interpolate and compute most probable values for them. The DEM thus generated is in raw state and does not contain geo-referencing information. So, the DEM needs to be georeferenced.
Workflow
Accuracy assessment
The accuracy assessment has been done by using the best reference DEMs (USGS, Toposheets etc.) available. The analysis of error in the output DEM can be done by the following two methods:
- Comparison of every pixel in the raster with the reference DEM at same resolution
- Random checks at the spot height points given in toposheets
The accuracy of the DEM has been expressed as linear error (LE) and circular error (CE). Linear error represents the vertical accuracy while circular error is for horizontal positional accuracy. Linear error gives an estimate of the overall accuracy of the DEM at the said confidence level (it is assumed that X, Y are correct and a comparison in Z is performed). An LE90 of 25 implies that 90% of the pixel within the DEM varies from the reference DEM by 25m or less. Circular Error (CE) is used to measure the combined errors in X and Y of the final DEM. This is a circular radius in meters, which will include 90% of the positional errors of the final DEM versus the reference DEM. (in other words, if 100 shots have been fired at a target, how large a circle should be drawn in order to accommodate 90 shots around bull’s eye). The radius of that circle is the CE 90 for the final DEM.
A model has been created in Erdas imagine software to compute the pixel level elevation differences both for negative and positive values as positive number.
Results & discussion
Several combinations of GCPs and TPs have been attempted during the process of DEM generation but only the best results obtained are being discussed here.
The RADARSAT derived DEM
The minimum elevation in the output Radarsat DEM (Fig. 1) was –80m while the maximum was 3580m. This elevation range shows that this area is a rugged terrain with high relief. The output Radarsat DEM was validated with respect to the USGS 30m DEM. The vertical accuracy LE80 (failed area constituting 15% was not used for error computation) was 65m with positional accuracy CE90 at 25m and standard deviation in overall error of 45m. The failure areas are restricted either in high escarpment areas or in the mountainous terrain. The reason for such failures appears to be principally due to layover and shadow effects of such terrains.
Accuracy of the Radarsat DEM can be further improved by using an additional image for the same area. For example in high relief areas the Layover and shadow effect is more pronounced in the used stereo pairs (S1 and S6), as the difference in Local Incidence angle is very high and which, probably accounts for DEM failure in these areas. To get a good DEM in all such areas, along with S1 image, we may use another image of S2/S3 beam mode that have close incidence angles in order to get a better DEM.
Fig. 1: A) S1 beam mode, Ascending , B) S6 beam mode, Ascending, C) DEM at half resolution (25m) and D) DEM (Thematic at 50m interval).
The following graph shows horizontal distance in meters (X-axis) for a selected cross section and Y-axis shows the elevations derived from Radarsat image and USGS DEM elevations at same points. It may be observed that both of the lines maintain the same pattern with the elevations on moderate slopes exhibiting very close match, whereas on hilltops the difference increases to about 80m higher than actual.
The Radar images are processed in two different ways viz. interferometric method and stereoscopic method. Interferometric method yields better results in a relatively low relief area but not so in moderate relief areas, whereas, the stereoscopic method gives better results for moderate relief areas and failing in flat terrains [3].
Alternatively, stereo pairs obtained by satellites, operating in optical mode yield more consistent results over different kinds of terrains. The only method of processing these images is the stereoscopic method of measuring parallax by pixel matching.
The SPOT derived DEM
The DEM extracted from SPOT stereo images (Fig: 2) contain the elevation range of 1m to 1722m. The DEM created at full resolution (i.e., 10m) has resulted into large areas of failure and hence the same has been regenerated at half resolution (20m), which has shown significant improvement in the output with less than 5% failure area. The resulting DEM was validated with USGS DEM of 30m resolution. At 32m was the value at LE90, and CE90 was at 19.5m having a standard deviation also of 32m.
Fig. 2: A) SPOT PAN Left Image (Clip), B) SPOT PAN Right Image (Clip), C) Digital Elevation Model at 20m resolution and D) DEM (Thematic at 50m interval).
As can be observed in the graphic presentation of results (above) that the SPOT derived DEM shows variable results in different slope/terrain conditions. Consistently low values can be seen in the diagram when compared to USGS DEM (30m resolution).
The IRS 1C derived DEM
The Dehradun area is a mixed terrain of moderate relief of the outer Himalayas and Gangetic plains. The IRS stereo pair (Fig. 3) used for this terrain yielded a minimum elevation of 418m to a maximum of 1475m. The DEM thus obtained, was validated with DEM extracted from 1:50,000 scale, Survey of India topographic maps. The LE 90 was 35m and 18m positional accuracy (CE90). The standard deviation in error this case was 34.5m. In this case also the failure area is less than 5% of the complete area.
Fig. 3: A) IRS-IC PAN Left Image (Clip), B) IRS-IC PAN Right Image (Clip), C) Digital Elevation Model at 25m resolution and D) DEM (Thematic at 50m interval)
The following distance vs elevation graph for IRS Pan derived DEM and topographic map derived DEM shows significantly close values for both of the sources whereas secondary undulations in the terrain are depicted better in the IRS derived DEM.
The ASTER derived DEM
The elevation range of the DEM obtained from the ASTER stereo images (Fig. 4) was 1000m to 3696m. The ASTER DEM was validated with the DEM extracted from contours obtained from 1:100,000 scale, Russian topographic maps. The vertical accuracy (LE90) was 30m with 20m positional accuracy (CE90). The standard deviation in the measured error w.r.t. reference DEM was 28m. The failure pixels were significantly less (less than 5%), which probably is because of the near synchronous acquisition of stereo images (difference of 4 seconds only) by the satellite.
Fig. 4: A) ASTER VNIR (nadir), Clip, B) ASTER VNIR (backward looking) Clip, C) Digital Elevation Model at 15m resolution and D) DEM (Thematic at 50m interval)
The following plot of distance (X-axis) vs elevation (Y-axis) also exhibits a good agreement of the ASTER DEM with topographic map derived DEM with minor variation in positional accuracy.
The following points have been important, while creating good DEM:
- Collecting most of the GCPs in low altitude areas and very less (1 or 2 only) on maximum elevation. The reason being high altitude areas have lean and their image position does not reflect true ground position and hence X-Y references taken for such points results into large RMS error
- Collection of GCPs on few hilltops (especially on the maximum elevation point in the area) is a must in order to define the full range of elevation in the output DEM. Otherwise, the high elevation areas yield large errors (>150m) due to lack of calibration
- The GCPs must be well spread over the entire area
- Large number of TPs help in getting better oriented epipolar images, thus the enhancing the chances of good raster match and hence good DEM with less failure areas
- Running a second DEM process after modifying some GCPs and adding few more TPs (if required) after random evaluation first round of DEM generation helps in improving the accuracy and containing the failure areas
*In the case of Radarsat LE 80 has been considered
Conclusion
Good quality DEMs can be created from different satellite image stereo pairs. The number and distribution of GCPs and TPs play a crucial role in the final accuracy of the DEM.
Layover and shadow effects in used Radar images were pronounced, rendering it a not so viable option for high relief areas. However, judicious selection of beam mode combination and by using more than one pair would give better results.
Study reveals that, of the three different pairs of optical sensor data processed, the ASTER images yield the best results. The elevation range and relief covered by ASTER image in this sample study was also significantly large (2700m approx.). The better results appear to be due to the negligibly small temporal difference between the stereo pair (only four seconds), whereas pairs acquired through other satellites in optical mode have comparatively large temporal variation. Further, the unique fixed b/h ratio yields more consistent and better results over various terrains. The net cost of DEM derived from ASTER images also is very less as compared to IRS and SPOT images, which cost nearly three times more than ASTER.
Since the output DEM reflects a pixel level true picture of the terrain, the DEMs thus created can be used for various application areas of GIS like Telecom, Watershed delineation, 3-D analysis etc.
References
- Thierry Toutin, “Generating DEM from Stereo-Images with a Photogrammetric Approach: Examples with VIR and SAR data,” EARSeL Advances in Remote Sensing, Vol. 4, No. 2, p. 110-117, 1995.
- Marty Marra, Kelly E. Maurice, Dennis C. Ghiglia, Heinrich G. Frick,” Automated DEM Extraction Using RADARSAT ScanSAR Stereo Data”, From Web
- Enrico C. Pedroso, Alvaro P. Crósta and Carlos R. de Souza F, “Comparative Analysis of RADARSAT and JERS-1/SAR DATA for Geological Mapping and Gold Exploration in The Tapajós Region, Brazilian Amazon”, From Web.
- Thierry Toutin and Ph. Cheng, “DEM Generation with ASTER Stereo Data”, Earth Observation Magazine, Vol. 10, No. 6, pp. 10-13, June 2001.
Thierry Toutin, “3D Topographic Mapping with ASTER Stereo Data in Rugged Topography”, Personal communication to the authors, 2001.
PCI Geomatics, “OrthoEngine SE 3D, v8.2” in Reference Manual, Version 8.2 (October19, 2001).