Automatic DEMs & Orthoimages from SPOT Stereo-Pairs.
G. Petrie & N. Al-Rousan
It was very interesting for us to read the article by Dr. Krupnik on the tests that he has carried out on the automated extraction of DEM data from SPOT stereo-pairs using commercially available packages - as reported in the May issue of GIM. In parallel with his work, a similar set of accuracy tests have been carried out by us using two quite different commercial packages - PCI EASI/PACE and R-WEL DMS - as well as the ERDAS OrthoMAX package used by him. Our tests therefore complement and extend the information given by Dr. Krupnik.
The tests have been carried out over an area of hilly and stony desert (called the Badia) in North East Jordan. Thus the area has some similarities to parts of the BeerSheva area in Israel used by Dr. Krupnik. The Badia test area is however much larger covering 12,000 sq. km. with an elevation range of 1,500m across the area. Much of the terrain surface is covered by volcanic lava flows and salt pans with only a very small area of marginal agriculture. The area is covered by five SPOT Level 1B Pan stereo-pairs having excellent base:height values between 0.86 to 0.98.
Reference Data Sets
The reference data sets were supplied by the Royal Jordanian Geographic Centre (RJGC) and comprised the following data.
(a) A network of 130 ground control points (GCPs) was established using Ashtech dual-frequency GPS sets and employing differential GPS methods. This network covered the area of the five SPOT stereo-pairs used in the tests. 60 of these GCPs were located in a single reference stereo-model. The remaining 70 points were distributed as evenly as possible over the other four stereo-models. The coordinate values of all the points were available in the WGS84, geographic, UTM and Jordan TM systems. The accuracy of these GCPs is estimated to be s E = s N = s H = ±1m.
(b) 150km of accurate elevation profiles were also established using kinematic GPS techniques with the roving GPS mounted in a Land Rover vehicle. These profiles were located along two old roads that cross the area. They provided 15,000 height points that could be used to validate the DEM elevation data.
(c) Digitized versions of the 1:50,000 and 1:250,000 scale contoured topographic maps of the area that had been compiled photogrammetrically from aerial photographs at 1:40,000 scale were also supplied. The measured contour data provided another data set that could be used to validate the DEM elevation data.
The EASI/PACE and DMS packages were run on PCs: the OrthoMAX package was run on a Sun workstation. Both EASI/PACE and OrthoMAX employ a somewhat similar solution. This is based on an initial modelling of the satellite orbital track based on the ephemeris data supplied as a header file along with the SPOT image data. This provides the initial 3D coordinates of the perspective centres for each individual line of the SPOT image, together with the corresponding attitude (tilt) data. The photogrammetric solution then employs classical analytical methods based on the use of 3D collinearity equations - albeit within the constraints imposed by the SPOT linear array geometry. By contrast, DMS first uses a polynomial procedure to rectify and fit the individual SPOT images planimetrically to the GCPs in the terrain coordinate system. This is followed by the measurement and computation of the elevation values using quite simple parallax formulae.
Geometric Accuracy Tests
The first series of tests concerned the geometric accuracies achieved after the absolute orientation of the SPOT Level 1B stereo-pairs to fit the GCPs located in the reference model using manual/visual measurements. The results are given in terms of RMSE values in plan and height in Table I. Similar results were achieved with the other four stereo-models using smaller numbers of GCPs. The results, especially in height, are very good for the EASI/PACE and DMS packages. The results in height from OrthoMAX marked with an asterisk (*) are almost certainly erroneous. However, as will be seen later, the origin of this error appears to be purely statistical - since the DEM heights obtained from the image matching operation carried out with OrthoMAX after the orientation gave the expected results.
System Control Points Check Points
No. Plan (m) Ht. (m) No. Plan (m) Ht. (m)
EASI/PACE 45 ±6.8 ±4.7 - - -
23 ±6.9 ±4.5 25 ±8.0 ±5.8
OrthoMAX 49 ±14.7 ±1.0* - - -
39 ±7.9 ±1.3* - - -
DMS 45 ±11.5 ±4.5 - - -
20 ±12.8 ±4.5 25 ±11.9 ±6.2
Table I - Results after Absolute Orientation
After absolute orientation, the DEMs were extracted from the SPOT stereo-pairs utilizing the automatic image matching procedures that were available on all three systems. All are based on the use of area-based correlation techniques
Comparison - DEM Values v. GCP Values
The first test of the DEM values involved a comparison of the given elevation values of the GCPs with the corresponding height values generated by the image matching procedures that had been used to create the DEM. Again the results are given in Table II in terms of the RMSE values in elevation for the reference stereo-model only.
System Control Pts. Check Pts.
No. Ht. (m) No. Ht. (m)
EASI/PACE 47 ±3.7 - -
22 ±3.8 25 ±3.7
OrthoMAX 39 ±1.0* 25 ±5.6
DMS 45 ±4.5 - -
19 ±4.5 25 ±6.2
Table II - DEM Values v. GCP Values
Comparison - DEM Values v. GPS Profile Values
This has involved the comparison of a much larger sample of the elevation values from the DEMs with the corresponding height values of the same points given by the GPS profiles. The results are given in Table III in terms of standard deviation values. It will be noted that all the GPS height values were too high by circa 22 to 24m due to an overall datum error.
System No. of Mean St. Dev. (m)
EASI/PACE 1,248 24.0 ±6.1
528 24.0 ±6.0
OrthoMAX 528 22.4 ±8.4
DMS 1,248 25.6 ±9.3
528 15.6 ±9.0
Table III - DEM Values v GPS Profile Values
As one would expect, the results achieved in this particular comparison using the values obtained along the elevation profiles measured using kinematic GPS were somewhat less good than those obtained at the more restricted number of GCPs where the reference height values had been obtained using the GPS sets measuring in a static mode.
Comparison - DEM Values v. Contour Values
A comparison has been made between the DEM data and the digitized contours from the 1:50,000 scale maps that had been produced by aerial photogrammetric methods. The contours with a 10m interval were superimposed over the DEM produced from the SPOT stereo-pairs. Then the respective elevation values were compared for selected contours. The results are given as RMSE values in Table IV.
System No. of Pts. Ht. (m)
EASI/PACE 257 ±4.9
OrthoMAX 325 ±8.9
DMS 220 ±8.1
Table IV - DEM Values v 10m Contour Values
A similar test was carried out using the digitized contours at the 50m interval derived from the 1:250,000 scale maps covering the test area. Again the comparison was made with the DEM elevation values derived from the SPOT stereo-pairs. It will be seen from Table V that, somewhat surprisingly, the RMSE values are only slightly less good than those obtained with the 1:50,000 scale contours - in spite of the very large differences in scale between the maps.
System No. of Pts. Ht. (m)
EASI/PACE 719 ±7.0
OrthoMAX 712 ±8.9
DMS 589 ±9.9
Table V - DEM Values v 50m Contour Values
Comparison with Dr. Krupnik's Results
Overall the results achieved with the EASI/PACE, OrthoMAX and DMS packages over the Badia test field with RMSE values of ±6 to 9m in elevation lie in a similar range to those achieved by Dr. Krupnik for desert and mountain areas. These values are consistent with contouring at 20 to 30m intervals.
Comparison of Contours
The contours derived from the DEMs were superimposed on the contours digitized from the 1:250,000 scale maps. They showed an excellent agreement for all three packages that were tested.
Tests of the Orthoimage Data
Orthoimages with a 20m pixel size were created for the whole of the test area using the DEMs generated by each of the packages from the SPOT data. The complete orthoimage mosaic amounted to 56 Mbytes of data. An independent check of the planimetric accuracy of the orthoimage of the reference stereo-model was then made using a simple linear transformation of the image coordinates measured at the GCPs and comparing the results with the known planimetric coordinate values of the GCPs. The results are given in Table VI. These gave RMSE values in planimetry of ±12m for EASI/PACE; ±16m for DMS; and ±18m for OrthoMAX.
System No. of Pts. Plan (m) Plan (pixels)
EASI/PACE 47 ±12.3 ±0.6
OrthoMAX 43 ±17.7 ±0.9
DMS 45 ±16,0 ±0.8
Table VI - Planimetric Accuracy at the GCPs
From this test, it will be seen that the orthoimages generated from all three packages would meet the planimetric accuracy specifications for 1:50,000 scale mapping - where the standard error of ±0.3mm is equivalent to 15m on the ground. The radiometric quality of all the orthoimages was excellent and they all merged smoothly to produce the final orthoimage mosaic of the Badia area.
Analysis and Conclusions
As in the case of Dr. Krupnik's tests, the series of tests reported in this article show that useful results can be obtained that are compatible with the accuracy requirements of small-scale topographic mapping. These can be obtained with the latest generation of software using automated techniques for the generation of DEMs and orthoimages from SPOT stereo-pairs for desert and mountain areas with little vegetation cover. In other areas with large areas of vegetation, the results may not be so favourable. In particular, tests carried out by the authors in other areas have shown that difficulties occur in those areas where considerable changes in the appearance of the vegetation cover, cultivated areas and hydrology are experienced due to seasonal (dry/wet or summer/winter) climatic changes. If the individual images making up the stereo-pairs have been taken some months apart, then the quite different appearances of these areas will prevent image matching and the generation of the DEMs.
A large number of additional tests have been carried out over the Badia area using the SPOT stereo-pairs, besides those summarized here. They include the results of comparative testing of SPOT Level 1A and 1B stereo-pairs. Details are given in the following papers:-
Al-Rousan, N., Cheng, P., Petrie, G., Toutin, Th. and Valadan Zoej, M.J., 1997 - Automated DEM Extraction and Orthoimage Generation from SPOT Level 1B Imagery. Photogrammetric Engineering & Remote Sensing, 63 (8): 965-974; and
Al-Rousan, N. and Petrie, G., 1998 - System Calibration, Geometric Accuracy Testing and Validation of DEM and Orthoimage Data Extracted from SPOT Stereo-pairs Using Commercially Available Image Processing Systems. International Archives of Photogrammetry & Remote Sensing, 32 (4): 8-15.
Biographies of the Authors
Professor Gordon Petrie holds academic qualifications in geography (Aberdeen University),
surveying (University College London) and photogrammetry (ITC). Since 1978, he has been Professor of Topographic Science at the University of Glasgow, finally retiring in 1995. Currently he is U.K. national delegate to ISPRS Commission IV.
Dr. Naif Al-Rousan has a B.A. in geography from Alexandria University; a Diploma in photo-
interpretation from the ITC; an M.Sc. in physical geography from the University of Jordan; and a Ph.D. in photogrammetry and remote sensing from the University of Glasgow. Currently he is a lecturer at Mu'tah University near Kerak in Jordan.