Methods for Assessing the Quality of A Digital Elevation Model Generated using Full-Scale Images
Dr. S.A.B. Kim* Dr.wonyou Park , Prof. Tag-Gon Kim**
*Satellite Technology Research Center,
**Dept. of Electrical Engineering,
Korea Advanced Institute of Science and Technology,
373-1 Kusung, Yusung, Taejeon, S. Korea 305-701.
Tel: +82-42-869-8634. Fax: +82-42-861-0064.
Email:-sbkim@krsc.kaist.ac.kr
Keywords digital elevation model, accuracy assessment, geodetic datum
Abstract The most viable way to assess the quality of a digital elevation model, (DEM) generated from satellite images would be to compare with digitized map. Complexity in quality assessment (QA) arises due to difference in geodetic datum: the most accurate DEM from satellite images requires WGS84 (World Geodetic System) datum whereas the digitized maps are given on a local datum (Tokyo Mean datum for Korea). This complexity has not been addressed so far. How much the datum difference may impact QA results is investigated in this work. The test region is 30 km by 40 km area in Korea, and a DEM is generated from 10-mresolution SPOT satellite images. QA is performed, first, on WGS84 datum. Second, QA is performed on Tokyo Mean datum and the difference between the two QA results are analysis. The rootmean-square (RMS) of the differences over the entire test site is merely 4.5 cm (252,000 grid cells). There are, however, spots with significant difference. At 24 grid cells the difference is > 0.5 m. The RMS of their difference is 4.6 m. Correlation analysis supports that steeper terrain and the horizontal distance between the two datum lead to greater difference between the two QA methods. Since most countries use non-WGS84 local datum, the results from this study can be applied universally.
Introduction
A digital elevation model (DEM) is the most effective medium for terrain analysis and constructing geographical information system on a large scale (e.g., Moore et al., 1991). An essential step in DEM production is the assessment of DEM quality. A coomonly-used method for such assessment is to compare with digitized maps. GPS (global positioning system) measurement, an alternative, suffers from insufficient sampling of terrain though by far more accurate than digitized maps.; Complexity in the quality assessment (QA) arises due to difference in geodetic datum. That I, when satellite images are used, the datum of the most accurate DEM would be WGS84 (World Geodetic System). This is because such DEM would require ground control points to be obtained by GPS measurements and GSS84 is the datum of GPS records. Whereas the digitized maps are given ona local datum (Tokyo Mean datum for Korea). This complexity has not been addressed so far. Nor has been the issue of how
much the datum difference may impact QA result, thus the issue is investigated in this work.
Limitation of using GPS Recores as Truth Data
It is demonstrated the GPS records are not appropriate as truth, due to sampling deficiency of the measurements. The sampling deficiency issue was hinted by previous authors (Sasowsky et al., 1992) but not in the context of comparisons using GPS records, nor in a quantitative manner. To demonstrate this, three sets of differential GPS measurements are compared with 3"-DTED (digital terrain elevation data) over S. Korea (NIMA, 1999). The GPS data are: (a) 31 records measured by Korean National Geographic Institute (NGI). These are evenly distributed over S. Korea, thus would be less biased to any particular terrain type (b) 161 records measured using a Trimble receiver over about 100 km by 200 km region (c) the same method as in (b), but 25 records over about 60 km by 60 km region.
The difference between DTED and GPS over the three regions are -1 m, -17 m and 20 m (Table 1). Closer examination reveals that the magnitude of the mean error is due more to the sampling deficiency than to the actual difference between GPS and DTED (Kim et al., in preparation). Another way of explaining the sampling deficiency in GPS measurement is by the Nykist criterion. The characteristic wavelength of a terrain is about 300 m, after the variogram analysis by Collins and Woodcock (1999) of 30-m resolution Landsat image over 24 by 24 km forest region in Oregon. The GPS records should has a horizontal resolution of O(2), which is impractical. Since the only feasible source with sufficient sampling is the digitization of a conventional map.
Table 1. Comparison of GPS DTED over S. Korea.
| Area |
GPS - (DTED+geoid(m) |
| Mean |
SD (m) |
| A |
1.3 |
6.5 |
| B |
-17.3 |
11.0 |
| C |
19.6 |
9.3 |
