Multi-Resolution approach to Radargrammetric Digital Elevation Models Generation
DEM Generation
The underlying principle for generating the DEM is to find, for each X,Y coordinates, a height H where the normalized cross correlation of two image patches, one from each images, is highest within a range of heights. The location of each image patch is computed from the refined imaging model. A problem in correlating SAR imagery is that there are usually some very dominating high intensity features amongst the majority of average intensity features. When we choose any patch, the correlation results will be biased to these strong intensity features, even though they may be at the far edges or corners of the patch. We reduce this problem by introducing a weighted block window to minimize the contribution of the edge pixels in the correlation computation. The weighted block window chosen is the Welch window [4] in two dimensions. The weighted window w
i in one
dimension of pixel i can be expressed as follows:
wi = 1 - [ ( i - N - 1) / ( N + 1 )]2 (7)
where N is the half window size in one dimension. The full window size in one dimension is 2N+1. Figure 5 shows the weighted window in one dimension with N = 19. The weighted window wij in two dimension ( i, j ) is:
wij = wi × wi (8)
Figure 5. The weighted window in one dimension with N = 19.
Hierarchical Multi-Resolution Approach
To improve the robustness of the DEM generation processing, a hierarchical multi-resolution approach (fig.6) is implemented. The original stereo pair of imagery (12.5m/pixel) were averaged over 16´16 (200m/pixel), 4´4 (50m/pixel) and 2´2 (25m/pixel) pixel-windows. A preliminary DEM is first computed from the correlation of the lowest resolution (200m/pixel) image pair. For the higher resolution computation, the DEM computed at the previous step is used to carry out a geometric correction of the image pair to enhance correlation, hence more accurate height determination. The advantage of using this hierarchical multi-resolution matching technique is that it is easier to find the correct matching of the reduced resolution images because of the removal of the speckle noise and detailed features in the block window.
Figure 6. Hierarchical multi-resolution DEM approach
Test Site and DEM Generation
The hierarchical multi-resolution DEM approach was tested with two descending RADARSAT images of S2 and S6 beam modes (fig. 7 and 8) over Brunei. The detailed image description is listed in table 1. The pair of imagery (12.5m/pixel) were averaged over 16´16 (200m/pixel), 4´4 (50m/pixel) and 2´2 (25m/pixel) pixel-windows. The lowest resolution (200 m/pixel) DEM (fig. 9) of the area was first generated from the 200m/pixel image layers. The 50m/pixel resolution DEM (fig. 10) of the same area was later generated from 50m/pixel image layers by using the terrain information of previous DEM. Finally, the highest resolution DEM (fig. 11) of 20m/pixel of the same area was generated from 25m/pixel image layers by using more detailed terrain information of previous DEM layer. An attempt to generate 10m/pixel resolution DEM by using the original imagery did not produced satisfactory results, probably because of the heavily scattered speckle noise at that resolution that causes de-correlation of the image patches.
Table 1. The description of Radarsat stereo imagery used in the DEM generation
| Stereo Pair
|
Image Format
|
Beam Mode
|
Incidence Angle (degree)
|
Orbit
|
Date
|
Image Pixel (12.5m/pixel)
|
| Image No. 1
|
SGF
|
S2
|
23.8 – 31.1
|
12110
|
28 Feb 1998
|
8953×9000
|
| Image No. 2
|
SGF
|
S6
|
41.4 – 46.7
|
11381
|
8 Jan 1998
|
8675×9000
|
Figure 7. (above) Radarsat image (S2 mode) acquired on 28 Feb 1998. © CSA 1998
|
Figure 9. DEM with 200m/pixel resolution in an area of 80×100 km2.Map projection: UTM 50N Datum: GRS 1980
|
Figure 8. (above) Radarsat image (S6 mode) acquired on 8 Jan 1998.© CSA 1998
|
Figure 10. DEM with 50m/pixel resolution in the same area.Map projection: UTM 50N Datum: GRS 1980
|
Figure 11. DEM with 20m/pixel resolution in the same area.Map projection: UTM 50N Datum: GRS 1980
|
