International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
2.2.3 Patch Size Optimal Backprojection
To reduce the interpolation error when terrain variation is
considered, it would be better that the patch size changes
according to the terrain characteristics. The patch size should
be large for rolling terrain and, on the contrary, small for
rugged one. Quadtree structure (Mather, 1999) that segments
the terrain coverage with a given elevation range is a
straightforward and yet effective way to characterize the terrain
variations. Based on this consideration, we select quadtree
structure to optimal the window size in the patch backprojection.
The method discussed in this section is a modified one stated in
the previous section. In the quadtree analysis, the allowed
elevation range is selected according to that the model error
will not exceed the tolerance. The analysis of the model errors
will be given in the next section. After performing the terrain
subdividing, tiles with different sizes are projected into the
image space one by one using the process described in the
previous section. Figure 3 illustrates the concept of the patch
size optimal backprojection.
Figure 3. Illustration of patch size optimal backprojection
(a) Optimal tiles size by quadtree structure
(b) Anchor point generation for the top layer
(c) Anchor point generation for the bottom layer
(d) Interpolation
2.3 Model errors estimation
Patch size is crucial when doing computation. The larger the
patch size, the faster the computation time, but the model error
increases as well. Hence, patch size optimal should consider
the model error of the proposed method. Two steps, including
affine transformation for anchor point sets and linear
interpolation in patch backprojection could introduce model
errors. Considering the imaging geometry, patch size, tilt angle,
and elevation range are the three most important factors to be
considered in the analysis of model errors. Simulations are
employed on SPOTS supermode images. The model error is
defined as the difference of the proposed method and the
rigorous point-by-point backprojection.
The evaluation items and parameters are shown in Table 1. The
evaluation items includes: transformation model error, variation
of terrain model error and tilt angle model error. The tilt angle
is within 30?. The elevation ranges from 0m to 2000m. The
patch size is within 40m by 40 m and 3000m by 3000m for each
evaluation item.
Table 1. SPOTS model error analysis
Transformation Interpolation
Model Error Model Error
Evaluation Item Affine Variation of | Tilt angle
Variable transformation terrain
Tilt angle (deg 30 30 1-30
Elevation (m) 0 and 2000 0~2000"" 0 and 1000
Patch size (m * m) |80*80- 80*80~ 80*80~
2000*2000"* 2000*2000"' |2000*2000'"
"Tilt angle step (deg): 1, 2, 4, 8, 16, 20, 25, and 30.
"Elevation step (m): 80, 160, 320, 500,640, 750, 1000, and 2000.
"Patch size step (m*m): 80, 160, 320, 640, 1280, 1500, and 2000,
588
2.3.1 Affine transformation model error
Given a tilt angle of 30? and elevation in 0m and 2000m, the
errors due to affine transformation are shown in Figure 4. It
indicates that when the patch size of 1500pixel by 1500pixel is
selected, the model error is within the tolerance, i.e., 0.05 pixels
in this investigation.
,
Transformation model error
0.05
Model Error(pixel)
0.00 eet
0 500 1000 1500 2000
Patch Size(m*m)
Figure 4. Transformation modcl error for SPOTS
2.3.2 Variation of terrain model error
Given a tilt angle of 30°, the errors due to terrain variation are
shown in Figure 5. It indicates that when the variation of
terrain is less than 1000m and patch size is less than 640m by
640m, the model error is within the tolerance, i.e., 0.05 pixels,
in this investigation.
Model Error for Variantion of Terrain
0.25 Patch Size
(m*m)
= 020 >
e
>
a. | ? 160
-— 0.15
S 320
=
LL ne —610
5 À
=
o —8—- 1280
z
0.05 , — 1500
x —G-—- 2000
0.00 SK :
0 500 1000 1500 2000
Variations of Terrain (meter)
Figure 5. SPOTS model error for variation of terrain
Intern
2.3.3
Given
angle
size iS
than 0
rror (Pixel)
Model E
The te
GCPs
scale t
centim
Figure
repres«
DTMs
The re