The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
437
The nearest neighbor, bilinear, cubic convolution, and sine
interpolators for resampling the slave image were all
investigated. The 2D sine lengths include both odd and even
numbers, varying from 2 to 8. Both sine add-ons: modulation
and Hann window were demonstrated and applied.
The coregistration performance was evaluated against the
coherence of the two coregistered SAR images and the
accuracy of the InSAR DEM. Our own program was developed
to compute coherence. The results were compared to the
coherence as computed by commercial software. The InSAR
DEM was generated through commercial software.
4. RESULTS AND ANALYSIS
The average coherences computed from the master SAR image
and the resampled slave SAR images were recorded into tables.
The coherence comparison and analysis were conducted side by
side.
4.1 Coherences of Original and Coarse Coregistered SAR
Images
Searching Subpixel Tie Points and
Offsets
# of
Pars
Interpola
tor
Coherence
Method
Rate
Kernel
North
South
Original
N/A
N/A
0
N/A
0.2554
0.2542
Coarse
N/A
N/A
2
N/A
0.3181
0.3517
Table 1 Coherences of coarse coregistered SAR images
Table 1 contains the coherences out of the original SAR images
and coarse coregistered SAR images. If the coherence image is
computed from the original master and slave images, the
average coherence is 0.2554 for northern Indiana and 0.2542
for southern Indiana. After the slave SAR image was shifted
based on coarse coregistration, the average coherence became
0.3181 for northern Indiana and 0.3517 for southern Indiana, a
little higher than original SAR images.
4.2 Coherences of Fine Coregistered SAR Images
More factors were examined in fine coregistration for
coherence evaluation. First, the approach of oversampling
cross-correlation function peak was tested. The oversampling
rate was 10. The oversampling kernel was Spline. 4-parameter
transformation equations were used. The interpolators included
Nearest, Bilinear, Cubic, and Sine with the lengths from 2 to 8.
Modulation was applied to all sine interpolators.
Searching Subpixel Tie Points and
Offsets
#of
Pars
Interpolat
or
Coherence
Method
Rate
Kernel
North
South
Fit Peak
10
Spline
4
Nearest
0.4297
0.3980
Fit Peak
10
Spline
4
Bilinear
0.4390
0.4596
Fit Peak
10
Spline
4
Cubic
0.4450
0.4461
Fit Peak
10
Spline
4
Sinc2x2
0.4482
0.4177
Fit Peak
10
Spline
4
Sinc3x3
0.4500
0.4201
Fit Peak
10
Spline
4
Sinc4x4
0.4507
0.4217
Fit Peak
10
Spline
4
Sinc5x5
0.4506
0.4218
Fit Peak
10
Spline
4
Sinc6x6
0.4500
0.4213
Fit Peak
10
Spline
4
Sinc7x7
0.4499
0.4213
Fit Peak
10
Spline
4
Sinc8x8
0.4498
0.4211
Table 2 Coherences of oversampling cross-correlation function
After fine coregistration, the coherence increased significantly,
even with only nearest neighbor interpolator. The coherence of
nearest neighbor is 0.4297 for northern Indiana, much better
than the coherence out of coarse coregistration (0.3181). The
coherence of nearest neighbor is 0.3980 for southern Indiana,
also better than the coherence obtained from coarse
coregistration (0.3517). Bilinear yields higher coherence than
nearest neighbor for both northern and southern Indiana
(0.4390>0.4297 and 0.4596>0.3980). Cubic yields slightly
higher coherence than bilinear (0.4450>0.4390) for northern
Indiana, but slightly lower coherence than bilinear
(0.4461 <0.4596) for southern Indiana, so there is not much
difference in performance between bilinear and cubic in
interpolation.
Theoretically, sine interpolation should have higher coherence
than cubic, bilinear, and nearest neighbor methods; and the
longer sine length should have the higher coherence, but those
are not always true. For northern Indiana, all sine interpolations,
even the shortest one, 2-point sine (0.4482), have higher
coherence than cubic and bilinear. The longer length has
slightly higher coherence. After Sinc4x4 (0.4507), the
coherence is reduced as the sine length is increased, i.e. sinc4x4
seems to be the best interpolator in this table. If sine length is
longer than 8, the coherence starts to fluctuate, although the
overall trend is rising. For southern Indiana, the coherences of
bilinear (0.4596) and cubic (0.4461) interpolation are
significantly higher than all sine interpolation with bilinear
providing the best results. The trend of sine interpolation is the
same as northern Indiana, though the length of the highest sine
coherence is 5.
Table 3 shows the comparison between oversampling the cross
correlation function by 10 times and by 100 times.
Searching Subpixel Tie Points and
Offsets
#of
Pars
Interpolat
or
Coherence
Method
Rate
Kernel
North
South
Fit Peak
10
Spline
4
Sinc4x4
0.4507
0.4217
Fit Peak
100
Spline
4
Sinc4x4
0.4507
0.4217
Table 3 Coherences of different oversampling rates
100 times oversampling rate does not have higher coherence
than 10 times oversampling rate. The results confirm the
average random phase error could be much bigger than 1/100 of
a wavelength and the higher oversampling rate may not benefit
significantly.
Table 4 is the comparison among different number of
parameters.
Searching Subpixel Tie Points and
Offsets
#of
Pars
Interpola
tor
Coherence
Method
Rate
Kernel
North
South
Fit Peak
10
Spline
4
Sinc4x4
0.4507
0.4217
Fit Peak
10
Spline
6
Sinc4x4
0.4514
0.4217
Fit Peak
10
Spline
12
Sinc4x4
0.4511
0.4218
Table 4 Coherences of different number of parameters
From Table 4, one can find the number of parameters has no
significant effect on the coherence. 4-parameter has been
sufficient. 6 parameter transformation equations can model the
small deformation with respect to azimuth pixels. The 2 nd
ranked 12 parameter transformation equations could model
more distortion for SAR images, although that is not necessary
for this ERS SAR data.