high level of noise in either the source or the reference images.
For the images where the method was successful in more than
90 percent of the scenes, they are within sub-pixel accuracy.
Robustness
One of the advantages of this paper’s method as compared
to others is the requirement of a fewer numbers of good
GCP
s
in a pool of checkpoints where there could be quite a few
erroneous points. The causes for errors as stated earlier could
be due to errors in the reference data or due to mismatches
in automated checkpoint identification process, which are
caused due to the presence of large homogeneous patches in
the data or the presence of cloud patches. Hence, to prove the
robustness of our methodology, images corresponding to small
islands in the Indian Ocean region were chosen. The land por-
tion was significantly less as compared to the scene size, and
the scene contained no significant land or man-made features.
Table 4 describes the scene characteristics indicating the
percentage of land covered with respect to complete scene,
and the number of checkpoints identified by
SIFT
. A plot
showing the
RMSE
in meters is depicted in Figure 10. IRS
Cartosat-1 imagery was the target image and
LDCM
/
OLI
with
15 meters spatial resolution image was used as a reference
image. Figure 9 shows the overview of some of the scenes
tested. Each image shows the complete scene and the high-
lighted bounding box shows land mass. The summary of this
experiment is that the process worked satisfactorily on all the
difficult data sets with a similar level of product accuracy,
proving the robustness of the methodology to operate in an
autonomous mode.
Implementation
The software for the entire process of automatic ortho-recti-
fication is implemented in
ANSI
C/C++ on a Linux platform.
Open source libraries OpenCV and GALib (Wall, 1992) were
used and the
RSM
for IRS Cartosat provided by Radhadevi
et
al.
(2008) and Srinivasan
et al.
(2009) are used for product
generation, and the software is integrated with the necessary
customizations of GALib and OpenCV libraries. Although the
application is developed and tested on system configuration
with specifications 2 × 2.0
GHz
Intel Xeon E2560, 16
GB
RAM
, 1
Figure 10. Corresponding RMSE in meters.
T
able
4. R
esults with
M
inimal
L
and
C
over
D
ata
S
ets
S. No
Percentage of land covered
w.r.t full scene
No. of Checkpoints generated
by SIFT for GCPs
1
1
20
2
20
221
3
3
42
4
13
62
5
33
15
6
10
37
7
9
68
8
7
20
9
7
20
10
8
3
Figure 8. RMSE in meters for different sensors.
Figure 9. Sample scenes corresponds to the scene indicated in Table 4 (land coverage is highlighted in box).
386
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
