approximately 16 kms × 10kms, and corresponds to a city
region, having 862 m of minimum height; 937 m of maximum
height, and average height is 900 m. The
LDCM
/
OLI
ortho-im-
age at 15m resolution from Path 144 and Row 51 was chosen
as the reference ortho-image. A total of six
GCP
s were used
(i.e., the chromosome length for
GA
), and
GA
was run for 20
generations with population size of 10 in each generation.
SIFT
feature-based matching at intermediate resolution of 5
m was used to estimate the transformation coefficients using
scale-space parameters for
SIFT
, i.e., (a) number of octave (
n
oct
= 4), (b) number of scales per octave (
n
spo
= 3), (c) sampling
distance (
σ
min
= 0.5), (d) the level of blur (
σ
min
= 0.8), and (e)
assumed level of blur (
σ
in
= 0.5). Sample matched points ob-
tained in Stage1 of
GCP
identification are depicted in Plate 1a
using
LDCM
/
OLI
as reference and Cartosat-2A at 5 m resolution
as source image. After estimating the transformation coef-
ficients the matching was performed tile-wise in the original
resolution, i.e., 0.8 m. Matched points in one such tile are
depicted in Plate 1b (Stage 2). The layout of all the points
and the
GA
-based selected
GCP
s along with its convex hull is
shown in Plate 1c. The corresponding model error was 4.5 m
in latitude direction and 7.75 m in longitude direction. After
refining
RSM
, an ortho-image was generated with 0.8 m output
resolution and Automatic Product Evaluation (
APE
) using
SIFT
-
based matching was performed between the same reference
ortho-image (which is used for checkpoint identification) and
the obtained Cartosat-2A ortho-product. A total of 48 conju-
gate points were generated. Note that the points generated
through
APE
are new set of tie points and are different from
the checkpoints identified for
GCP
selection. Plate 1e shows
a few of the tie points of
APE
process, and Plate 1f shows one
of the tie points in original resolution. The Table in Plate 1d
shows the final
RMSE
of the ortho-product generated, which
was 9.8 m in latitude and 10.8 m in longitude. The same data
set was subjected to ortho-rectification using GoogleMap
extracted images as reference data, where the resolution of the
reference image was 1 m. A total of 90 checkpoints were ob-
tained and the same process as discussed above was repeated.
The product accuracy obtained was 2.2 m and 4 m in latitude
and longitude, respectively. Similarly, the experiments were
conducted on various data sets as shown in Table 2 with dif-
ferent types of scenes and sensors. The summary of the above
experiments is that the proposed methodology resulted in
relative accuracy of less than a pixel of the reference image
resolution.
Reliability
More than 200 data sets covering various regions in the
Indian sub-continent acquired between January 2007 and
December 2014, were executed in the autonomous mode for
computing system reliability. Among these data sets; approxi-
mately 48 belonged to Cartosat-2/2A/2B, and 156 data sets
were of Cartosat-1. For all these data
LDCM
/
OLI
band with 15
m spatial resolution was used as the reference data sets. The
experimental results are shown in Figure 8 and the summary
is tabulated in Table 3. As can be seen, excluding very few of
the scenes, most of the images resulted in less than 15 m er-
ror, which is the sub-pixel resolution of reference image used.
The success percentage for Cartosat-1 was around 95
percent, while for Carto-2/2A/2B it was around 75 percent.
Totally, about 20 out of 200 scenes could not be processed,
which upon careful inspection revealed that the scenes had
significant changes in the terrain features as compared to ref-
erence image (due to different vintages) and also contained a
Plate 1. Results of Experiment-1: Source Image from Cartosat-2A and ETM Ortho image as the Reference.
T
able
3. S
ummary
of
R
esults
O
btained
from
A
ll
the
S
cenes
Sensor
System
Total
Scenes
Mean checkpoints/
scene
Mean GCPs
per scene
RMSE
Lat [m]
RMSE
Long [m]
Scenes successfully
processed
Success
rate (%)
Within one pixel
error of reference
C-2A/B
48
126
6
10
16
35
75
32 (92%)
C-1
156
178
6
14.47
1.53
148
95
143(97%)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
385
