and replaced with mere scaling and translation as both the
images are ortho-products, which would be free from scale
and orientation variations. After the tie points are generated,
the absolute coordinates are computed using corresponding
projection transformations. Product error
p_error
is computed
as
CE90
of all differences in meters at each tie point. Similarly,
the model (
RSM
/
RFM
) error
m_error
is the error computed after
refining the model using
GA
-based
GCP
selection, which is the
RMSE
at the best 50 percent of all the checkpoints. The
p_error
is compared with model error (
m_error
), and if the error is
found to be within the limits, then the product is accepted
else the process is repeated through
GCP
collection process.
The iterations are repeated until acceptance criterion is met
or maximum number of iterations is reached. The acceptance
criterion used in our approach is “the difference between
p_
error
and
m_error
is less than the resolution of coarser image.”
The final product error is a function of three main parameters.
i.e., platform characteristics (such as stability, precision, and
uncertainty of measuring devices on board), reference da-
tabases being used (spatial resolution and errors associated
with it), and terrain type (hilly, plain, etc); hence a uniform
acceptance criterion for all cases would be unrealistic. There-
fore, rule-base is proposed based on the three sources of errors
discussed above. The rule-base could be different based on
the combination of mission, type of terrain, and reference data
used, which can be flexible so that it gets evolved.
Autonomous Mode of Ortho-Rectification
An autonomic computing model requires the system to be
operated unattended and should have features such as fault
tolerance, robustness, reliability, control loop, and switching
to safe mode of operations in case of non-compliance to the
prescribed quality. The hardware and software can be made
fault tolerant and robust by providing sufficient redundancy in
hardware, and building the software using accepted software
engineering practices. However, the other major source of er-
rors/faults for the ortho-rectification process can come from the
data itself, i.e., errors in reference data, poor quality images,
large homogeneous patches in the data, and presence of cloud
cover in the data. Hence, to cater to the errors in the data, four
fault tolerant mechanisms are used at different levels.
1. In
SIFT
matching, two levels of filtering are applied, the
first is matching left-to-right and right-to-left, and the
second is applying
RANSAC
to eliminate the gross level
mismatches.
2. Evolutionary algorithms by their characteristics strive
to select the best/optimal points consistent with the
model. It is also clear from the controlled exercise that
the choice of objective function “A/E50” showed simi-
lar results as that of ‘no errors’ case despite introducing
errors in the data.
3. Control loop back based on decision rules as shown
in Figure 7 ensures that both model error and product
errors are consistent among themselves, which assures
the generated ortho-product is always correct.
4. Finally, as a “Safe Mode” operation of applying System
Level Model upon non-convergence always ensures
that the ortho-image is generated with at least system
level accuracy (see Figure 7).
To substantiate these arguments and to prove the reliability
and robustness of the proposed method, various tests have
been conducted, which are described in the next section.
Although the process is tested for IRS Cartosat series of mis-
sions, the process is generic and works for all the satellites
provided the corresponding
RSM
/
RPC
model is available.
Results
This Section is categorized into three Sub-Sections. The
first sub-section discusses the controlled exercise, where the
datasets are carefully chosen from different satellites, with the
scenes covering different types of terrain. Also, different refer-
ence databases are used for the same data. This section also
compares one of the results with some of the existing meth-
ods. The second section is aimed at arriving at the reliability
of the proposed methodology with a large number of scenes
from different satellites. The third section covers the robust-
ness of the proposed method, where the method is applied on
difficult scenes such as tiny islands where the scene contains
less features and has large amounts of homogeneous patches
predominantly covered with water and clouds. For all the
results the IRS Cartosat series of satellites are considered, and
the model as described in Radhadevi
et al.
(2008), and Srini-
vasan
et al.
(2009) are used as the
RSM
/
RFM
model.
Controlled Exercise
Two data sets of Cartosat-2A, one each from Cartosat-2B and
Cartosat-1 with different characteristics, i.e., hilly regions,
water bodies, partial cloudy, urban region, developed and
un-developed areas. The reference data used are switched
from one of the ETM, LDCM8/
OLI
and Google Earth extracted
images. Table 2 shows the list of datasets used, experiment
details, and obtained results. The complete process of ortho-
rectification is totally autonomous with no user input; how-
ever the intermediate stage-wise results for one of the scenes
is shown in Plate 1a through 1f.
From the CARTOSAT-2A image from orbit number
8011 acquired on 22 October 2009, Bengaluru City in In-
dia is considered for the experiments. The image extent is
T
able
2. S
ample
D
ata
S
ets
U
sed
for
E
xperiments
and
T
heir
R
esults
Data
Set Id
Reference
Image Used Image Characteristics
Check Point
from SIFT
Model used
RSM or RFM
Model Errors after
GA (Lat, Lon) in m
Product Errors
(Lat, Lon) in m
C-2A LDCM/OLI
*
Urban
69
RSM
(4.5,7.75)
(9.8,10.8)
C-2A Google Earth
#
(same as above)
90
RSM
(1.41,3.9)
(2.2,4)
C-2B LDCM/OLI
*
50% cloudy , Hilly and partial cloudy
45
RSM
(7.8, 5.2)
(10.6, 6.9)
C-1 LDCM/OLI
*
Long strip of 80kms
$
40
RSM
(6.2, 5.1)
(7.1, 5.4)
C-1 LDCM/OLI
*
Long strip of 100 Km.
25
RSM
(5.7, 6.9)
(8.8,6.1)
C-1 LDCM/OLI
*
Hilly 25KmX25Km
45
RFM
(6.6,7.3)
(4.5,4.7)
C-2 LDCM/OLI
*
Plain
113
RFM
(5.3,4.4)
(6.7,6.8)
C-2A: CARTOSAT 2A, nominal spatial resolution 0.8m *LDCM 8/OLI, spatial resolution 15m
C-2B: CARTOSAT 2B, nominal spatial resolution 0.8m # Google image, spatial resolution of 1m
C-1: CARTOSAT 1, spatial resolution 2.5m
$ 80Km long North-West orientation partially water covered
C-2 : CARTOSAT 2, nominal spatial resolution 0.8m
384
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
