PE&RS May 2016 - page 377

Autonomous Ortho-Rectification of
Very High Resolution Imagery Using SIFT
and Genetic Algorithm
Pramod Kumar Konugurthi, Raghavendra Kune, Ravi Nooka, and Venkatraman Sarma
Abstract
Ortho-rectification of very high resolution imagery from agile
platforms using Rigorous Sensor Model / Rational Functional
Model is quite challenging and demands a fair amount of in-
teractivity in Ground Control Point (GCP) identification/selec-
tion for refining the model and for final product evaluation.
The paper proposes achieving complete automation in the
ortho-rectification process by eliminating all the interactive
components, and incorporating fault tolerance mechanisms
within the model to make the process robust and reliable.
The key aspects proposed in this paper are: two stage Scale
Invariant Feature Transform ( SIFT) based matching to obtain a
large numbers of checkpoints using much coarser resolution
images such as Landsat/ETM+, followed by a GA to select the
right combination of minimal GCPs based on minimizing Root
Mean Square Error (RMSE) and maximizing the area covered
under GCPs, and finally, a decision rule based product evalu-
ation to make the process operate in an “autonomous closed
loop mode”. The method is generic and has been tested on
hundreds of Cartosat-1/2 images, and has achieved above
90% reliability with sub-pixel relative error of reference data.
Introduction
The raw satellite images taken from very high resolution agile
satellites, such as Cartosat-2, Ikonos, QuickBird, EROS etc.,
introduce high relief displacements in hilly terrain and also
suffer from high geometric distortions due to platform stabil-
ity, variations in sensor mechanics, etc. and the errors are
also compounded due to high oblique imaging (Toutin, 2004).
Hence, such data demands geometric corrections before being
put to use for analysis and measurements. Ortho-rectification is
a standard process in remote sensing for correcting the geomet-
ric distortions and relief displacement errors introduced by the
payload system during the imaging time. Ortho rectification
of satellite imagery requires a precise Rigorous Sensor Model
(RSM) or Rational Function Model (RFM) which are refined us-
ing well-distributed Ground Control Points (GCPs). RSM links
the object space to image space and is satellite/sensor specific.
The model is derived using complete knowledge of acquisition
system’s internal details such as sensor characteristics, align-
ment of CCDs within the focal plane, offset angles with respect
to payload cube normal, which are in general known to agency/
agencies which assemble the satellite. Alternative to RSM is a
sensor neutral RFM or Rational Polynomial Coefficients (RPCs),
a generic model that abstracts the satellite/sensor details to the
end users (Tao et al., 2000). The raw RPC coefficients and RSM
without use of any GCPs would have limitations due to platform
stability, imaging sensors and other on-board measuring sen-
sors such as star sensor, gyroscope, GPS etc., and would result
in a large amount of geometric errors (Toutin, 2004). These
errors are quite often of the orders of around few tens to a
hundred meters depending upon the accuracy and precision of
the onboard recording devices and satellite/sensor calibration
mechanisms. GCPs (precise Latitude, Longitude, and Height of a
particular feature on image) can be used to refine the RSM/RFM
(Chen et al., 2009), and thereby can reduce the geometric errors
in the model and subsequently the final product. The primary
sources of obtaining GCPs are either through ground surveying
using GPS, or a combination of ortho-rectified reference datasets
such as Landsat/ETM+ ( Landsat-7), Landsat/OLI (Landsat) and a
Digital Elevation Model (DEM) derived from Shuttle Radar The-
matic Mapper (SRTM) (Farr et al., 2007) /ASTER Global DEM
(Fujisada, 2012). However, the satellite systems such as
Worldview-3 or Pleiades which have better onboard devices
and better calibration mechanisms provide better accura-
cies without use of any GCPs. Therefore, refinement of RSM
using GCPs from Landsat-7/ Landsat may not be required, as it
would deteriorate the product.
Figure 1 explains the standard process of ortho-rectifica-
tion being adopted by the satellite data processing commu-
nity. In case of RFM, the raw images and the corresponding RPC
coefficients form the basic inputs for the ortho-rectification
process, while for RSM ephemeris data and raw images are
provided as inputs. First, the checkpoints are identified
from the reference ortho-image databases such as Landsat/
ETM+ or LDCM/OLI and DEM from SRTM or ASTER Global DEM .
Subsequently, a subset of checkpoints is selected as GCPs to
refine the model. The process of GCP subset selection is done
iteratively so that the errors are minimized.
During these iterations some of the already identified
checkpoints may be considered as GCP s or vice-versa based on
heuristics or trial and error. The error metric generally used in
such cases is RMSE or CE90. Once the GCPs and the model are
accepted, ortho-rectification is a well known process and can
be generated as shown in the block marked as “$” in Figure
1. After the product is generated, the product is evaluated by
identifying the new set of tie points in both the reference im-
age and the ortho-product. The product error (RMSE or CE90) is
computed, and if the error is found to be within the accept-
able limits, then the product is accepted, or else the process
is repeated by refining the GCPs. The identification of GCP s for
refinement of RSM/RFM, and tie points for product evaluation
is generally done in such a way that the points are evenly
distributed across the scene.
The shaded blocks (*, +, #) in Figure 1 indicate the interac-
tive/semi-automated processes, which consume a lot of time
Authors are with Advanced Data Processing Research
Institute (ADRIN), Department of Space, Government of India,
Manovikas Nagar, Secunderabad-500009, India
(pramod@adrin.res.in).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 5, May 2016, pp. 377–388.
0099-1112/16/377–388
© 2016 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.5.377
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2016 377