PE&RS May 2016 - page 387

TB SATA hard disk, and RHEL OS. The time taken for a typi-
cal scene of IRS Cartosat series of satellites is less than five
minutes with RFM. A detailed break-down of the timelines is
indicated in Table 5.
Table 5. Time Taken for Each Module
Module Approximate Time taken (in Minutes)
for a Typical Scene of 12 K × 20K
Candidate Point for GCP
identification using SIFT 2 to 3
GA based GCP selection 20
generations, 10 Populations
using RSM
10 to 12
GA based GCP selection 20
generations, 10 Populations
using RFM
1
Ortho rectification 2 to 3
SIFT based Automatic
Product Evaluation 2 to 3
Total Time with RSM 16-20 minutes
Total Time with RFM Less than 5minutes
Conclusions
Conventional ortho-rectification process for remote images
involves considerable human intervention in identifying GCP s
and selecting good GCP s for refining the RSM or RFM, especially
when the resolution of reference image used are much coarser
as compared to the target image. In this paper, three major
contributions are made to generate ortho-images in autono-
mous mode for very high-resolution images. First, a two stage
SIFT-based checkpoint identification using reference databases
such as Landsat/ETM+, LDCM /OLI. Second, a GA-based GCP
selection methodology for refining the RSM/RFM model. Third,
a decision rule based Automatic Product Evaluation (APE )
for the final product acceptance or switching to safe mode of
operation upon reaching maximum number of iterations. The
proposed autonomous ortho-product generation methodology
was applied to several datasets pertaining to IRS Cartosat-
1/2/2A/2B satellites. Both RSM and RFM were tested and the
methodology worked as expected. The final relative error
of the ortho-product thus generated was found to be within
a pixel resolution of reference ortho-product. The method
worked successfully for almost all the images where the refer-
ence dataset was available and the quality of the images was
reasonably good. The reliability of the method for Cartosat-1
is well over 97 percent while for Cartosat-2/2A/2B was above
90 percent. The process was carried out on some difficult
data sets which were predominantly covered by either water,
clouds, or both, and the method was robust and worked satis-
factorily with similar reliability and accuracy.
The two stage SIFT-based check point generation produced
reliable and accurate checkpoints for the target VHR images
using much coarser HR images such as Landsat/ETM+ , LDCM/
OLI, as well as using VHR images from Google Earth as refer-
ence dataset. GA was used to select the best combination
of GCPs from the SIFT identified checkpoints. Five different
candidate objective functions were evaluated and all of them
are resulted in reasonably similar errors with minor variations
when the data sets had no induced errors. However, the objec-
tive function “Area/E50” was found to be more appropriate as
it selected good quality GCP s spread across the scene, resulted
in faster convergence, and most importantly found to be more
accurate under erroneous conditions.
The APE process involves computing RMSE from the SIFT
identified tie points between the reference image and the
ortho-product. In some rare cases mismatches were removed
either manually or applying RANSAC. In the future, a more
sophisticated filter mechanism at the APE level needs to be
worked out. At present, the decision rule base is implemented
using a look up table. In the future, more heuristics need to
be added to accommodate different types of satellites, terrain
types, availability, and accuracy of the reference data. With
our experience, it is understood that RSM needs to be cali-
brated regularly as the alignment of payload, detector arrays,
reaction wheels, which can accumulate small biases over a
long period of time. Hence, a feedback mechanism to auto-
calibrate RSM by analyzing the error patterns regularly can be
a possibility for future work. The complete process of ortho-
image generation as proposed in this paper takes less than five
minutes for a standard scene from IRS Cartosat-1/2/2A/2B
with RFM on the latest Xeon Quad Core processor. Process
optimization to parallelize and generate the ortho-products in
near real-time is another direction of future work.
Acknowledgments
The authors would like thank Director ADRIN Sri. Santanu
Chowdhury and former Director ADRIN Smt. Geeta Varadan,
for their continued support, and encouragement during the
research phase. Authors would like to express sincere thanks
to Sri. R. Ramachandran, former Associate Director ADRIN
for his valuable guidance during this research. The authors
would like to place on record the efforts of Sri. V. Srinivas,
Smt. M. V. Srivally and Sri. M.Janardhana Rao of ADRIN,
who have identified and provided the required imagery for
conducting the experiments. The authors also would like to
thank Dr. Saraswathi Puranik, who has readily accepted our
request to proof read and improve the readability of the paper.
References
Abdel Hakim, A.E., and A.A. Farag, 2006. CSIFT - A SIFT descriptor
with color invariant characteristics, ACM Digital Library,
Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, CVPR 2006, pp.
1978–1983.
Botterill, T., S. Mills, and R. Green, 2009. New conditional sampling
strategies for speeded-up RANSAC, Proceedings of the British
Machine Vision Conference.
Cai, G.R., P.M. Jodoin, S.Z. Li, S.D. Wu, S.Z. Su, and Z.K. Huang,
2013. Perspective-SIFT: An efficient tool for low-altitude remote
sensing image registration, Signal Processing. 93: 3088–3110.
Chen, Q.H., X.G. Liu, W. Gao, and T.L. Liu, 2009. An automatic
ground control point matching based on GCP chip database
for remote sensing images, Proceedings of the International
Conference on Image Analysis and Signal Processing, Taizhou,
China, doi:10.1109/IASP.2009.5054640.
Chen, L.C., C.Y. Lo , C.L. Liu, and A.J. Chen, 2000. Orientation
modelling by matching image templates of a GCP database,
Proceedings of the 21st Asian Conference on Remote Sensing,
ACRS, 21(2):1999–2004.
Chunlei, H., P. Chunhong, H. Leigang, and Z. Zhixin, 2012.
Multilevel SIFT matching for larger-size VHR image registration,
Proceedings of the IEEE Geoscience and Remote Sensing Letters,
9(2):171–175.
Farr, T.G., E. Caro, R. Crippen, R. Duren, S. Hensley, M. Kobrick ,
M. Paller , E. Rodriguez, P. Rosen, L. Roth, D. Seal, S. Shaffer, J.
Shimada, J. Umland, M. Werner, M. Oskin, D. Burbank, and D.
Alsdorf, 2007. The Shuttle Radar Topography Mission, Reviews
of Geophysics, 45, RG2004, doi:10.1029/2005RG000183.
Feng G., J.G. Masek, and R.E. Wolfe, 2009. Automated registration
and orthorectification package for Landsat and Landsat-like data
processing, Journal of Applied Remote Sensing, Vol. 3, 033515,
doi: 10.111 7/1.3104620.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2016 387