and human effort. With the increase in the number of images
being acquired on a daily basis, the development of automat-
ed techniques for geo-coding/ortho-rectification has become
absolutely essential and critical. In this paper, we compare the
existing implementations of automated registration/rectifica-
tion methods and present an alternative approach which auto-
mates the entire process of generating ortho-products for Very
High Resolution (
VHR
) images using High Resolution (HR)
and/or less accurate reference databases, with robust blunder
detection and quality
GCP
selection methodologies.
Major contributions in this paper to achieve the goal of
complete automation include, (a) robust matching to obtain
large number of checkpoints in
VHR
target images using much
coarser resolution reference images compared to target im-
age (15 m ortho-images as a reference to rectify 0.8 m target
images), (b) selection of good quality
GCP
s by incorporating
an evolutionary approach with built-in fault tolerance in the
model, and finally (c) a control loop based on decision rule
base for evaluating the final product.
In the following section we describe the related work in the
areas of
GCP
identification and use of
SIFT
and Genetic Algo-
rithms (
GA
) in the area of remote sensing and ortho-rectification,
followed by the proposed methodology of generating an ortho-
image in an autonomous mode. The results of the methodology
applied on the IRS Cartosat series of images are presented lead-
ing to the conclusions and proposed future directions.
Related Work
The ortho-rectification process involves refinement of
RSM
us-
ing
GCP
s, followed by geometric and relief displacement correc-
tions using the refined
RSM
and
DEM
. There are two sources of
obtaining
GCP
s, that is, either through direct
GPS
measurements
or by identifying them from known reference ortho-databases
such as Landsat
ETM+
or
LDCM
/
OLI
and
DEM
from
SRTM
or
ASTER
Global
DEM
. Chen
et al.
(2009) proposed automated
GCP
identi-
fication based on image matching techniques using
GCP
chips
database. Image matching methods are generalized into two
categories, i.e., region-based matching such as cross correla-
tion, mutual information, and feature matching such as
SIFT
(Li
et al.
, 2009), and Harris
et al.
(1988). Feature based matching
methods are supposedly better compared to region-based meth-
ods as they are scale and rotation invariant, and also are less
prone to noise and illumination variations in the data.
Krauss
et al.
(2013) proposed a fully automatic processing
system CATENA for high-resolution images, where the target
images considered were SPOT4, SPOT5, and the IRS-P6 LISS
camera. Feng
et al.
(2009) proposed Automated precise Reg-
istration and Ortho-rectification Package (
AROP
), which is for
both Landsat and Landsat-like images. Liu
et al.
(2009) and
Gianinetto
et al.
(2008) proposed automatic
GCP
identification
using high-resolution imagery. In all of these, the resolution
of reference images used are either similar or better than the
target image resolution. For Very High Resolution images such
as IRS Cartosat-2, Ikonos, etc., availability of comparable or
better resolution reference images globally is not ensured and
also involves a enormous cost.
Chen
et al.
(2000) proposed automatic
GCP
s identification
using a template-based image matching/wedge corner match-
ing from pre-identified chips stored in the database, which
are matched using normalized cross correlation and least
square matching. Yang
et al.
(2009) proposed a fast rectifica-
tion using feature control point database. The disadvantages
with these methods is twofold, first the
GCP
collection and
database creation is a manual process, and second, the match-
ing processes are sensitive to seasonal changes and cannot be
extended to very different resolution images such as Landsat
ETM and Cartosat-2 (i.e., HR to
VHR
). Hence, to improve accu-
racy multi-resolution and multi-seasonal
GCP
chips (Liu
et al.
,
2009) were proposed as schema based on feature extraction
and matching using
SIFT
(Yi
et al.
, 2009) and
SURF
(Teke
et al.
,
2010). However, the database creation process still remains a
manual process. Additionally, they require usage of Very High
Resolution images such as Spot5, QuickBird, etc. to collect
GCP
s, which in itself would be a difficult task. Leprince
et al.
(2007) proposed a method for auto-rectification and registra-
tion based on iteratively refining the rough selection of
GCP
s
using an inverse ortho-rectification model and sub-pixel
correlation matching. However, the process of identifying
the rough selection of
GCP
s (referred as image control points)
is again a manual process. Xing
et al.
(2000) proposed im-
age matching using
GA
and Least Square Matching (
LSM
) to
improve the time taken for matching. Here, the
GA
is used to
find approximate match point within 2 to 3 pixels, followed
by an
LSM
which would improve the matching to sub-pixel
accuracy. All these works can be categorized into two major
limitations: first, these methods demand use of similar resolu-
tion or better resolution imagery for automated collection of
Figure 1. Ortho-Rectification Process.
378
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
