PE&RS May 2016 - page 379

GCPs and creation of GCP database, and second they are based
on area matching methods, which are prone to errors and are
susceptible to temporal/seasonal variations.
Chunlei et al. (2012), and Li et al. (2009) proposed SIFT
matching for large-size high-resolution image registrations,
where the images are of similar resolution and orientation.
However, some of the challenges for GCP identification for
very high-resolution images like Cartosat-2 include change in
orientation angles caused by agile satellites for imaging the
specified region, higher levels of noise in the data and most
importantly, significant variations in resolutions in the glob-
ally available reference database for GCP identification such
as Landsat/ETM+ and LDCM/OLI. Lowe et al. (2004) proposed
a scale and rotation invariant feature-based matching using
SIFT for extracting distinctive invariant features from im-
ages that can be used to perform reliable matching between
different views of an object or scene with different scale and
orientation. Jiang et al. (2012) compared major feature point
detection algorithms such as Forstner, SUSAN, Harris, and SIFT,
and proposed that SIFT produces more consistent, stable, and
accurate points and is less prone to noise in the data. Later,
many variants of SIFT were proposed in PCA-SIFT, CSIFT, GLOH,
SR-SIFT and Robust-SIFT to make it more effective and suit-
able for remote sensing images. Speed Up Robust Features
(SURF) also produces similar feature points and consumes less
time. However, it produces minimal effective points (Li et al. ,
2009). Cai et al. (2013) discussed Perspective SIFT, for register-
ing remote sensing images having severe perspective distor-
tions by extending SIFT and Affine- SIFT (Morel et al. , 2009).
However, all of these are applied on similar resolution images
and for image-image registration, but they do not discuss
cases, when (a) there is a severe scale variation in the refer-
ence and target images, (b) when there are mismatches that
can arise due to presence of large homogeneous regions in the
satellite images, and (c) there are geometric distortions that
are introduced by very high resolution agile satellites.
All the above cases may lead to a large number of mis-
matches using SIFT-based matching, especially when much
coarser resolution images are used as reference for correcting
much higher resolution images. Hence, there is a requirement
for a strong homologous point detection methodology which
filters out mismatches, or alternatively a much better strategy
is a strong selection mechanism, which selects the best subset
of required number of GCPs from a large number of points.
Note that for both of these methods, i.e., homologous point
selection and selection of minimal good, GCPs are different.
For identifying the inliers, many variants of RANSAC (Fischler
et al., 1981) and BaySAC (Botteril et al., 2009), Optimized
BaySAC (Zhizhong et al., 2014), can be used. Ren et al. (2014)
discussed SAR image-image registration using SIFT for real
time navigation using similar resolution images as refer-
ence and also stated that SAR images due to the multiplica-
tive noise characteristics in the data ,always produces fewer
match points using SIFT. Valadan et al. (2002) proposed use of
GA for optimal selection of coefficients in global polynomial
and multi-quadratic 2D mathematical models for 2D geomet-
ric correction of Ikonos images in the absence of RSM/RFM.
However, in this work the RSM is being used and is assumed
to be available. Senthilnath et al. (2014) discussed the usage
of Genetic Algorithm (GA) in various practical scenarios for
image registration, and subsequently discussed the issues
related to SAR image registration with optical image. They
proposed usage of SIFT and multi-objective GA with angle,
distance, and vicinity criteria as fitness function (2014) for se-
lection of tie points. However, in all these variants of RANSAC
and GA that are discussed, similar resolution images are used
and are applied for image-image registrations and emphasized
matching consistency rather than the product accuracy. The
two important limitations in all these methods are: lack of
selection criterion based on the spatial distribution of con-
trols, which is very crucial for ortho-rectification of images
acquired from high resolution agile satellites using RSM/RFM,
and an assumption that similar or better resolution reference
images are available and are also accurate, which may not be
possible at all locations. Here, we propose an evolutionary
method based on GA for selection of good GCPs which consid-
ers distinct factors in the fitness function, such as minimal
RMSE, higher spatial distribution of GCP s, as well as sufficient
margins for errors in both data and processes, which is essen-
tial for design of autonomous systems.
Methodology
The proposed methodology of ortho rectification remains the
same as discussed in Figure 1. However, the key interactive
elements namely “Identify/Refine GCP *”, “Evaluate Model
Errors+”, and “Evaluate Product#” are automated with the
changes given below.
1. SIFT-based checkpoint identification,
2. GA driven GCP selection, and
3. Decision Rule based Product Evaluation
SIFT-based Checkpoints Identification
SIFT is a robust feature detection algorithm, which is scale and
orientation invariant and produces richer, more stable, and
accurate matches. However, use of SIFT for checkpoint iden-
tification for very high-resolution images using much coarser
level reference database involves issues in matching with
high scale variations, high orientation changes, etc. The best
high-resolution known reference database of ortho-images
available globally is from Landsat-ETM+ and LDCM/OLI , which
has a spatial resolution of 15 m, while the target images under
consideration are of Cartosat-1 and Cartosat-2/2A/2B, hav-
ing spatial resolution less than one meter. Empirical analysis
resulted in low confidence in matching when SIFT was used
directly on these remote sensing images. Hence, we adopted a
two stage matching, first matching at intermediate resolution
as the pre-cursor, followed by matching at original resolu-
tion of input images, using the cues from the first stage. The
process is depicted in Figure 2. First, the exact extent pertain-
ing to that of target image from the reference ortho-image is
extracted using ephemeral RSM, followed by oversampling the
reference image and sub-sampling of the target image to an
intermediate resolution.
The selection of intermediate resolution is new, and we
believe it is a crucial step which provides the cues for the later
stage matching and produces good and consistent tie points.
After constructing the corresponding images, SIFT is applied
for feature extraction. A bi-directional feature matching (i.e.,
Left-to-Right and Right-to-Left) is proposed to eliminate the first
level of mismatches (if any), followed by RANSAC to eliminate
gross level outliers in the matching process. The resulting tie
points are used to estimate the transformation coefficients (i.e.,
the offset and orientation) which are used in the second stage.
In the second stage, block-wise SIFT features extraction and
feature matching is performed in original resolution of input
image. The estimated transformation coefficients from first stage
are used for extracting the exact extent pertaining to each block.
Block-wise matching serves two purposes, i.e., to keep the im-
age sizes small so that the processing is easier as well as keep-
ing the memory requirements for computation low, and also to
restrict the search space for feature level matching within the
block so that the scope for mismatches is less. Note that SIFT
produces a large number of match points for remote sensing
images, hence we propose to select the best points based on a
feature matching threshold, so as to constrain the problem space
for GA. Since the tie points would be in image coordinates (scan,
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2016 379