IEEE Geoscience and Remote Sensing Magazine - December 2020 - 28

3) The algorithm should be robust enough to avoid any
false point matching and consequent misregistration
mainly due to too many servo errors and errors in scan
mirror mechanism repositioning during acquisition.
4) The software should be generalized enough for possible
future modifications, whenever necessary.
5) Because the software is to be incorporated in the operations software, it is expected that the complete process
takes a minimum amount of time to execute.
SIMILARITY MEASURES, USAGE, AND MATCHING
CRITERION
As mentioned in the problem definition, a sensed image
is to be registered with a reference image. Referring to Figure 4, the first step in image registration is identifying corresponding (similar) features between the reference and a
sensed image. In general, nonlinear distortions will be present in the incoming satellite imagery. Hence, establishing
the points of similarity between the reference and a sensed
image is a key step in the process of data registration. It is
done using a similarity measure that quantifies the degree

Reference Image

Sensed Image

Similarity Measures
and Matching Criterion

of correspondence between the pair of images. Typically, the
similarity is calculated at a certain number of points in the
images. The points at which the similarity metric value is
higher are considered the match points. It is clear from Figure 4 that, in turn, the set of match points calculated from a
similarity measure controls the accuracy level of data registration because the transformation parameters are estimated
for the established match points between the reference and
sensed image. Hence, the adopted similarity measure should
be robust, accurate, and objective. In this implementation,
two classical measures, i.e., normalized cross correlation coefficient (NCC) and minimum sum of squared difference
(MSD), and two measures based on statistical dependence,
i.e., the cluster reward algorithm (CRA) [12] and mutual information (MI) [13], [14], have been tested for the data.
To this end, we implemented a moving-window algorithm to automate the match-point identification process as
follows. As shown in Figure 5, a small target window (T) is
taken at a GCP in the reference image. A larger search window (S) in the sensed image is considered at the same location, i.e., coordinates of the GCP in the reference image. The
sizes of the target and search windows are heuristically chosen, inspired by fundamentals in classical image-processing
theory [15]. The target window is moved within the search
window pixel by pixel and scan by scan, as shown schematically in Figure 5. For each combination of the two windows,
the similarity measure is calculated and stored. At the end of
the process, the best matching position of the target window
is found based on the matching criterion of the similarity
measure. The entire process is repeated for all GCPs.
We now define the similarity measures along with their
matching criterion. A very first measure is sum of squared
difference (SSD). This measure calculates squared differences between pixel brightness values in a reference and a
sensed image. The SSD [16] is defined as

SSD (i, j) = /

/ 6rij - s ij@2,(1)
N

i = 1 j =1

Transformation
Parameters Estimation

Resampling of Sensed Image
and Transformation

Registered Image
FIGURE 4. A methodology for image registration of KALPANA-1

and INSAT-3A.
28

where N is the total number of rows or columns in the image; rij
and sij are the brightness values at reference and sensed images,
respectively; and i, j are the image coordinates. It is evident from
(1) that the point at which the SSD is at a least value gives the
location of the match point. Thus, it is also called MSD.
The next classical measure is NCC [17], which is defined in
(2), shown in the box at the bottom of the next page, where
μr and μs are the means of target and search windows, respectively. It is the well-known cross-correlation coefficient whose
value lies between -1 and + 1. The point at which the NCC is
maximum gives the location of the match point.
We tested two statistical measures to find the match
points in a sensed image. First is the CRA [12], which is
based on a joint histogram of reference and sensed images.
Let H RS (k, l) be the joint histogram of the reference and
the sensed images, and H R (k) and H S (l) be the histograms
of reference image and sensed image, respectively. It is defined as
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2020

IEEE Geoscience and Remote Sensing Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2020