IEEE Geoscience and Remote Sensing Magazine - December 2021 - 133

subregion [30]. A detection approach is employed to extract
the specified number of feature points. Tie points affect the
registration accuracy but are not the sole influencer.
TRANSFORMATION MODEL PERFORMANCE
The transformation model abstractly represents the geometric
mapping function from a sensed image to a reference image.
The actual between-image geometric distortion is difficult to
obtain without prior information, and the estimated transformation
approximates the real geometric relationship between
images. One part of the N pairs of tie points is taken
for mapping function estimation through the least-squares
method, assuming N matched feature points. The left part
in the sensed image is employed as the test point to be transformed
into the reference image system [188]. The distance
between the transformed coordinate and the corresponding
point in the reference image is calculated as the residual, the
mean of which is a representation of the estimated transformation
model:
RMS
Nte- =
NT xHxy Hy
1
NT
- / (( -+ -
j
=
-
1
where H denotes the estimated transformation model by T
pairs of tie points, (, )xy and (','),xy which represent the corresponding
points in the reference and sensed images, respectively.
Furthermore, a
| goodness-of-fit test may be applied [188]
2
to analyze whether the residuals are equally distributed across
all quadrants. However, " overfitting " may yield zero error for
a mapping model with sufficient degrees of freedom; this is a
well-known phenomenon in numerical analysis. Under this
circumstance, the registration results may not be optimal.
ALIGNMENT ERROR
The oldest method for estimating registration accuracy is visual
assessment by a domain expert, which is still in use and
remains the most effective technique, although it cannot be
quantified [16], [188]. At present, this is performed using
professional software, such as ENVI and ArcGIS, with shutter
tools. Similarity metrics in area-based registration, such as
MI, NMI, CC, and so on, are frequently employed to evaluate
alignment accuracy [59]. The indicators are easily influenced
changes in the information with development and differences
in radiation. To quantitatively present the alignment error, the
RMSE is calculated using feature points manually extracted by
a specialist employing (4) [85]. Since image registration aims
to achieve the relative spatial alignment of two different images,
there is no gold standard reference image with which to
evaluate the registration accuracy. When evaluating outcomes
according to at least three criteria, the most indicative results
point to the best registration, as different assessments have
their own advantages and disadvantages.
FUTURE TRENDS
There has been a large number of independent studies on remote
sensing image registration, and much effort has been
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
COMBINED APPROACHES FOR IMAGE REGISTRATION
With the development of imaging sensors, the resolution
of remote sensing images has increased, and local deformation
has become obvious. For example, the geometric
distortion caused by terrain relief and high-rise buildings
leads to inaccurate registration [36], introducing difficulties
for remote sensing image applications. The reference
and sensed images cover the plain and mountainous regions
simultaneously in Figure 13(c). Calculating the displacements
of corresponding pixels for spatial registration,
the enlarged displacements in the specified rectangular regions
are shown in Figure 13(d) and (e). The magnitude
and direction of the displacements in the plain region are
similar, but they differ in the mountainous region. Here,
multistage registration with a global mapping function
cannot exactly describe the spatial relationship between
the reference and sensed images, and neither can the local
transformation model.
Given that displacements vary in different terrain regions,
dividing images into a series of regions and registering with
a specified approach may yield a high-precision alignment,
indicating a combination of different techniques. Concretely,
133
') (')),
22
(5)
put into constructing robust feature descriptors and eliminating
mismatched features. With the development of sensor
technology and application requirements, some novel
opportunities and challenges must be addressed for remote
sensing image registration. To us, it seems likely that
the future of this field will include accelerated, combined,
heterogenous, cross-scale, and smart remote sensing image
registration techniques, which are introduced in detail
in the following.
ACCELERATED REMOTE SENSING
IMAGE REGISTRATION
With the ongoing development of sensor technology, the
spatial resolution of remote sensing images increases, resulting
in a growing number of features with distinctive
details. The huge number of features lengthens the distance
to the real-time registration of remote sensing images,
causing inefficiency when aligning large-scale images.
Thus, constructing descriptors and matching the detected
features is time-consuming for general images, especially
WFV ones. As proposed in [52], to achieve real-time registration
to the greatest extent possible, remote sensing image
registration can be operated on a cloud platform based
on finite-state chaotic compressed sensing theory. Similarly,
cloud computing [91] and some hardware systems may also
be effective for accelerating image registration. At present,
parallel computing [139] is the easiest path to implementation.
Here, an image is divided into several subregions,
and the image features in each one are simultaneously extracted,
based on the same principles, on different parallel
processors, as is the transformation model construction.
The parallel commands are easy to implement on MATLAB
and other platforms.
IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021
