IEEE Geoscience and Remote Sensing Magazine - December 2021 - 122

were presented early in the field's development, with remote
sensing image registration accounting for a minimal
percentage of annual WoS publications. More recently, a
considerable number of studies have been published, peaking
in 2019. Thus, comprehensive analysis is necessary to
identify unsolved problems for the rapid development of
this field.
In this article, we summarize various classical approaches
to remote sensing image registration as well as
recent methods based on deep learning, optical flow estimation,
and image registration software. We also point out
interesting aspects and analyze development trends from
our perspective, without describing specific approaches
in detail. Concretely, the registration approaches can be
classified into three categories, namely,
intensity-based,
3,500
3,000
2,500
2,000
1,500
1,000
500
feature-based, and combination registration, as detailed in
Figure 4. The intensity-based technique directly uses pixel
intensity information to register images, including the conventional
area-based approach and optical flow estimation.
The geometrical and advanced features used to register images
instead of intensity information are defined as featurebased
approaches. Combination registration mainly consists
of the integration of feature- and area-based methods
as well as two geometric feature-based techniques. Many
detailed classifications are presented in each category.
All registration approaches must undergo coordinate
Published Papers
Proportion
2019: 3,154
Year
FIGURE 3. The number of papers about remote sensing image registration on WoS, per year.
transformation and resampling to ultimately acquire the
aligned image, as demonstrated in Figure 5. Before this step,
transformation models for coordinate recalculation other
than optical flow estimation should be constructed. In general,
transformation models, such
as the affine, projective, piecewise
linear, and thin spline models, are
derived from global or local parametric
models. To calculate these models,
images are preprocessed to extract
representative features through techniques
including geometrical- and
advanced-feature extraction and
matching. Given that intensity information
is directly utilized in areabased
registration, feature extraction
is omitted, and the transformation
model is constructed when matching
the intensity information. Since most
approaches prefer to contribute to the
preliminary steps (e.g., feature extraction,
feature matching, and mismatched
feature elimination) rather
0.14
0.12
0.1
0.08
0.06
0.04
0.02
* Feature- and Area-Based Methods
Combination
Method
* Two Geometric Feature-Based
Methods
Intensity-Based
Method
Remote
Sensing Image
Registration
* Area Based
Frequency Domain
Spatial Domain
* Optical Flow
Dense Optical Flow
Sparse Optical Flow
Feature-Based
Method
* Geometrical Feature Based
Points, Lines, and Polygons
Feature Matching...
FIGURE 4. The remote sensing image registration algorithms. GAN: generative adversarial network.
122
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021
* Deep Learning
Siamese Network
GAN
Published Papers
1979
1981
1983
1985
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
2007
2009
2011
2013
2015
2017
2019
2021
Proportion (%)

IEEE Geoscience and Remote Sensing Magazine - December 2021

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