IEEE Geoscience and Remote Sensing Magazine - March 2021 - 68

For the test case at RR, both MS and PAN images are properly degraded, in agreement with Wald's protocol, as for the
Toulouse data set.
BENCHMARK
Several state-of-the-art algorithms are employed for comparison purposes:
◗◗ EXP: MS image interpolation using a polynomial kernel
with 23 coefficients [43]
◗◗ CS methods
** BT-H [59]
** BDSD [33]
** C-BDSD [34]
** BDSD-PC [35]
** GS [24]
** GSA [30]
** C-GSA [32]
** PRACS [29]
◗◗ MRA methods
** AWLP [61] with revised statistical matching between PAN and MS bands [31]
** MTF-GLP: GLP [43] with MTF-matched filter [47],
unitary injection model, and revised statistical
matching between PAN and MS bands [31]
** MTF-GLP-FS: GLP [43] with MTF-matched filter [47]
with a new FS regression-based injection model [133]
** MTF-GLP-HPM: GLP with MTF-matched filter [47]
with HPM injection model [143] and revised statistical matching between PAN and MS bands [31]
** MTF-GLP-HPM-H: GLP with MTF-matched filter
[47] with HPM injection model [143] and haze correction [59]
** MTF-GLP-HPM-R: GLP with MTF-matched filter [47]
and HPM injection model [143] with a new preliminary regression-based spectral matching phase [131]
** MTF-GLP-CBD: GLP [43] with MTF-matched filter
[47] and regression-based injection model [132]
** C-MTF-GLP-CBD: context-based GLP [43] with
MTF-matched filter [47] and regression-based injection model [132] with local parameter estimation
exploiting clustering [32]
** MF: nonlinear decomposition scheme using MFs
based on half gradient [55]
◗◗ VO methods
** FE-HPM: filter estimation based on a semiblind deconvolution framework and HPM injection model
[50]
** SR-D: pansharpening based on sparse representation of injected details [88]
** PWMBF: model-based fusion using PCA and wavelets [79]
** TV: pansharpening based on TV [67]
** RR: model-based reduced-rank pansharpening
[72], [134]
◗◗ ML methods
** PNN: proposed in [91]
68

** PNN-IDX: PNN augmenting the input by including
several maps of nonlinear radiometric indexes as
proposed in [91]
** A-PNN: proposed in [95]
** A-PNN-FT: proposed in [95].
A more detailed description of the methods can be
found in the " Review of Pansharpening Methods " section and in the related references. For the sake of brevity,
the setting of the parameters for each approach into the
benchmark can be found by having a look at the default
parameters in the related MATLAB toolbox distributed to
the community.
REDUCED-RESOLUTION PERFORMANCE ASSESSMENT
This section focuses on the presentation of the experimental results at RR. The comparison will be performed class
by class, splitting the section into four subsections, one for
each class. For a wider (interclass) discussion, interested
readers can refer to the " Discussion " section, which completes the discussion of the experimental results.
COMPONENT SUBSTITUTION
The products obtained through the CS methods have always been appreciated for the high quality of the spatial
information, which is accurately transferred from the PAN
image to the pansharpened image. The analysis conducted
in this study confirms this consideration but highlights that
a proper detail-extraction phase is required to obtain high
performance. In fact, the GS method, which uses a simple
average for constructing the low-pass PAN image from the
MS channels, is not able to bear comparison with the other
CS approaches.
A second crucial consideration regards the remarkable
advances that have been obtained by the most recent implementations of the classical CS methods. In particular, partitioning the image in segments allows a more specific evaluation of the coefficients required by both the BDSD and
the GSA method. However, the context-adaptive BDSD approach does not solve the difficulty of estimating the larger
number of parameters involved in the fusion of the eightband WV-3 images. On the contrary, the introduction of the
constraints with the BDSD-PC technique allows us to significantly increase the robustness of the method, avoiding physically unjustifiable solutions. Similarly, taking into account
the characteristics of the image acquisition process permits
us to achieve a high-quality fused product through the BT-H
method, while the original BT version was classified as a lowperformance approach in the previous study [14].
The visual analysis of the images, as in Figures 7 and 8,
highlights that the improvements obtained by the recent
approaches are related to a more targeted introduction of
the PAN information into the original MS image, which
is the result of taking into account the features of the acquisition sensors and the spatial characteristics of the illuminated scene. As a consequence, the details are more
easily identifiable in the final products due to a higher
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2021
IEEE Geoscience and Remote Sensing Magazine - March 2021

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