IEEE Geoscience and Remote Sensing Magazine - March 2021 - 38

For the eight-band WorldView-2 and WorldView-3 satellite
images, the overall performance is relatively poorer than
the four-band Ikonos, QuickBird, GF-1, and WorldView-4 satellite images. This can be seen in most of the quantitativeevaluation indices, such as the ERGAS, SAM, and Q2 n . On
the one hand, the PRACS, MTF-GLP, MTF-GLP-HPM, MTFGLP-CBD, and MSDCNN methods show overall satisfactory performance for the WorldView-2 and WorldView-3 images. In addition, the BDSD method shows a good result in the
WorldView-2 experiment. On the other hand, the PRACS,
MTF-GLP, and MTF-GLP-HPM methods show inconsistent
performance in the reduced- and full-resolution experiments for the WorldView-2 and WorldView-3 images. Particularly, the PRACS and MTF-GLP-HPM methods have excellent performance in terms
of the full-resolution quantitative-evaluation indices
FEW PUBLICLY AVAILABLE
and relatively poorer perforBENCHMARK DATA SETS
mance in the reduced-resolution experiment. Conversely,
FOR PANSHARPENING ARE
the MTF-GLP method shows
AVAILABLE, ESPECIALLY
a satisfactory result in the reLARGE-SCALE ONES.
duced-resolution experiment
and relatively poorer performance in the full-resolution
experiment. The MTF-GLP-CBD method shows competitive results in the reduced- and full-resolution experiments
for the WorldView-2 and WorldView-3 images.
On the whole, the experiment results show that the pansharpening methods have slightly different performances
on various remote sensing satellites, especially between
the four- and eight-band images. Most of the pansharpening methods have a relatively poorer performance for
the eight-band images than for the four-band ones. The
BDSD and MSDCNN methods generally show robust
and excellent performance in the reduced- and full-resolution experiments. The MTF-GLP and MTF-GLP-HPM
-PP methods generally show satisfactory performance
in terms of the quantitative-evaluation indices in the reduced-resolution experiment and poorer performance in
the full-resolution experiment. Conversely, the MTF-GLPHPM generally shows satisfactory performance in terms
of D m, D s, and the QNR in the full-resolution experiment
and relatively poorer performance in the reduced-resolution experiment.
EXPERIMENTAL RESULTS BASED ON DATA SETS FOR
DIFFERENT THEMATIC SURFACE FEATURES
The experiment results based on different thematic
data sets are shown in Figures 12-15 and Tables 10-13,
respectively. For each figure, the horizontal axis represents different pansharpening methods, and the vertical axis denotes the statistical results. In addition, the
top five results for each evaluation index are marked by
dark colors. In the experiments, to avoid the influence
of the different spectral range coverages between the HR
38

Pan and LR MS images, only thematic images from Ikonos, QuickBird, GF-1, and WorldView-4 with four-band MS
were employed. Therefore, 305 pairs of HR Pan/LR MS
images in the urban data set, 203 pairs of HR Pan/LR
MS images in the green vegetation data set, 153 pairs of
HR Pan/LR MS images in the water-scenario data set, and
949 pairs of HR Pan/LR MS images in the unlabeled data
set were utilized.
As shown in Figure 12 and Table 10, the BDSD and
MSDCNN methods show excellent performance based on
the urban data set in the reduced- and full-resolution experiments. The ATWT, MTF-GLP, and MTF-GLP-HPM-PP
methods show satisfactory results in the reduced-resolution
experiment and relatively poorer performance in the fullresolution experiment. Conversely, the MTF-GLP-HPM and
MTF-GLP-CBD methods have satisfactory performance in
the full-resolution experiment and poorer performance in
the reduced-resolution experiment.
Figure 13 and Table 11 provide the experiment results
based on the green vegetation data set. It is shown that
the BDSD and MSDCNN methods have excellent performance in the reduced- and full-resolution experiments;
the MSDCNN method has the best results. Similar to
the results based on the urban data set, the ATWT and
MTF-GLP methods have better results in the reduced-resolution experiment, and the MTF-GLP-HPM and MTFGLP-CBD methods achieve better performance in the
full-resolution experiment.
The experiment results based on the water-scenario
data set are shown in Figure 14 and Table 12. The performance of almost all the pansharpening methods is
poorer than for the urban and green vegetation data sets,
which can be observed in the Q2 n index. This is because
the edges and detailed structures of the images in the
water-scenario data set are generally missing, and the advantages of the HR Pan cannot be comprehensively utilized. In addition, the results show that the SFIM, MTFGLP-HPM-PP, and MTF-GLP-HPM methods have poor
robustness with a wider confidence interval in terms of
the ERGAS.
For the unlabeled data set with mixed surface features,
the BDSD and MSDCNN methods show excellent performance, and the MSDCNN has a competitive advantage
over most of the other pansharpening techniques in the
reduced- and full-resolution experiments. In addition,
similar to the results based on the water-scenario data
set, the SFIM, MTF-GLP-HPM-PP, and MTF-GLP-HPM
methods display poorer robustness in terms of the ERGAS index.
On the whole, the experiment results based on the
proposed large-scale thematic data set show that different pansharpening methods have a relatively larger difference for various thematic images. In addition, the SFIM,
MTF-GLP-HPM-PP, and MTF-GLP-HPM methods with the
same HPM high-frequency injection model generally show
poorer robustness.
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