IEEE Geoscience and Remote Sensing Magazine - December 2017 - 58

ill-posedness of inverse problems involving the estimation
of high-resolution subspace coefficients. Regularization on
subspace coefficients has been extensively explored. Sparsity-promoting regularization has been commonly adopted
for subspace-based methods [222], [225], [226], assuming that there is a limited number of endmembers at each
high-resolution pixel. HySURE solves a convex optimization
problem with vector-total-variation-based regularization
of the spatial distribution of subspace coefficients, leading
to implicit denoising effects in the fusion results [229]. Wei
et al. developed a Bayesian HS-MS fusion methodology in
which different types of regularization terms on subspace
coefficients can be designed flexibly, based on information
from the prior distribution in the observed scene [230]. A
Sylvester equation-based explicit solution was further integrated into the Bayesian methodology to speed it up, leading to fast fusion based on Sylvester equation (FUSE) [231].
Veganzones et al. [232] introduced local image processing
into an unmixing-based approach, where well-posed inverse
problems are solved for each small patch, assuming that the
number of endmembers in each patch is smaller than the
number of multispectral bands.
One interesting finding of the comparative study in
[215] was that hypersharpening methods and unmixingbased methods (CNMF and HySURE), which are entirely
different approaches, showed comparably high numerical
performances, although the characteristics of the resolution-enhanced HSIs are different. This finding implies that
hybrid methods combining different approaches can be expected to further improve fusion performance.
Table 5 shows the overall quantitative assessment results
of 12 HS-MS fusion algorithms for eight simulated HS-MS
data sets used in [215], including those based on the Pavia
University and University of Houston data sets. The quantitative assessment of fusion performance was carried out based
on a version of Wald's protocol presented in [232] and [215].

The reconstruction performances of algorithms has
posed the greatest concern for researchers; however, their
general versatility, computational costs, and impacts on
applications are also important for users. Because each
method has advantages and disadvantages, it is essential
to choose a method based on the fusion and analysis scenarios. The impact of HS-MS fusion on applications has
been recently investigated via classification and unmixing
[233], [215]. Further research on real data is still necessary to verify the practicability of HS-MS fusion for future
-hyperspectral satellites.
CHALLENGES IN RESOLUTION ENHANCEMENT
The main challenges in the resolution enhancement of HSIs
relate to practical issues. The development of multisensor
superresolution algorithms has recently surged; however,
very few publications in the literature discuss experiments
on real data. Studies regarding the impact of HSI resolution
enhancement on applications are still lacking. It is necessary to clarify benefits and address practical issues of resolution-enhancement technology to promote its operational
application for future hyperspectral satellite missions. For
instance, temporal mismatches included in the input images raise challenges for HSI resolution enhancement. Furthermore, a no-reference image-quality assessment for HSIs
needs to be developed to provide reliability information
regarding resolution-enhanced hyperspectral data products
for end users.
In multisensor superresolution, the problem remains
that large spectral distortions are inevitable when the mismatch between the two imaging sensors is large in either
the spatial or spectral domain (e.g., a large GSD ratio or hyperspectral pan-sharpening). To address this issue, a possible future direction for performance improvement lies in
developing algorithms that exploit a spectral library or spatial information of a high-resolution image.

TABLE 5. THE AVERAGE QUALITY MEASURES FOR EIGHT DATA SETS FROM [215].
CATEGORY

METHOD

PSNR

SAM

ERGAS

Q2n

GSA [206]

39.221

2.063

1.885

0.8796

MRA

SFIM-HS [207], [217]

41.074

1.707

1.789

0.8867

GLP-HS [208], [217]

41.322

1.664

1.733

0.8984

CNMF [224]

42.092

1.607

1.681

0.9060

ECCV '14 [225]

38.101

2.658

4.561

0.8431

ICCV '15 [227]

40.470

1.672

1.902

0.8847

HySURE [229]

42.336

1.602

1.766

0.9089

MAP-SMM [219]

40.008

1.822

1.997

0.8635

FUSE [231]

40.568

1.881

1.908

0.8693

FUSE-S [231]

41.177

1.803

1.804

0.8764

Unmixing

Bayesian

Ideal

Bold numbers indicate the best performances. HS: hyperspectral; ECCV: European Conference on Computer Vision; ICCV: International Conference on Computer Vision; FUSE-S:
sparse FUSE.

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2017

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017