IEEE Geoscience and Remote Sensing Magazine - June 2017 - 48

observation affirms that denoising postprocessing is essential
in the evaluation methodology to fairly compare the resolution-enhancement performance of different approaches
whenever HS-MS fusion simulations take into account additive noise for input images.

is room for improvement by performing more detailed
spectral grouping (e.g., [28] and [31]) and synthesizing
high-resolution images as hypersharpening. The method
of grouping remains an open issue. ECCV'14 shows good
PSNR values in the spectral ranges covered by MS bands.
On the other hand, they dramatically decrease in the other spectral ranges, which is possibly caused by inaccurate
estimation of high-resolution abundances.

srF oVerLaP
To further investigate the impact of SRF overlap on the quality of fused images, HS-MS fusion was performed on the
AVIRIS Cuprite data using the following three scenarios of
WV-3 MS bands: 1) VNIR-SWIR 16 bands, 2) VNIR eight
bands, and 3) SWIR eight bands. Figure 17(a) and (b) presents PSNR and SAM, respectively, with these three scenarios
of MS bands. As can be expected, in the second and third
scenarios, the fusion performance of all the methods decreases compared to those in the first scenario. In particular,
the Bayesian methods and ECCV'14 show significant degradation in both PSNR and SAM, whereas the pan-sharpening-based methods CNMF, ICCV'15, and HySure present
relatively stable results. This finding suggests that HS-MS
fusion should be performed on the original HS features
rather than on linearly transformed features, such as principal components, if the SRF overlap is relatively small.

ROBUSTNESS ANALYSIS
The robustness of all of the methods under comparison
against noise, SRF overlap, GSD ratios, and misregistration
is systematically analyzed in the following.
noise roBUstness
Figure 16 shows the impact of denoising on the overall fusion
quality, using PSNR and SAM as quality measures. It demonstrates that the denoising postprocessing consistently
improves the performance of GSA, SFIM-HS, GLP-HS,
and MAP-SMM while neither improving nor worsening the fusion results produced by the other methods.
This observation can be explained by the fact that CNMF,
ECCV'14, ICCV'15, HySure, FUSE, and FUSE-S all operate
on a subspace, which means that the dimension of their
fusion result is only as large as the dimension of the corresponding underlying subspace. In the case of unmixingbased methods, this dimension is given by the predefined
number of endmembers. Any denoising method reduces
the dimension of the supposedly high-dimensional noisy
data to some extent. Now, subspace-based methods already
naturally imply denoising powers. Therefore, depending
on the denoising technique at hand, postdenoising does
not affect those fusion results. On the other hand, this

Gsd ratio
We examined the robustness of the ten HS-MS fusion algorithms against the GSD ratio using the Hyperspec Chikusei
data set. Figure 17(c) and (d) shows the PSNR and SAM for
the ten fused images with three cases of the GSD ratio, i.e.,
three, six, and ten. As the GSD ratio increases, the fusion
performance generally decreases. For both quality measures, the hypersharpening methods CNMF and HySure
show the most robust results.

3.5

-0.05
-0.1
-0.15

2.5
SAM (°)

PSNR (dB)

2
1.5

-0.25
-0.3
-0.35

-0.4

0.5
0

-0.2

-0.45
1

4
5
6
Data Set Number
(a)
GSA
SFIM-HS

GLP-HS
CNMF

ECCV'14
ICCV'15

-0.5

HySure
MAP-SMM

4
5
6
Data Set Number
(b)

FUSE
FUSE-S

FIGURE 16. The impact of denoising on the quality assessment of fused data using (a) PSNR and (b) SAM.

ieee Geoscience and remote sensing magazine

june 2017

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