IEEE Geoscience and Remote Sensing Magazine - September 2015 - 38

(a)

(a)

(b)

(c)

(d)

(e)

(c)

(d)

(e)

(f)

(g)

(h)

(f)

(g)

(h)

(i)

38

(b)

(j)

(i)

(j)

Figure 6. Details of original and fused Camargue dataset HS image in the visible domain. (a) reference image, (b) interpolated HS
image, (c) SFIM, (d) MTF-GLP-HPM, (e) GSA, (f) PCA, (g) GFPCA,
(h) CNMF, (i) Bayesian Sparse, (j) HySure.

Figure 7. Details of original and fused Camargue dataset HS image in the SWIR domain. (a) reference image, (b) interpolated HS
image, (c) SFIM, (d) MTF-GLP-HPM, (e) GSA, (f) PCA, (g) GFPCA,
(h) CNMF, (i) Bayesian Sparse, (j) HySure.

particular, because of the nature of the Garons dataset (village with a lot of small buildings) and the chosen ratio of
5, worse results have been obtained than for the two first
datasets since a lot of mixing is presented in the HS image.
A visual analysis of the result shows that most of the
fusion approaches considered in this paper give good results, excepted two methods: PCA and GFPCA. PCA belongs to the class of CS methods which are known to be
characterized by their high fidelity in rendering the spatial
details but their generation of significant spectral distortion. This is clearly visible in Figure 6 (f), where significant
differences of color can be observed with respect to the
reference image, in particular when examining the different fields. GFPCA here also performs poorly. Compared
with PCA, there is less spectral distortion but the included
spatial information seems to be not sufficient, since the
fused image is significantly blurred. Spatial information
provided by PCA is better since the main information of
HS image (where the spatial information is contained) is
replaced by the high spatial information contained in the

PAN image. When using GFPCA, the guided filter controls
the amount of spatial information added to the data, so
not all the spatial information may be added to avoid to
modify the spectral information too much. For the Moffett field dataset, GFPCA performs a little bit better since,
in this dataset, there is a lot of large areas. Thus blur is less
present whereas, in the Garons dataset, GFPCA performs
worse since this image consists of numerous small features,
leading to more blurring effects. As a consequence, in this
case, GFPCA performs worse than PCA.
To analyze the spectrum in detail, chosen thanks to
RMSE percentiles, some additional information about the
corresponding pixels are needed. Fig. 9 corresponds to a
pixel in the reference image which represents a red building. Since in the HS image this building is mixed with its
neighborhood, we do not have the same information between the reference image ("pure" spectrum) and the HS
image ("mixed" spectrum). Fig. 8 corresponds to a pixel in
a homogeneous field area, no mixing is present and very
good results have been obtained for all the methods. For
ieee Geoscience and remote sensing magazine

september 2015
Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2015
