The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1229
D _ ^Li v n
U MBTi ~ n D H
2Л
/■=1
images, А/, is the mean radiance value of original image and
(3) RMSE(Bj) is defined as:
RMSB,B i ) = •¡(M i - FU§) 2 (6)
Where n is total number of LSR images.
3. EXPERIMENTT RESULTS AND ANALYSIS
In this section, we consider two different data sets acquired by
Quickbird and IKONOS satellite sensors, respectively. The
purpose of selecting two types of sensors is to demonstrate that
selected fusion methods work for both very high spatial
resolution image. In order to assess the effectiveness of the
proposed method, МВТ is compared with the fusion methods
which can also fuse either individual band LSR images, such as
SFIM and DWT. Three fusion methods are all achieved on the
ERDAS IMAGE 8.7 platform. SFIM and МВТ are achieved
under spatial modeler. DWT based on ERDAS IMAGE 8.7
platform is a modification of the work of King (King et al,2001)
with extensive input from Lemeshewsky (Lemeshewsky, 1999a;
Lemeshewsky,2002b). The spectral quality and spatial quality
of the fused images will be evaluated by the following indexes.
3.1 Information Preservation Evaluation
3.1.1 Spectral Quality of the Fusion Methods: The fused
images gain spectral information from LSR images, so, the
spectral quality of the fusion methods will be evaluated by the
comparing their spectral quality with that of the original LSR
images. The comparing is performed by the following indexes.
Spectrum Difference (SD) between original LSR images and
the fused images.
SD =
1 ^ "LSR jk -FUS jk
MNjih LSR jk
(4)
Where FUSi represents the radiance of the /th band fused image.
The lower value of SD and spectral ERG AS indexes, the higher
the spectral quality of the merged images. The values of SD and
ERGAS indexes are zero in ideal condition, and fusion method
has optimal spectral information preservation when the values
of SD and ERGAS indexes are closer to zero.
3.1.2 Spatial Quality of the Fusion Methods: The fused
images gain spatial information from HSR image, and the
spatial quality of the fusion methods will be evaluated by the
relationship between the fused images with HSR image using
the indicators of Correlation Coefficient (CC) and spatial
ERGAS (Lillo-Saavedra et al,2005). CC is got by:
M N
> уj )- f) №,, уj)-o)
cc=
'■=1 J
(7)
M N
YY.(m,y,)-n 2 (o(x„y,)-öf
M j=\
Where F(x, y) and 0(x, y) are the pixel grey values of fused and
HSR images respectively, f is the mean value of fused image,
and O is the mean value of HSR image.
Spatial ERGAS is similar with spectral ERGAS. A spatial RMSE
index is included in spatial ERGAS definition, and spatial
RMSE is got by:
-M HSR f +(SD n -SD HSR f (8)
Where MN is the total number of pixels of original and fused
images, FUS jk is the radiance value of pixel j in the Afii band of
the fused image, and LSR jk is the radiance value of pixel j in the
Mi band of original LSR image.
The index SD can only evaluate the spectral quality of
individual band. The spectral ERGAS (Wald,2000) index has
been selected to estimate the global spectral quality of the
fusion methods, and spectral ERGAS is defined as:
Where M Fi and M HS r are the means of fused image and HSR
image respectively, SD Fi and SD hsr are, respectively, standard
deviation of fused image and HSR image.
Higher CC or lower spatial ERGAS values imply that fusion
method has higher spatial quality. In ideal condition, the values
of CC and spatial ERGAS are 1 and zero, respectively. Fusion
method has optimal spatial information preservation with HSR
image when the values of CC are high and the values of spatial
ERGAS are lower.
ERGAS = 100
h 1
y J-ZC
I V n M
RMSE{B i Y
M 2
)
(5)
Where h is the spatial resolution of HSR image, / is the spatial
resolution of LSR image, n is the total number of the fused
3.2 Results on Quickbird Images
A scene of Quickbird images (1600x1600 pixels for HSR image)
taken on 21 Nov.2002 were selected as one test data in this
paper. The test region of Quickbird covers urban areas of
Sundarbans in India. Small areas (308x152 pixels for LRS
image) of original HSR and LSR images are shown in Fig. 3.