1102
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
image display. Therefore, one cannot conclude that one
image is better than another if the display condition is
not the same. Unfortunately, no display conditions were
clearly described in many visual comparisons, including
those in the IEEE GRSS 2006 Data Fusion Contest. This
ambiguity in display conditions significantly reduced the
reliability of the visual comparison results.
• For quantitative evaluation, different evaluation
results can often be obtained when different quantitative
measures or indicators are selected for the evaluation.
Therefore, whether, or not, a given quantitative index
can measure image fusion quality or measure quality
difference between two images is still an open question.
Among numerous quantitative evaluation indicators, the
Mean Bias (MB), Variance Difference (VD), Standard
Deviation Difference (SDD), Correlation Coefficient
(CC), Spectral Angle Mapper (SAM), Relative
Dimensionless Global Error (ERGAS), and Q4 Quality
Index (Q4) have been often used in image fusion
publications. They were also used in the IEEE GRSS
2006 Data Fusion Contest for quantitative evaluation.
Therefore, this paper focuses on the evaluation and discussion
of how display conditions affect visual comparison and whether,
or not, the seven often used quantitative indicators (MB, VD,
SDD, CC, SAM, ERGAS, and Q4) can provide convincing
results to tell the quality difference or similarity of two images.
This evaluation is conducted based on the assumption that
(1) if two images of the same area can present
identical information, including colour, spatial detail and
image depth, under the same visualization condition, and
(2) if the two images can also provide the same
classification result using the same classifier under the
same processing condition,
the two images can be defined and accepted as having the same
image quality.
This assumption is true for remote sensing imagery and remote
sensing applications, because the two foremost important
applications of remote sensing imagery are (1) visualization and
(2) classification. If two images can provide the same results for
visualization and classification under the same condition, they
will not make any difference for remote sensing applications,
and they can be equally accepted by remote sensing users.
For the evaluation and discussion, some testing images having
the same image quality are generated; the seven quality
indicators are applied to the testing images to check their ability
to measure the quality similarity among the images; and the
fusion quality evaluations by Alparone, et al. (2004) and
Alparone, et al. (2007) are reviewed and analyzed to see
whether, or not, the quality indicators of the evaluations
provided convincing results.
2. TESTING IMAGES
An original Ikonos MS image of Fredericton, NB, Canada,
collected on October 1, 2001, is used for the evaluation. The
image contains 4 spectral bands and is stored in 16 bits. For
visual comparison purpose and to test the performance of the
seven quantitative indicators, the original Ikonos image (Ik-Orig)
is altered through mean shifting, histogram stretching, and
histogram stretching plus mean shifting, resulting in a mean
shifted Ikonos image (Ik-Shift), a histogram stretched image
(Ik-Str), and a histogram stretched and mean shifted image (Ik-
Str-Shift). The detailed alteration of the Ikonos image is
described in Table 1.
2.1 Visual comparison
To prove that the four images (Ik-Orig, Ik-Shift, Ik-Str, and Ik-
Str-Shift) have the same image quality for visualization, they
are displayed under the same display conditions and compared
with each other. The histogram stretchings used are zero
stretching (i.e. no stretching), linear stretching, root stretching,
adaptive stretching, and equalization stretching (Figure 1). It
can be seen that all of the four images appear very dark without
any histogram stretching. And, all of the four images appear
exactly the same when they are stretched using the same
histogram stretch, regardless what stretch is applied (compare
images in the same column of Figure 1). This comparison
demonstrates that the four images have the same quality for
visualization and visual interpretation.
On the other hand, it can also be seen from Figure 1 that the
same image can be displayed and interpreted differently as if the
source image had different qualities, if the image is not
displayed under the same condition. For example, the same
original Ikonos image (Ik-Orig) in Figure 1 appears
significantly differently under different display conditions.
Some appear darker than others, and some look noisier than
others. If the image source information and the image stretching
information were not given in Figure 1, one must say that the
images in different columns of Figure 1 have different qualities.
Ik-Orig
Ik-Shift
Ik-Str
Ik-Str-Shift
Band 1
B
B+100
Bxl.5
Bxl.5+100
Band 2
G
G+100
Gxl.5
Gxl.5+100
Band 3
R
R+100
Rxl.5
Rx 1.5+100
Band 4
NIR
NIR+100
NIRxl.5
NIRx 1.5+100
Table 1. Alteration of the spectral bands of the original Ikonos MS (Ik-Orig) to obtain other testing images
(Ik-Shift, Ik-Str, and Ik-Str-Shift)