The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1108
question: how can the GLP-CBD QuickBird fusion result
published in the contest outcome paper in 2007 appear
significantly better than that submitted to the contest in 2006?
UNB fusion result (0.7m)
UNB fusion result (0.7m)
Original Pan (0.7m)
Original Pan (0.7m)
Figure 4. Subsets of the QuickBird fusion results of UNB-Pansharp submitted to the IEEE GRSS 2006 Data Fusion Contest (UNB-
Pansharp can produce fusion results either with or without feature enhancement. The fusion results with feature enhancement were
submitted to the contest. All images in this figure are displayed under the same image stretching condition.)
The inconsistency and irregularity in the evaluation of IEEE
GRSS 2006 Data Fusion Contest also raised the question on the
capacity of the seven quantitative indicators (MB, VD, SDD,
CC, SAM, ERGAS, and Q4) for quality measurements between
images. 6
6. CONCLUSIONS
This paper analyzed and evaluated three cases of image quality
comparisons using visual and quantitative methods. The three
cases are (1) visual and quantitative analysis of the four testing
images generated for this study; (2) review and analysis of the
fusion quality evaluation by Alparone et al.,(2004), which
received the 2004 IEEE Geoscience and Remote Sensing Letter
Best Paper Award (Alparone et ah, 2007); and (3) review and
analysis of the evaluation of the IEEE GRSS 2006 data fusion
contest. The quantitative methods evaluated are the seven
frequently used indicators—Mean Bias (MB), Variance
Difference (VD), Standard Deviation Difference (SDD),
Correlation Coefficient (CC), Spectral Angle Mapper (SAM),
Relative Dimensionless Global Error (ERGAS); Q4 Quality
Index (Q4)—which are also the quantitative measures of the
IEEE GRSS 2006 Data Fusion Contest.
In the visual and quantitative analysis of the four testing images
generated for this study, it was found:
• The four testing images generated through mean
shifting and/or histogram stretching provide the same
visualization and classification results under the same
display and classification conditions. This demonstrates
that mean shifting and histogram stretching (within the