Comparison between Fused Image and Full-Resolution Image

Although the result of the fusion looks very satisfactory to the eye, its spectral truth remains to be checked quantitatively.

When the real satellite imagery becomes available, data fusion in order to yield high resolved color images will certainly be performed. However, the accuracy of the estimated color values will remain unknown, and thus also the errors which propagate through later image processing steps such as land cover classification, change detection, NDVI computation, etc.

In contrast, with the airborne / simulated imagery we are in a position to immediately check the deviation between the multispectral imagery which is truly sampled with 1 m ground resolution and the one interpolated from 4 m resolution by data fusion with 1 m panchromatic resolution. The following quantitative comparisons can be made:

• We can directly regress the true intensity or reflectance values against the fusion-estimated values for each spectral band and determine the correlations.
• We can determine the root mean square deviation in reflectance between true and fusion-estimated values.
• The NDVI-values (normalized difference vegetation index, used for vegetation health monitoring) can be computed and compared to the full-resolution imagery.
• The true and the fusion-derived imagery can be classified into land cover classes [Richards (1993)]. Between the two results, a confusion matrix can be calculated and a statistical indicator of agreement ('Kappa coefficient', [Congalton (1991)]) on the overall classification accuracy of the fusion-estimated imagery in comparison to the true 1 m imagery.

In the following, the results of the above comparisons are described.

 
• Correlation between Original and Fused Image
• Normalized Difference Vegetation Index