International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
4. VALIDATION OF FUSION ALGORITHMS 
THROUGH SIMULATED DATASETS 
To verify the radiometric fidelity of the approach, a systematic 
validation was carried out based on image data which were 
acquired with the airborne version of the MOMS imaging 
system (DPA) in the framework of the MOMS-2P research 
program. In July 1997, a high resolution multispectral dataset 
was recorded at one of our forest study sites (Hillesheim, Eifel 
Mountains). We produced a validation dataset that corresponds 
to the geometric resolution configuration of IRS-ID imagery 
(Figure 3, above). Starting from a dataset comprising all 
spectral bands with 5m spatial resolution we produced a 
degraded multispectral dataset with a pixel size of 25m. 
Additionally, a panchromatic image with 5m resolution was 
generated by averaging all spectral bands (the NIR channel was 
included with respect to the spectral range of some existing and 
future panchromatic channels which also cover substantial parts 
of the NIR range). 
As a reference, we implemented standard fusion algorithms 
such like multiplication, Brovey, IHS, PCA, HFA, HFM 
(Sparkle) as well as more complex methods such as LUT 
recoding and Wavelet Transform. The Wavelet Transform was 
implemented under IDL 5.0, according to the algorithm of 
Garguet-Duport et al. (1996). We would like to mention that 
our wavelet fusion algorithm is only of preliminary nature, 
therefore the result might not be fully representative for this 
approach. However, in view of the manifold possibilities to use 
the wavelet domain for data fusion, a unique implementation 
does not exist. 
Our validation strategy permits to directly compare the results 
of fusion algorithms with the true high resolution multispectral 
images, such that a rigorous quality assessment can be achieved. 
The assessment permits a visual comparison as well as 
quantitative verification approaches, such as the comparison of 
the grey value distribution (histogram analysis), the calculation 
of the global correlation, global regression, RMS-error and 
difference images between fused and true images (Wald et al., 
1997). Particularly, the difference images allow the analysis of 
the fusion accuracy in relation to spatial structures associated to 
specific landcover types. 
To validate the quality of the LCM approach regardless of any 
substitution algorithm (see above), we have created a mask 
containing all areas with acceptable local correlation levels 
between the degraded panchromatic band and the multispectral 
channels (r > 0.66). All quantitative assessments presented here 
are derived from this mask which, in case of the NIR channel, 
contains 54% of the image and includes the major part of the 
forest. For the visible bands, more than 80% of the image is 
included. 
5. RESULTS 
Our validation indicates that the LCM approach performs 
significantly better than all other techniques included in this 
study. Especially the result for the NIR band, which is generally 
considered the most problematic channel for image fusion, 
confirms that this method reconstructs better the radiometric 
properties of the true image. The analysis of the global 
correlation between restored and true images (Table 2), for 
example, produces correlation coefficients of 0.975 (blue), 
0.989 (green), 0.973 (red) and 0.946 (NIR) for the LCM 
approach. The HFM algorithm and the LUT technique show 
better than average results as well, while the frequently used 
component substitution techniques (IHS, PCS) and the Brovey 
algorithm achieve only moderate results. The observed trend 
applies not only to the correlation and RMS-error (Table 2), but 
also to histogram parameters, global regression coefficients and 
difference images, which cannot be presented in this paper due 
to the page limitation. 
The quality of the restored high resolution image using the 
LCM approach becomes also visually apparent. Figure 3 shows 
enlarged forest subsets of the LCM result, compared to the true 
image, the fusion input and the HFM, Brovey and IHS results. 
Because the introduced texture is locally adjusted to the 
Fusion technique and implementation references Correlation RMS error 
ch 1 
ch 2 
ch 3 
ch 4 
ch 1 
ch 2 
ch 3 
ch 4 
Multiplication (Filiberti et al., 1994) 
0.944 
0.959 
0.941 
0.860 
2.499 
3.673 
4.120 
12.750 
Brovey (Roller and Cox, 1980; Vrabel, 1996) 
0.820 
0.927 
0.920 
0.922 
4.345 
4.520 
4.778 
9.663 
IHS transformation (Haydn et al., 1982)* 
0.791 
0.933 
0.924 
0.874 
5.784 
5.424 
5.410 
13.340 
PCS (Shettigara, 1992) 
0.893 
0.910 
0.888 
0.912 
3.287 
4.909 
5.438 
11.146 
HFA (Schowengerdt, 1980) 
0.932 
0.975 
0.940 
0.926 
2.717 
2.576 
4.049 
9.059 
HFM (Filiberti et al., 1994; Vrabel, 1996) 
0.955 
0.989 
0.968 
0.935 
2.580 
2.128 
3.216 
8.696 
Wavelet Transform (Garguet-Duport et al., 1996)** 
0.935 
0.953 
0.934 
0.873 
2.739 
3.916 
4.445 
11.911 
LUT recoding (Price, 1987) 
0.960 
0.986 
0.950 
0.943 
2.094 
1.988 
3.984 
8.563 
LCM (Tom, 1986; Hill et al., 1998) 
0.975 
0.989 
0.973 
0.946 
1.667 
1.774 
2.769 
8.006 
* channel 1-3 via IHS 321, channel 4 via IHS 421 ** preliminary results 
Table 2. Fusion result versus true image: correlation and RMS-error (calculation window = area with local correlation between 
degraded panchromatic band and multispectral channels > 0.66).