and the 
ic image 
  
  
  
cated on the straight line which is also plotted for orien- 
tation. The correlation values and the mean deviation of 
the fusion-sharpened from the true 1m reflectance val- 
ues are given in Table 2. 
While the correlation between both images is very good 
(97%) in the red and green bands, the near infrared 
seems to be less correlated (93%) for the simple fusion al- 
gorithm. In the scatter plot (Fig. 4, top left) it is obvious 
that there exist different spectral clusters which need 
specific alignment factors in order to preserve the mean 
reflectance value of the original MS4m-image. Applica- 
tion of the class-specific fusion algorithm (section 3.4) 
then indeed improves the spectral truth significantly 
(see Fig. 4, lower half, and Table 2). 
4.2 Normalized Difference Vegetation Index 
The normalized difference vegetation index (NDVI) is a 
feature of great importance for vegetation monitoring. It 
is defined as 
NIR — R 
DVS a Ri br 
where NIR is the near infrared band and R is the red 
band. It is obvious that the NDVI depends only on the 
relative relation of the spectral reflectance values and 
is thus invariant to scaling factors applied to all spe- 
tral bands. The panchromatic sharpening which pre- 
serves the relative spectral contributions thus does not 
affect the NDVI values. The NDVI depends on the chro- 
matic information only, and cannot profit from additional 
panchromatic data of higher resolution. Still, the com- 
parison of the NDVI values of the coarse 4m MSam- 
image and the fine 1m MS,m-image shows a correlation 
of > 90% (Table 2). 
4.3 Local Spectral Variance 
Particularly for the classification of vegetation it is use- 
ful to consider texture features. One of the simplest tex- 
ture features is the spectral variance in the local neigh- 
borhood A/(z) of a pixel x. Here we compute the root 
mean variance from all neighboring pixels z/ € A/(x) 
around x as 
Js X 3. s(e' — al) [ri(a’) — (ro)? 
i e'CA/(z) 
  
for the three spectral bands 4 € (NIR,R,G), where 
(ri(z')) is the weighted mean value in the neighbor- 
hood N(x), using a Gaussian weighting g(|zr' — z|) 
as a function of the distance |z' — x| of each neighbor- 
ing pixel x’ to the central pixel =, and a normalization 
factor N = Zaren) (lr. — z|). For the distance 
weighting we choose a Gaussian distribution width of & 
= 1.83 pixel, so that the influence is vanishing outside an 
11 x 11 pixel window. 
As an example, the local spectral variance is high for 
forest areas, lower for meadows, and low for water and 
smooth artificial surfaces such as parking lots. In gen- 
eral the texture is highly correlated through the spectral 
bands, and thus captured well in the panchromatic band. 
Therefore we find that this feature improves dramati- 
cally by the fusion-sharpening (Table 2). After fusion, 
the correlation with the true, 1 m resolved local spectral 
variance is > 90%. 
Simple Fusion 
Near Infrared : Red 
  
Reflectance Fused Image 
Reflectance Fused Image 
  
  
0.02 04 06 08 1.0 
Reflectance Original Image 
  
Green 
Reflectance Fused Image 
  
00 02 04 06 08 10 
Reflectance Original Image 
  
  
  
Fusion Using Unsupervised Clustering (16 Classes) 
  
  
   
   
  
Near Infrared Red 
1.0 
o o 
S 08 S 
tt E 
© © 
$ 0.6 $ 
2 = 
u u. 
[] o 
2 04 € 
f s 
3 © 
= 0.2 = 
a a 
0.0 0.0 
00 02 04 06 08 1.0 0.0-02..04 086 08 10 
Reflectance Original Image Reflectance Original Image 
Green 
1.0 
o 
9 08 
En 
© 
$ 06 
2 
u 
9 
2 04 
S8 
® 
e 0.2 
oc 
0.0 : 
00 02 04 06-08 10 
Reflectance Original Image 
  
  
  
Figure 4: Scatter plots comparing the reflectance values of the 
true 1m MS,m-image and the panchromatic-sharpened multi- 
spectral image for two fusion algorithms (Sections 3.3 and 3.4). 
The corresponding correlation values are given in Table 2. 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 289