04 
late 
ple 
hey 
ber 
zed 
zed 
asy 
[he 
the 
on- 
ion 
the 
ush 
and 
and 
for 
age 
  
A0 304 
  
S 
o 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
The only interaction that is still required is the NDVI threshold 
selection for the required feature classes. Although some 
procedures exist to develop automated threshold selection based 
on local minima detection, we decided to use an interactive 
process so that the user can immediately see the selected classes 
on top of the image data. Mistakes and erroneous thresholds can 
be interactively corrected. Once the selected NDVI classes have 
been verified by visual analysis, the image is separated into 
independent layers and processed similar to the previous 
process based on the GIS input. 
Table 1 presents the NDVI thresholds and the respected feature 
classes. Figure 9 shows the NDVI masks and Figure 10 the 
result of the feature based enhancement process. The steps are 
the same as those of chapter 3.1. It should be noted that for the 
water areas, another band combination (3, 2, 1) was employed 
for better feature separation. 
  
  
  
  
  
  
  
Class NDVI Value 
Water NDVI x -0.12 
Open/Beach -0.12 « NDVI x 0.00 
Open/Inland/Built-up 0.00 « NDVI x 0.19 
Vegetation 0.19 « NDVI 
  
  
Table 1. Selected enhancement classes with NDVI values 
  
Figure 9. Selected NDVI classes (pseudo color coded) 
For a better comparison, Figure 11 presents the same subset as 
shown in Figure 7. Again, the level of detail demonstrates the 
superiority of the local enhancement procedure. 
401 
  
Figure 10. NDVI enhanced Quickbird image 
  
Figure 11. The subset of the NDVI enhanced image (left) 
shows a higher level of detail compared to the globally 
enhanced image (right) (contrast stretch with +20) 
4. CONCLUSIONS 
The result of the NDVI based enhancement seems almost better 
than the one that is based on GIS information. The reason is that 
the selected contrast enhancement is based on the underlying 
image information. At large magnifications, however, 
discontinuities in the selected classes become visible. In 
contrast to GIS feature classes, NDVI classes are not 
contiguous and may contain single pixels and small pixel 
groups that are differently enhanced than their neighbors. This 
can result in image noise in certain areas. Image processing 
such as filtering or blow and shrink operations may be 
employed to create more contiguous image masks. At standard 
resolutions, however, this method shows results that prove the 
validity of the presented approach. 
Both procedures work well for the display of multispectral 
images. As individual band selection can be incorporated in this 
enhancement process, the extension to rapid hyperspectral 
image display is possible. Known optimum band selection can 
be combined with spectral enhancement in this procedure. The 
method can also be automated to a large degree using a flow 
chart environment or scripting language. With more 
investigations in the future; some of the interactive steps will be 
replaced by default values.