4. PRE-PROCESSING OF MULTIPLE IMAGES 
Multiple images are either (a) taken simultaneously in various 
spectral bands (MSS), with different polarizations, with different sensors, 
or (b) taken sequentially. In addition there are manipulations dealing 
with images and combining them with non-image data (or synthetic images), 
Appropriate pre-processing of the individual images may precede the work 
with multiple images. We group the various methods according to their 
applications as follows: 
- image registration; 
- data compression and enhancements} 
- data display. 
The digital classification and clustering methods generally concern auto- 
matic image analysis and therefore are considered to fall outside the 
Scope of this review. 
4.1 Image Registration 
This is & technique of matching overlapping images (or natural and 
Synthetic images) that differ both geometrically and radiometrically. 
Image registration may be achieved much in the same way as geometric 
rectification: homologue features are identified in both images and two 
pairs of coordinates are measured for each feature: X,y in image 1 
("master", "search" or "reference" image) and X,Y in image 2 ("slave" or 
"input" image). The two sets of coordinates for each feature define a 
geometric transformation between the two images: 
x = g(X,Y) 
¥ = h(X,Y) o) 
Often, this transformation (1) is denoted by rubber —sheet stretch or 
warping function . 
The identification of homologue features may be manual or automatic; 
the latter methods fall into two categories: 
(a) sequential similarity detection algorithms (SSDA), 
(b) correlation methods. 
SSDAs are less expensive than correlation methods (Barnea et al., 1972). 
The latter can be carried out most economically using Fast Fourier Trans- 
forms (FFT) methods (Anuta, 1970). Even then they are more expensive than 
similarity measures. However, correlation methods are statistically more 
satisfying. 
Details of registration procedures are not to be discussed in the pre- 
sent context since they are not important for interpretation. However, the 
procedures enable one (a) to join overlapping images into mosaics,(b) to detect 
Stereo parallaxes automatically, (c) to composite an MSS-image with N bands 
and one with M bands into an image with N + M bands, (d) to merge multi- 
temporal images for change detection. There is fairly clear evidence that 
the time dimension is very important in many interpretation tasks: multi- 
temporal images with few spectral bands may be more useful than images with 
& large number of spectral bands all taken at the same time. In the former 
there is less redundancy than in the latter, because of dynamics in seasonal 
vegetation, hydrology and others. 
Registration of images with non-image data, e.g. of digital terrain 
  
     
   
   
   
   
   
    
    
    
    
   
     
    
    
     
   
    
    
   
  
  
  
  
  
  
       
   
r 
o 
T 
& 
  
as 
SU 
si 
nc 
wi 
pe 
mu 
..