IAPRS & SIS, Vol.34, Part 7, *Resource and Environmental Monitoring", Hyderabad, India,2002 
  
make this area an ideal test site for evaluation of the impact of 
data compression on mineral abundance maps. 
3.0 DATA PROCESSING 
An outline of the data processing is given in Figure 1. The 
different steps include data compression and de-compression, 
atmospheric correction and post-processing of spectra, 
endmember selection and spectral unmixing, and fidelity 
assessment. The original data and de- compressed data cubes 
are processed separately using the same data processing 
techniques. 
  
HSOCVO Dara 
Compression and 
Data Decompression 
  
  
  
       
   
  
        
     
   
AWIRIS Radiance 
Data 
de-compressed 
AVIRIS Radiance 
Data 
orginal 
  
Atmospheric Correction 
Post-Processing 
and 
orginal i i de-compressed 
Surface 
Ketlectance Data 
== 
Endmember Selection 
  
  
    
  
   
  
  
and 
Spectral Unmixing 
original | ; de-compressed 
Endmembhers, 
Fraction Maps 
| ed 
Fidelity 
Assessment 
  
  
  
        
  
  
  
  
Figure !. Data processing scheme 
3.1 Data Compression 
The AVIRIS radiance data cube was compressed by factors of 
10, 20, and 40 with the HSOCVQ, which is a VQ based data 
compression technique with a characteristic of self-organising 
clusters. This technique clusters spectra in the data cube while 
it compresses them. It applies VQ to each of the clusters and 
ensures each of the vector-quantized clusters to have a fidelity 
better than the given threshold. If the fidelity of the cluster is 
not better than the threshold, the spectra in the cluster are 
adaptively re-clustered. As a result of the unique way of 
clustering, the code vectors trained in this process are very fast 
and efficient. High reconstruction fidelity can be attained with a 
relatively small codebook. One of the unique features of this 
technique is the guarantee that the reconstruction fidelity of 
each spectrum in the compressed data cube is better than the 
threshold. This feature allows HSOCVQ to preserve spectral 
signatures of small targets in the scene of a hyperspectral data 
cube. Subsequently, the compressed data were de-compressed 
in order to reconstruct the cubes using the codebooks generated 
for each cluster. 
3.2 Surface Reflectance Retrieval 
Prior to the correction of the atmospheric effects, the 
wavelengths covering the strong atmospheric water absorption 
regions at 1380 nm and 1870 nm were eliminated from further 
processing due to the dominance of noise in these areas. For the 
same reason, the first six bands and the last five bands were 
also excluded resulting in a reduced wavelength coverage of 
428 nm to 2458 nm. 
Surface reflectances were then computed from at-sénsor 
radiance (original) data and de- compressed data cubes, 
compensating for atmospheric absorption and scattering effects. 
The procedure is based on a look-up table (LUT) approach with 
tunable breakpoints as described in Staenz and Williams 
(1997), to reduce significantly the number of radiative transfer 
(RT) code runs. MODTRANA.2 was used in forward mode to 
generate the radiance LUTs, one each for a 5% and 60% 
reflectance. These LUTs were produced for five pixel locations 
equally spaced across the swath, including nadir and swath 
edges, for a range of water vapour contents, and for single 
values of aerosol optical depth (horizontal visibility) and terrain 
elevation. The specification of these parameters and others 
required for input into the MODTRAN4.2 RT code are listed in 
Table 1. 
Atmospheric model US standard 76 
Aerosol model Desert 
Date of overflicht June 12, 1996 
Solar zenith J 158° 
Solar azimuth le 183.7“ 
Sensor zénith angle Variable 
Sensor azimuth angle Variable 
Terraim elevation (above sea level) [524 km 
Sensor gititude (above sea level) 20.100 km 
Water vapour content. variable 
Ozone column as model 
CO mixing ratio 300 ppm 
Horizontzi visibility 50 km 
  
Waveleneth erid interval | em 
Table |. Input parameters for MODTRANA code runs 
    
  
   
   
    
   
  
  
  
     
   
   
    
   
     
    
   
    
  
  
   
     
     
     
    
   
    
    
   
   
  
  
  
  
   
    
For th 
adjust: 
vapoui 
al, IS 
were 
vapou 
an iter 
surfac 
descril 
The n: 
irregul 
may h 
from t 
selecti 
errors 
using 
of the: 
3.3 Er 
Endme 
from t 
Iterati 
step, t 
data s 
produ 
distan 
the av 
calcul: 
spectr 
find tl 
that a 
selects 
first eı 
averag 
, find tk 
the se 
numbe 
In this 
Once 
unmix 
and Sı 
linear 
compc 
1) for 
AVIR 
selecti 
4.0 F 
The : 
compr 
levels. 
betwe! 
(scale: 
RAS 
where 
DNo | 
numbe 
cube, 
pixel €