has 
for 
to 
(1) 
on, 
of 
re, 
as 
rds 
ich 
les 
'ed 
are 
eir 
les 
ho 
for 
me 
ins 
cal 
les 
10t 
are 
till 
  
Peter Doucette 
3 EXPERIMENTS WITH IMAGERY 
The test image shown in fig. 4 is a 1.0m HYDICE strip flown over the U.S. Army’s Yakima Training Center (YTC), 
located near Yakima, Washington. The HYDICE aerial sensor is a push broom system that generates images with a 
spectral resolution of 210 bands within the range of 0.4-2.5um (10nm nominal bandwidth). An optimized band 
reduction technique described in Lundeen et al. (1996) was applied by the provider to condense the original image to 44 
bands. A relatively rural 300x300 pixel test area is used in this study are so indicated in fig. 4. An appropriate band is 
selected from each of the RGB regions to simulate a true color image. 
  
     
Figure 4. Selected bands from a 1.0m HYDICE strip with 300x300 pixel test area outlined. 
  
3.1 Spectral Classification 
Supervised spectral classification is a relatively time consuming image processing step. The trade-off for a time- 
economical approach is generally somewhat higher noise levels in the classification results. The idea was to input the 
results from a relatively ‘low-cost’ spectral classification of the image into our road vectorization algorithm to 
determine how robustly it could perform. A maximum likelihood classification (MLC) was performed on the first three 
layers from a PCA transformation of the 44-layer image. The eigenvalues indicated an accounting of 99% of the image 
variance in the first three PCs. Fig. 5 shows the spectral classification results for the paved road class that is used as 
input for our vectorization algorithm. Presence of both clutter and omission noise is apparent. The former is due to 
spectrally inseparable samples of rooftops and driveways, and the latter is largely attributable to a decrease in feature 
homogeneity with higher spatial and spectral resolution. 
  
— 
y ud 
  
  
  
Figure 5. Paved road class pixels from unedited maximum likelihood classification 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 249