International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
   
     
    
7 2 
Figure 4. Fourth MNF band of a hyperspectral image. 
The outcomes of  parallel- pipe analysis are 
preliminary classification results. A quick 
verification with known ground truth samples 
indicates that the overall accuracies are 60% to 70%. 
At first sight, this may not seem adequate. Further 
investigation reveals that the low accuracy rates 
stem from two primary sources. As of an example 
shown in Fig. 5, most of the target of interest within 
the polygons created from previous field 
investigation have been successfully identified 
during the spectral analysis phase. However, there 
seem to be a lot of commission errors. Part of the 
reason for this is because the GIS layer used as 
ground truth samples does not cover the entire 
study area. In addition, the creation dates of the GIS 
data are at least one year old and the target of 
interest is spreading very fast in the study site. 
Figure 5. Preliminary result (red) from spectral 
analysis and known ground truth data (blue). 
74 
Another source of the errors comes from the fact 
that Leucaena Leucocephala may exhibit different 
spectral patterns due to intrinsic characteristic 
differences. Indeed, there are more than one kind 
(species or sub- species) of Leucaena Leucocephala 
exist within the study area. Consequently, the 
variations may cause significant confusion to 
spectral classifiers. 
Albeit all of these, the accuracy can still be 
improved, especially for errors from the second 
source (intrinsic characteristic variations). The 
approach employed in this study to remedy this 
defect is via texture analysis of high resolution 
images as described below. 
3.2 Analysis of High Resolution Images 
Detail examination of the classification results 
produced in the first analysis phase indicates that 
spectral classifiers did a poor job in recognizing 
Leucaena Leucocephala with varied characteristics in 
spectral domains. A primary reason for this defect is 
because the hyperspectral image has a very broad 
ground resolution (30m) comparing to the physical 
sizes of the tree and canopy structures of the target 
to identify. Therefore, introducing spatial analysis of 
high resolution data should be a right approach to 
solve this problem. Accordingly, texture analysis 
procedures of high resolution images were 
employed to improve the classification results in 
spots where spectral analysis failed to produce 
accurate outcomes. 
    
   
Pi 
Figure 6. Texture of target plant in high resolution 
images. 
An example of such texture analysis is shown in Fig. 
6. The greens in this image are in fact all Leucaena 
Leucocephala. Because of the difference in 
population density, trunk heights, and perhaps 
canopy structures or health statuses, they appear to 
have different degrees of greeness. This may cause a 
spectral analyzer to classify them as different 
classes. However, in spite of differences in spectral 
Internatio 
VIE EN? 
appearat 
the sam 
are very 
as the f 
same in 
texture 
accuracy 
increase: 
3.3 F 
First of 
analysis 
accompl 
experien 
will be 
achieve 
analysis 
been ev 
significa 
the enti 
become 
texture 
algorithi 
consider 
impleme 
Secondl 
MNF tre 
reasonal 
improve 
shall be 
creating 
analysis 
texture 
order t 
system. 
The im 
signific: 
biodivei 
situatio 
deterior 
an acct 
spreadi 
This s 
spectral 
texture 
an effi 
specific 
systeme 
applies 
hypers[ 
spectra 
further 
resoluti 
that th 
(spectre 
accurac 
from lc 
Park an