International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004 
  
of values that topographically characterize classes of interest. 
Selection of class topographical signatures is aimed just the same 
as selection of class spectral signatures; however, selection of 
topographical signatures is rather different than selection of 
spectral signatures because, they cannot be collected via visual 
interpretation. Topographical samples are gathered by selecting 
the elevation and slope pixels that spatially correspond to nimage 
pixels satisfying the ranges of values characterizing the spectral 
signatures. The reason for using the pixels satisfying the minimum 
and maximum ranges for spectral signatures instead of the original 
spectral training set was the need for collecting unbiased samples 
that better represent the topographical distribution. 
Frequency histogram is a valuable supplement in defining 
elevation or slope ranges where classes were most likely to occur. 
Data ranges representing class topographical signatures were 
determined with the help of histogram graphics. Histograms were 
truncated by removing the observations at the two tails of the 
histogram so as to exclude deviated region of the distribution 
profile. By this way, minimum and maximum ranges for 
topographical attributes associated with four classes were 
statistically refined. Box plot of the elevation (Figure 3) and slope 
(Figure 4) point up the different ranges of elevation and slope that 
characterize the land cover classes. 
Fourth phase: Redefinition or adjustment of training sets in this 
phase is critical. The effect of ancillary topographical parameters 
on classification accuracy is tested by means of two products; one 
is derived from spectral data and the other from both spectral and 
the topographical data. This is accomplished by classifying the 
multispectral image data by training set involving class spectral 
signatures only, to yield Product 1 (Pl) and; classifying 
multispectral image data and topographical data by means of 
training set involving both class spectral and topographical 
signatures to yield Product 2 (P2). Also a third product is 
generated (P3) to represent a conventional logical channel 
approach where multispectral data and topographical data are 
classified by means of training set involving class spectral 
signatures only. 
Two training sets were generated to satisfy the afore mentioned 
criteria; Training Set 1 (T1); involving class spectral signatures 
only and Training Set 2 (T2); involving both class spectral and 
class topographical signatures. 
The question is “is it possible to manually select training samples 
that would also represent topographical signatures, without 
deforming the class spectral signatures?” Answer to this question 
is possibly no, because collecting samples that can satisfy 
topographical signatures and do not change the characteristics of 
spectral signatures is manually impractical. Therefore an 
automated selection procedure was adopted. 
In order to implement automated selection, all of the samples were 
transferred to a database table and two queries one of which is for 
T1 and other for T2 were performed with respect to minimum and 
maximum ranges previously defined both for spectral and 
topographical signatures. This yielded two training sets TI and T2 
with class spectral statistics, mean and variance almost identical 
where; T2 represents topographical signatures as well. I f this was 
not achieved, it would be hard to state that the difference in 
between Product 1 and Product 2 is due to topographical effect. 
56 
  
1400 +— : c: AA 
1300 4 
  
  
  
  
  
  
  
  
  
  
d cT 
  
  
  
  
  
  
  
  
  
  
i sab 
1000 4 - 
= 
= 
S 
> 
o 
uj 900 
| Agriculture Range-shrub Range-herb Forest 
Figure 3: Box plot of elevation signatures data range 
50 
40 4 
Ss A | 
  
  
  
  
  
x. 
EET 
Agriculture Range-shrub Range-herb Forest 
  
  
  
  
Slope (degrees) 
  
  
  
  
  
  
Figure 4: Box plot of slope signatures data range 
Fifth phase: Maximum likelihood classification is performed to 
yield P1 (Figure 5), which is the result of classification of spectral 
data by means of Training 1 (Training set for spectral data only), 
second to yield P2 (Figure 6), which is the result of classification 
of both spectral and topographical data by means of Training 2 
(Training set for spectral and topographical data) and to yield P3, 
which is the result of classification of spectral data and additional 
topographical data by means of Training 1. 
Accuracy Assessment 
A certain amount of difference is identified between the products. 
However to understand the precise amount of disparity between 
the products, and their association with the real world; accuracy 
assessment of the products are needed. 
Error matrix is an effective way to represent the accuracy of 
classification; it provides both inclusion (commission error) and 
exclusion (omission error) for each class. Products were tested 
with the ground truth. Table 2 is the error matrix for Product 1, 
Table 3 is the error matrix for Product 2 and Table 4 is the error 
matrix for Product 3. 
Product 2 accomplishes overall accuracy of 73.6%; 10% greater 
than Product 1. The improvement can be observed in each single 
class. Product 3 provides slight amount of improvement in 
accuracy compared to Product 1.