"OR 
| four broad 
ng Landsat 
assification 
models did 
elationships 
land cover 
v additional 
only. 
procedure, 
ons derived 
containing 
ined. Most 
/ledge base 
rked in the 
nd Guiking, 
e compared 
a Landsat 
th a ground 
erived from 
tion of 10 
ework of a 
d use" an 
vations and 
dscape and 
lassification 
Is study are 
supervised 
  
classification only based on remote sensing, (2) 
topographic ^ normalization prior to supervised 
classification and (3) application of rules after supervised 
classification. 
Approach | Arch 2 { Arch 3 : 
Supervised Classification [Topographic ! Cmtextuel Information !Grdexued Information 
1 3a Export knowledge 
13b Statistical analysis 
  
  
  
  
  
  
' 
i 
' 
| 
' 
i 
i 
| ' 
| i 
' ' 
' 
{| Emo muraux i 
f AFL ' 
| 
i 
anus of the nse ) 
  
  
  
  
Diassan and 
GC) pula Cundusuams 
| Aw din tion 
Figure | Overview of applied methods 
Approach 1 Supervised classification (standard) 
In this approach, a standard procedure of image 
processing, a type of supervised classification was used 
(Richards, 1993). Supervised classification is the process 
of using samples of known identity to classify pixels of 
unknown identity. For this process training fields are 
selected in different land cover classes. We used the 
generally applied maximum likelihood classification, 
since it uses algorithms in a statistically acceptable 
manner and is based on probability statistics. 
For evaluation of the classification accuracy an error 
matrix was calculated based on the independent geo- 
dataset. The classification was summarized in a 
classification accuracy report presenting overall accuracy, 
producers accuracy and users accuracy of the different 
land cover classes. These accuracy measures are defined 
as follows: 
e Overall accuracy = Z correct pixels/ Z Pixels in error 
matrix 
e  Producer's accuracy - X correct pixels per category/ X 
reference pixels 
e. User's accuracy — X correct pixels per category/ X 
classified pixels 
This error matrix is used to compare this relatively 
straightforward classification with error matrices 
produced by the other approaches using ancillary 
information. All image processing has been performed 
with Erdas Imagine. 
Approach 2 Topographic normalization before 
supervised classification 
In this approach the data were topographically normalized 
before classification. This normalization is meant to 
overcome the problem that slopes facing the sun are 
lighter than slopes without sun. 
A number of investigations which attempt to explain the 
topographic effect, especially for Landsat and Spot digital 
multispectral data, have been conducted. Most of these 
studies attempted to account for and correct the negative 
aspects associated with the topographic effect by 
modifying the surface radiance values recorded by the 
satellite sensors using the cosine of the angle of incidence 
and inclination, as calculated from a registered digital 
elevation data set. The topographic effect in digital 
imagery can be compensated for by transformation based 
on Lambertian or Non-Lambertian reflectance models. In 
this study both models are used. 
The Lambertian Reflectance model assumes that the 
surface reflects incident solar energy uniformly in all 
directions, and that variations in reflectance are due to the 
amount of incident radiation, which lead to simple 
formulas. 
The Non-Lambertian Reflectance model (Minaert 1961) 
supposes that the observed surface does not reflect 
incident solar energy uniformly in all directions. The data 
were transformed using equations developed by Colby 
(1991). The Minnaert constant (k), needed in the 
equations was found by regressing a set of observed 
brightness values from the remotely sensed imagery with 
known slope and aspect values, with the same type of 
land cover. On both normalized images again supervised 
classifications are performed and the accuracy is 
evaluated using error matrices and accuracy reports. 
Approach 3 Contextual information to improve 
supervised classification. 
Ancillary data can be incorporated in image classification 
during : 
1. Pre - classification scene stratification. 
2. Post - classification class sorting and 
3. Classification modification through prior probabilities 
(Hutchinson, 1982). 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 343