International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV. Part B2. [Istanbul 2004 
  
pem. 
PU m 
SN 2 
geese 
p 
Ye con ^ 
Wu: 
zi L 
   
  
Z mew 
AKGARAY; 
      
       
   
3 
jes | 
    
 owareanm | oes 
A m Avian sur! 4 d. 
    
Figure 1. Study area, Bartin and its close vicinity, 
shown in Turkish administration map and also by 
Landsat image of the area taken in July 2000. 
For the temporal analysis of forest cover changes, Landsat-5 
TM images covering the test site and taken on 19.05.1992 and 
12.07.2000 have been utilised. At the processing phase, only 
spectral channels 1, 4 and 7 of Landsat sensor were available for 
the experimental purposes. Reference datasets employed during 
the classification procedures of Landsat-5 TM images includes 
the management plans, stand and age class maps, topographic 
maps, aerial photographs and personal knowledge about the 
area. Another map used as a data belonging 1984 which has 
1:25000 scales and showing degree of canopy was obtained 
from local Forest Directorate. Classification procedure has been 
carried out using the related module of PCI Geomatica V8.2 
software package (www.pcigeomatics.com). For GIS analysis, 
ArcMap V8.3 software was used. 
However, before the image classification processes, geometric 
correction of Landsat images was completed. For this purpose, 
21 uniformly distributed GCPs digitized from the 1:25,000 
scales topographic maps of the interest area were used. 
Planimetric accuracy of these GCPs can be expected in the 
range of 7.5m. On Landsat images, linear features appeared 
sharp enough, so GCPs are mainly selected from road crossings 
and bridges. Digital image coordinates for GCPs were measured 
manually using the GCPWorks module of PCI system with the 
sub-pixel point determination. Then, affine transformation was 
applied between the GCPs's image and ground coordinates. 
Root means square errors for X and Y directions were found to 
be 0.69 pixels (20.7m) and 0.67 pixels (20.1m) respectively. 
After producing transformation function, for image registration, 
bilinear resembling method was used to determine the pixel 
values to fill into the output matrix from the original image 
matrix. 
3. Image Classification Procedure 
The overall objective of image classification procedures is to 
automatically categorize all pixels in an image into land cover 
classes or themes. Normally, multispectral data arc used to 
perform the classification and, indeed, the spectral pattern 
present within the data for each pixel is used as the numerical 
472 
basis for categorization. That is, different feature types manifest 
different combinations of DNs based on their inherent spectral 
reflectance and emittance properties. In this light, a spectral 
“pattern” is not at all geometric in character. Rather, the term 
pattern refers to the set of radiance measurements obtained in 
the various wavelength bands for each pixel. Spectral pattern 
recognition refers to the family of classification procedures that 
utilizes this pixel-by-pixel spectral information as the basis for 
automated land cover classification. 
One way of discriminating changes between two dates of 
imaging is to employ post classification comparison. In this 
approach, two dates of imagery are independently classified and 
registered. Then an algorithm can be employed to determine 
those pixels with a change in classification between dates. In 
addition, statistics (and change maps) can be compiled to 
express the specific nature of the changes between the dates of 
imagery. Obviously, the accuracy of such procedures depends 
upon the accuracy of each of the independent classifications 
used in the analysis. The errors present in each of the initial 
classifications are compounded in the change detection 
procedures (Lillesand and Kiefer, 1994). 
In a view of approach given above, image for each year was 
analyzed with supervised classification method since the 
authors have many reference materials and personal knowledge 
about the region. As a classification procedure, maximum 
likelihood method was selected for more reliability percentage. 
Table 1 is an error matrix that prepared to determine how well a 
classification has categorized a representative subset of pixels 
used in the training process of a supervised classification. This 
matrix stems from classifying the sampled training set pixels 
and listing the known cover types used for training (columns) 
versus the pixels actually classified into cach land cover 
category by the classifier (rows). Furthermore, classification 
results for each year are shown in Figure 2 with the respective 
colors. In this phase, main training sites arc assigned as forest, 
sca, agricultural and idle areas, settlement arcas, sandy-stony- 
rocky area and also cloud available in one image as an 
additional class. 
Inter.