nbul 2004 
er image) 
> thematic 
  
  
  
nages) 
using the 
oniferous 
sed into 
rejoining 
ept for the 
age) were 
erarchical 
classification procedure. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
The final image and classes are 
illustrated in Figure 8. 
Water 
Coniferous Forest 
A {Deciduous Forest 
^t ural areas 
i Forest Flt 
[Roads 
Figure 8. Final image created by using Hierarchical approach 
3.4.5 Spatial Reclassification: Spatial reclassification 
represents a comparatively simple way to examine the spatial 
variation in land-cover in remotely sensed images, and is easy 
to implement in most image processing systems. It can be 
performed in one of two ways. The first, named as kernel-based 
spatial reclassification (Barnsley and Barr, 1992), involves 
passing a simple convolution kernel across the land-cover 
image. In the second, referred to as object-based spatial 
reclassification, discrete objects (i.e., groups of adjacent pixels 
with the same class label) are identified within the initial image 
segmentation: information on the size, shape and spatial 
arrangement of these objects is subsequently used to determine 
the nature of the land-use in different parts of the image. 
In this study kernel-based procedure was applied to resultant 
land-cover image created using the hierarchical approach. 
Following contextual rules were used during spatial 
reclassification : 
* Pixels labeled as urban due to the spectral similarities can be 
reclassified and labeled as coast if they share a border with 
the lake. 
e Agriculture and forest pixels can be reclassified as urban if 
they are surrounded by a user-defined number of urban 
pixels. 
e Urban/developed pixels can be reclassified as forest if they 
are surrounded by a user-defined rate of forest area. 
A 3 X 3 kernel was used to detect the urban pixels neighboring 
Water as the first step of the spatial reclassification and 940 
coast pixels mislabeled as urban were detected. The thematic 
image was corrected and a new class was added to the 
classification scheme after these pixels were relabeled as 
“coast”. 
The second step of the spatial reclassification was to detect 
pixels labeled as any kind of forest or agricultural area in dense 
urban regions. À 5 X 5 kernel was used to find out forest or 
agricultural areas surrounded by urban pixels (Figure 9) because 
a 150 m. X 150m. area was found to be suitable for such a 
region which contains various land cover types with small 
parcel sizes. 
515 
F=Forest 
A= Agricultural areas 
U= Urban 
  
Figure 9. Spatial reclassification kernel 
The central forest (or agricultural) pixel was relabeled as urban 
if more than 14 of 24 neighboring pixels (almost 60 percent of 
the area) were labeled as urban. This threshold was found after 
comparing the thematic image and the orthophoto. 
Final 5 x 5 kernel was used to detect urban pixels within 
forested area but a negligible amount of pixels satisfied the 
criteria to relabel as forest. 
4. ACCURACY ASSESSMENT 
A number of randomly selected 265 reference points measured 
in the field survey were used in the accuracy assessment of the 
classification. The class values of the reference points were 
assigned during the field survey, except for the water class. The 
overall accuracy of the proposed hierarchical and maximum 
likelihood classifications were found to be 91.32% and 47.55% 
respectively. In order to compare different classification 
methods namely Hierarchical and Maximum Likelihood 
Classification techniques, Kappa coefficient of agreement as an 
accuracy measure for remote sensing classification is used. 
As it is given in Table 2, Kappa coefficients are obtained as 
0.94 for Hierarchical Classification and 0.37 for Maximum 
Likelihood Classification. This implies that the accuracy of the 
Hierarchical Classification, 91.32 percent, and the accuracy of 
the supervised classification, 47.55 percent, are better than the 
accuracy that would result from a random assignment. This 
result indicates the Hierarchical Classification is better than 
supervised classification in identifying the forestry areas from 
Landsat image. 
  
  
  
  
  
  
  
  
  
  
  
Hierarchical Supervised 
Classification Classification 
Classes 
Accuracy > Accuracy 
(%) Kappa (%) Kappa 
100.00 1.0000 70.59 0.6919 
Water 
Coniferous 95.60 0.8430 30.77 0.8985 
Deciduous 97.06 0.9044 94.12 0.9325 
Agricultural 90.24 0.9132 82.93 0.2902 
Grass 77.36 0.8619 16.98 0.0959 
Urban 90.91 1.0000 90.91 0.4103 
Roads 94.74 1.0000 5.26 0.2818 
Coast 100.00 1.0000 0.00 0.0000 
OVERALL 91.32 | 0.9403 47.55 0.3700 
RESULT 
  
  
  
  
  
  
  
Table 2. Overall classification results