and so also some of the finer classes like early and late 
pulses, deep and shallow water bodies etc. 
Table 1. Input parameters for multi-resolution segmentation 
of satellite data used in the study 
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
  
  
  
  
  
  
Study Homogeneity Segmen Scale No. of | Object 
Area & | Criteria tation image Mean 
data (0-1 range) Level objects Size in 
formed pixels 
Site-1, Color-0.9, Level-1 1 201932 1.9 
IRS- Shape-0.1 Level-2 5 10348 25.3 
LISS- (Smooth- Level-3 10 2902 90.3 
IH ness:0.9, Level-4 17 997 262.9 
Compact- Level-5 25 483 542.7 
ness:0.1) Level-6 28 386 679.1 
Level-7 35 242 1083.0 
Site-2 Color-0.8, Level-1 1 218476 1.2 
IRS- Shape-0.2 Level-2 2 73698 3.6 
LISS- (Smooth- Level-3 4 17204 15.4 
II ness:0.9, Level-4 6 8018 32.7 
Compact- Level-5 8 4778 54.9 
ness:0.1) Level-6 20 908 288.7 
Site-1, Color-0.7, Level-1 1 82888 3.6 
SAR Shape-0.3 Level-2 2 16072 16.3 
data (Smooth- Level-3 3 6738 38.9 
ness:0.9, Level-4 4 3617 72.5 
Compact- Level-5 4.5 2835 92:5 
ness:0.1) Level-6 5 2232 117.5 
Level-7 7 1087 241.2 
Level-8 8 842 311.3 
Level-9 11 438 598.5 
  
  
  
  
  
  
  
In the classification step, initially, broad land use land covers 
such as water bodies, vegetated areas, fallow fields and 
settlements have been classified at level-5 by Nearest 
Neighbour rule. However, for certain classes with poor 
separability from other classes manual classification 
approach has been followed. Settlement is one such class that 
has been manually classified based on the apriori 
information to avoid misclassification. Similarly, to avoid 
misclassification at finer levels of segmentation, certain 
classes like vegetation, fallow fields and water bodies have 
been declared as super classes at segmentation level-5 and 
class hierarchy has been established. Further, taking 
advantage of the distance criterion for contextual 
information, banana plantations could be classified 
satisfactorily in spite of its poor separability from cotton 
class. The classification accuracy computed in terms of 
kappa coefficient shows that the approach consistently 
performed better over per pixel classifier. Performance of the 
object based segmentation and classification has been tested 
on site-2 with fewer classes, allowing greater intra class 
variability. Comparison of results shown in figure 2 confirm 
the better performance of the adopted approach over the 
standard maximum likelihood classifier. In the present 
exercise, the utility of object oriented segmentation and 
classification could not be established with the RADARSAT 
SAR dataset used. However, it could be observed that banana 
plantations, bare fields and water bodies could be 
discriminated from cropped fields and settlements. 
It was also observed based on visual analysis and ground 
observations that classification accuracy computed by 
defining the homogeneous test areas corresponding to each 
of the classes, in general, leads to unrealistically higher 
classification accuracy. It is obvious that signature overlap 
occurs mainly at the class boundaries leading to 
misclassification, which is often difficult to account for at 
post classification evaluation stage. Misclassification of 
pixels by per pixel classifier was prominently seen at edges 
    
  
    
  
  
   
   
   
  
  
   
     
   
  
  
   
   
   
  
  
  
   
   
  
  
  
  
  
    
   
   
  
   
  
  
  
  
  
  
  
  
  
  
  
   
   
  
   
    
     
and class boundaries, which resulted in not so distinct class 
boundaries. Object based classification yielded clear class 
boundaries with reduced misclassification. 
  
  
Red soil 
Cross Plantation 
Other Ves. Tanks 
Figure 2. Results of per pixel (ML) classifier and object oriented 
multi-resolution segmentation based classification