Estimation of a Priori 
  
  
Probabilities 
  
of Landcover Categories 
for Bayes' Classifier 
Kunihiko Yoshino, Keiji Kushida 
Assistant Professor 
The Faculty of Agriculture 
University of Tokyo, JAPAN 
XVIII ISPRS Congress Vienna 1996 
Commission III 
KEY WORDS: Decomposition of mixed data,A Priori Probability, Landcover Classification, 
Bayes' Classifier,Landsat TM 
RBSTRRCT 
A priori probabilities of landcover 
categories in the study area improve the 
landcover classification accuracy, although 
the probabilities are very difficult to 
estimate in advance of the analysis. 
Algorithms for the decomposition of mixels to 
pure landcover categories were developed to 
estimate landcover area ratios in mixels. If 
the study area is supposed to be a very large 
mixel which 
several landcover 
categories, some of the 
contains 
decomposition 
algorithms of mixed data can be applied to the 
centroid vector of the study area. The area 
ratios of landcovers in the study area are 
equal to the a priori probabilities of 
landcovers. 
The algorithm of maximum likelihood 
estimation was applied to estimate the a 
priori probabilities of landcovers in the 
study area in this research. 
this research, the 
As a result of 
estimation algorithm 
worked well and the a prior probabilities of 
landcovers in seven small study sites were 
estimated very well. Moreover, those 
estimated a priori probabilities of landcovers 
improved the 
accuracy of landcover 
classification in the study sites. 
1. Introduction & Objectives 
Supervised classification has been used for 
most  landcover classification of remotely 
sensed images. The maximum likelihood 
classification algorithm has been generally 
used since it is simple and reasonable in 
terms of statistics. 
This algorithm has some assumptions in 
particular. One of the assumptions is that the 
Bayes' method needs the a priori probabilities 
of landcovers, but estimating the a priori 
994 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
probabilities of landcovers in the study area 
in advance of the analysis is very difficult. 
Therefore, most of the landcover research 
studies have used even weights of a priori 
probabilities for the maximum likelihood 
As Strahler (1980) reported, 
priori probabilities of 
landcovers in the study area for the maximum 
likelihood classifier 
classification accuracy. 
classification. 
the use of a 
improves the 
The a priori probabilities of landcovers in 
the study area are equal to the area ratios of 
landcovers in the study area following the 
definition of a priori probability. The area 
ratios of landcovers in the study area can be 
stimated from the mean valu es of 
multispectral data of the study area, applying 
the algorithm of the decomposition of mixed 
data. Then, we can improve the landcover 
classification of the maximum 
likelihood method by introducing the area 
ratios of landcovers in the study area. 
This paper has three objectives. The first 
objective is to show that the algorithm of 
decomposition of mixed data can be applied to 
accuracy 
estimate area ratios of landcovers as a priori 
probabilities in the study area. The second 
is to show that a priori probabilities of 
landcovers can be estimated well by this 
algorithm. The last is to show that the 
introduction of a priori probabilities of 
landcovers to the  Bayes' classification 
results in good 
classification. 
accuracy of  landcover 
2. Method to estimate the area 
ratios of landcovers 
The a priori probabilities of landcovers 
in the study area are supposed to be equal to 
the area ratios of landcovers in the study 
     
   
   
  
  
  
  
  
  
   
   
   
   
    
     
     
       
    
        
    
   
   
     
     
   
      
   
    
    
     
    
    
   
   
   
     
     
    
   
   
   
   
   
   
     
   
  
ar 
of 
SU 
tr 
ve 
pr 
9 à O 0 OH