Supervised Classification
    A procedure that requires interaction with the analyst

1. General procedures
        - training stage
            The analyst identifies the representative training areas and
            develops summary statistics for each category

        - classification stage
            Each pixel is categorized into a land cover class

        - output stage
              The classified image is presented in GIS or other forms

2. Minimum-distance classifier
        - calculate mean of the spectral values for the training set
            in each band and for each category

        - measure the distance from a pixel of unknown identify to the mean of
            each category

        - assign the pixel to the category with the shortest distance

        - assign a pixel as "unknown" if the pixel is beyond the
            distances defined by the analyst

        Advantage:
        - computationally simple and fast

        Disadvantage:
        - insensitive to differences in variance among categories

3. Parallelepiped classifier
        - form a decision region by the maximum and minimum values of
            the training set in each band and for each category

        - assign a pixel to the category the pixel falls in

        - assign a pixel as "unknown" if it falls outside of all regions
 

        Advantage:
        - computationally simple and fast

        - take differences in variance into account

        Disadvantage
        - perform poorly when the regions overlap because of high
            correlation between categories (high covariance)
 

4. Gaussian maximum likelihood classifier
        - assume the distribution of the training set is normal

        - describe the membership of a pixel in a category by
            probability terms

        - the probability is computed based on probability density
            function for each category

        - a pixel may occur in several categories but with different
            probabilities

        - assign a pixel to the category with the highest probability

        Advantage:
        - take into account the distance, variance, and covariance

        Disadvantage:
        - computationally intensive
 

5. Training
    Collect a set of statistics that describe the spectral response pattern
        for each land cover type to be classified

        - select several spectral classes representative of each
            land cover category

        - avoid pixels between land cover types

        - a minimum of n+1 pixels must be selected (n=number of bands)

        - more pixels will improve statistical representation, 10n or
            100n are common

        - spatially dispersed training areas throughout the scene better represent
            the variation of the cover types
 

6. Training set refinement
   Graphic representation
        - It is necessary to display histogram to check for normality and purity

        - Coincident spectral plot with 2 std dev from the mean is useful
            to check for category overlap

        - 2-D scatter gram is also useful for refinement

   Quantitative expression
        - divergence matrix, higher values indicate greater separability

   Self-classification
        - classify the training set as a preliminary classification

   Interactive preliminary classification

   Representative sub-scene classification
 

7. Reading chpt7
 
 
 
 
 
 
 
 
 
 
 

...Back to Ling Bian top page.