Unsupervised Classification

1. Principles
        - the process requires a minimal amount of initial input
            from the analyst
        - a numeric operation searches for natural grouping of
            the spectral properties of pixels
        - the analyst determines the information class for each
            spectral class

2. Chain method
        pass 1 builds clusters and calculates their mean vectors; pass 2
assigns pixels to clusters based on the minimum-distance rule

   pass 1: cluster building
        - R, a spectral radius used to determine whether a new
            cluster should be formed (e.g., 15 brightness)
        - N, the number of pixels to be evaluated between each major
            merging of the clusters (e.g., 2000)
        - C, a spectral distance used to determine merging clusters
            when N is reached (e.g., 30 brightness)
        - Cmax, the maximum number of spectral clusters (categories)
            (e.g., 20) to be identified

        - the operation evaluates pixels sequentially, combining
            successive pixels into a cluster if their spectral
            distance < R
        - a cluster is complete when N is reached
        - if the spectral distance between two clusters is < C, the
            two clusters are merged, until no clusters with distance < C
        - the new mean is the weighted average of the two original
            clusters

   pass 2: assigning pixels
        - assigns pixels based on the minimum distance classifier

        - manual modification based on knowledge of the area, co-
            spectral plots, and interactive display
          (F8-23,8-24,8-25, Jensen, 1996)

3. ISODATA method
        Iterative Self Organizing Data Analysis Technique

        Parameters required to run ISODATA
        - Cmax, the maximum number of spectral clusters
        - T, maximum % of pixels whose classes are allowed to be
            unchanged between iterations
        - M, the max number of times of classifying pixels and calculating
            cluster mean vectors
        - Minimum members in a cluster (%). For example, if the % <0.01, the
            cluster is deleted
        - Maximum Std Dev, when a std dev > specified Max-std-dev and the
            number of members > 2*Min members, the cluster is split
        - split separation: when the value  is not 0.0, it is used to
            determine the locations of the new mean vectors plus and
            minus the split separation value
        - minimum distance between cluster means. Clusters with a weighted
            distance < this value (e.g., 3.0) are merged

        - it uses a large number of passes
        - the initial means are determined based on the mean and std dev
            of each band (F8-26, Jensen, 1996)

   Iterations
        - assigns each pixel using the minimum distance classifier
           Second to Mth iteration
        - re-calculate the mean vectors,
        - examine Min members(%), Max std dev, split separation, and
            Min distance between clusters
        - the iteration stops when T or M is reached

4. Reading: chpt7
 
 
 
 
 
 
 
 

...Back to Ling Bian top page.