Number Pixels      (within class)    (within&between)
  Class Pair    Class-1 Class-2    Class-1 Class-2    Class-1 Class-2



  Number of data values in each class pair and number of values
   meeting class thresholds.
                                    Met Threshold1    Met Threshold1&2


                 Number Pixels      (within class)    (within&between)
  Class Pair    Class-1 Class-2    Class-1 Class-2    Class-1 Class-2    Number Features



  Number of data values in each class pair, number of values meeting
   class thresholds and number of features required for each class
   pair to meet optimization percent.
                                    Met Threshold1    Met Threshold1&2



 Feature extraction was stopped before the evaluation of all
  class pairs was completed.



  Eigenvectors could not be found while determining the feature extraction
 transformation matrix.



Projection Pursuit - Numerical Optimization

Number output features:

Bhattacharyya % change:

Projection Pursuit - Feature Selection

Projection Pursuit - First Stage

BU Number feature groups:

TD Number feature groups:

   Numerical optimization threshold is: %f%%


  Numerical Optimization will be used.


  Maximum number of output features is: %ld


   Both choices up thru %ld output features, then use random choice.


   Use random choice.


  Split feature groups with an odd number of bands:


   Begin with last feature grouping.


  Initial number of features: %ld


   Bottom-up threshold is: %f%%


   Top-down threshold is: %f%%


  Top-down/Bottom-up method used for First Stage


  Top-down method used for First Stage


  Uniform Channel Grouping used for First Stage


  Projection Pursuit Feature Selection Algorithm


  Projection Pursuit Algorithm


 Feature Extraction Preprocessing will be done with:


 The Last Feature Extraction Preprocessing results will be used.


 No Feature Extraction Preprocessing will be done.



 Feature Extraction will be done with:



 No Feature Extraction will be done; preprocessing only.



 A subset of the class pairs will be used.



 All class pairs will be used.



 Final Feature Extraction Transformation Matrix


  Optimized Preprocessing Channel Transformation Matrix


  After Numerical Optimization
  Minimum Bhattacharyya =  %9.4f



  Preprocessing Channel Transformation Matrix

  Channel Decision Tree Table
   Number of    Minimum     Min Class
   Features  Bhattacharyya     Pair      Channel Grouping



  The number of final features stopped at %ld because it can not grow larger than
   one fewer than the minimum number of class samples of %ld.


  The number of initial features was changed to %ld so that the number of bands in
   any one group would be at least one fewer than the minimum number of class
   samples of %ld.



Determining effective db feature matrix

Finding nearest pixel in other class

Classifying pixels

Computing final eigenvectors

Optimizing class

Class Pair

Calculating covariance inverses

Normalizing eigenvectors

Computing Sw(-1/2)Ev

Computing eigenvectors of Sw(-1/2)SbSw(-1/2)

Computing Sw(-1/2)SbSw(-1/2)

Computing eigenvector of Sw

Computing feature extraction matrix

Computing between-class scatter matrix (Sb)

Computing within-class scatter matrix (Sw)

Feature Extraction - decision boundary

Feature Extraction - discriminant analysis

Loading class statistics

  Class %3ld can not be used because all pixels are from cluster(s).


  Approximate maximum number of pixels per class = %ld.


  Class optimization threshold = %g percent.


  Between class threshold = %g.


  Within class threshold = %g.


  Minimum threshold number = %ld.



  Decision Boundary Technique



  Discriminant Analysis Technique


Feature Extraction

   - %ld pixels were used that were closest to the class boundary.



No transformation matrix was generated because the number of points
in at least one class was 0.  This is probably due to cluster fields
being used for the class statistics.