Accuracy Assessment

1. Post-classification smoothing
        majority filter: use a moving window to filter out the "salt and pepper minority pixels
        assign the majority category of the window to the center pixel of the window

2. Select test areas
        Select test areas to evaluate the accuracy of a classification
        selecting test areas that are representative categorically and geographically,
        sampling methods: uniform wall-to-wall, random, stratified random sampling
        sample size: 50 - 100 each category

3. Error matrix
A classification is not complete until its accuracy is assessed

   error matrix (confusion matrix, contingency table)
        - compares the ground truth and the results of the
            classification
        - can be used to evaluate the result of classifying the training set pixels and
            the results of classifying the actual full-scene

        - diagonal cells are correctly classified pixels
             
                                            all correctly classified pixels
          overall accuracy =  ------------------------------
                                                    total pixels

             - if the reference data are arranged as columns and the classified as rows
             - non-diagonal column cells are omission errors
             - non-diagonal row cells are commission errors

                                                        correctly classified in each category
        - producer's accuracy =  ------------------------------------------------------
                                                the total pixels used in the category (column total)
          omission error = 1 - producer's accuracy

      
                                                        correctly classified in each category
        - user's accuracy = ---------------------------------------------------------
                                               total pixels classified in the category (row total)
          commission error = 1 - user's accuracy
 

4. KHAT statistics
A measure of the difference between the actual agreement between reference
data and the results of classification, and the chance agreement between
the reference data and a random classifier

^       observed accuracy - chance agreement
k  = ---------------------------------------
             1 - chance agreement

        the KHAT value usually ranges from 0 to 1.
        0 indicates the classification is not any better than a random assignment of pixels;
        1 indicates that the classification is 100% improvement from random assignment

              r          r
          NS x_ii -  S (x_i+  ×  x_+i)
^          i=1       i=1
k = --------------------------
                      r
           N²  -  S (x_i+  ×  x_+i)
                    i=1
        r - number of rows in the error matrix
        x_ii - number of obs in row i and column i (the diagonal)
        x_i+- total obs of row i
        x_+i - total obs of column i
        N - total of obs in the matrix

- KHAT considers both omission and commission errors

5. Reading: chpt7
 
 
 
 
 
 
 

...Back to Ling Bian top page.