Remote Sensing Classification Accuracy

1
Remote SensingClassification Accuracy
2
1. Select Test Areas

• Selecte test areas in an image to evaluate the
  accuracy of a classification
• Test areas should be representative categorically
  and geographically
• Sampling methods uniform wall-to-wall, random,
  stratified random sampling        
• Sample size 50 - 100 pixels each category

3
http//aria.arizona.edu/slg/Vandriel.ppt
4
(No Transcript)
5
2. Error Assessment

• A classification is not complete until its
  accuracy is assessed
• Error matrix
• KHAT statistics

6
Error Matrix

• Also called confusion matrix and contingency
  table        
• Compares the ground truth and the results of the
  classification for the test areas
• Can be used to evaluate the result of
  classifying the training set pixels and the
  results of classifying the actual full-scene

7
Error Matrix

•                 

Diagonal cells are correctly classified pixels
                            correctly
classified pixels 1672 Overall accuracy
  ------------------------------- -------
84                               total pixels
evaluated 1992
8
Error Matrix

•                 

In this case, the non-diagonal column cells are
omission errors e.g. omission error for forest
43/356 12 The non-diagonal row cells are
commission errors e.g. commission error for corn
117/459 25
9
Error Matrix

•                 

                                  correctly
classified in each category producer's accuracy
  ---------------------------------------------- 
                          the total pixels used
in the category (col total) Omission error 1
(100) - producer's accuracy
10
Error Matrix

•                 

                                  correctly
classified in each category user's accuracy  
--------------------------------------------------
-----                         the total pixels
used in the category (row total) Commission
error 1 (100) - user's accuracy
11
(No Transcript)
12
(No Transcript)
13
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

14
KHAT Statistics

•       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

15
KHAT Statistics

•             r          r        N S xii - 
  S (xi    xi)          i1       i1 k
  -----------------------------------
                      r            N2  -  S (xi 
    xi)                   i1
• r - number of rows in the error matrix
• xii - number of obs in row i and column i (the
  diagonal cells)
• xi - total obs of row i
• xi - total obs of column i
• N - total of obs in the matrix

16
KHAT
17
KHAT Statistics

• KHAT considers both omission and commission
  errors

18
(No Transcript)
19
Readings

• Chapter 7