19
Remote Sensing Remote Sensing Classification Accuracy Classification Accuracy

Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Embed Size (px)

Citation preview

Page 1: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Remote SensingRemote Sensing

Classification AccuracyClassification Accuracy

Page 2: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

1. Select Test Areas1. Select Test Areas

► Selecte test areas in an image to evaluate Selecte test areas in an image to evaluate the accuracy of a classificationthe accuracy of a classification

► Test areas should be representative Test areas should be representative categorically and geographicallycategorically and geographically

► Sampling methods: uniform wall-to-wall, Sampling methods: uniform wall-to-wall, random, stratified random sampling random, stratified random sampling                

► Sample size: 50 - 100 pixels each category Sample size: 50 - 100 pixels each category

Page 3: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

http://aria.arizona.edu/slg/Vandriel.ppt

Page 4: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas
Page 5: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

2. Error Assessment2. Error Assessment

► A classification is not complete until its A classification is not complete until its accuracy is assessedaccuracy is assessed

► Error matrixError matrix► KHAT statisticsKHAT statistics

Page 6: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Error MatrixError Matrix

► Also called confusion matrix and Also called confusion matrix and contingency table contingency table               

► Compares the ground truth and the results Compares the ground truth and the results of the of the classification for the test areasclassification for the test areas

► Can be used to evaluate the result of Can be used to evaluate the result of classifying the training set pixels and the classifying the training set pixels and the results of classifying the actual full-sceneresults of classifying the actual full-scene

Page 7: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Error MatrixError Matrix                                   Classified                       Reference Data

Data     Water  Sand  Forest   Urban    Corn Hay Row Total Water 480 0 5 0 0 0 485        Sand 0 52 0 20 0 0 72 Forest        0 0 313  40 0 0 353   Urban 0 16 0 126 0 0 142 Corn 0 0 0 38 342 79 459 Hay 0 0 38 24 60 359 481Col Total  480       68     356 248 402 438 1992

Diagonal cells are correctly classified pixels Diagonal cells are correctly classified pixels

                                                        correctly classified pixels 1672 correctly classified pixels 1672 Overall accuracy =  ------------------------------- = ------- = 84%Overall accuracy =  ------------------------------- = ------- = 84%                               total pixels evaluated 1992                               total pixels evaluated 1992

Page 8: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Error MatrixError Matrix                                  

In this case, the non-diagonal column cells are omission In this case, the non-diagonal column cells are omission errors errors e.g. omission error for forest = 43/356 = 12% e.g. omission error for forest = 43/356 = 12%

The non-diagonal row cells are commission errorsThe non-diagonal row cells are commission errorse.g. commission error for corn 117/459 = 25%e.g. commission error for corn 117/459 = 25%

Classified                       Reference Data Data     Water  Sand  Forest   Urban    Corn Hay Row Total Water 480 0 5 0 0 0 485        Sand 0 52 0 20 0 0 72 Forest        0 0 313  40 0 0 353   Urban 0 16 0 126 0 0 142 Corn 0 0 0 38 342 79 459 Hay 0 0 38 24 60 359 481Col Total  480       68     356 248 402 438 1992

Page 9: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Error MatrixError Matrix                                  

                                  correctly classified in each category correctly classified in each category

producer's accuracy =  producer's accuracy =  ----------------------------------------------                           the ----------------------------------------------                           the total pixels used in the category (col total) total pixels used in the category (col total)

Omission error = 1 (100%) - producer's accuracy Omission error = 1 (100%) - producer's accuracy

Classified                       Reference Data Data     Water  Sand  Forest   Urban    Corn Hay Row Total Water 480 0 5 0 0 0 485        Sand 0 52 0 20 0 0 72 Forest        0 0 313  40 0 0 353   Urban 0 16 0 126 0 0 142 Corn 0 0 0 38 342 79 459 Hay 0 0 38 24 60 359 481Col Total  480       68     356 248 402 438 1992

Page 10: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Error MatrixError 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

Classified                       Reference Data Data     Water  Sand  Forest   Urban    Corn Hay Row Total Water 480 0 5 0 0 0 485        Sand 0 52 0 20 0 0 72 Forest        0 0 313  40 0 0 353   Urban 0 16 0 126 0 0 142 Corn 0 0 0 38 342 79 459 Hay 0 0 38 24 60 359 481Col Total  480       68     356 248 402 438 1992

Page 11: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas
Page 12: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas
Page 13: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

KHAT StatisticsKHAT Statistics ► A measure of the difference between the A measure of the difference between the

actual agreement between reference data actual agreement between reference data and the results of classification, and the and the results of classification, and the chance agreement between the reference chance agreement between the reference data and a random classifierdata and a random classifier

Page 14: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

KHAT StatisticsKHAT 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

Page 15: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

KHAT StatisticsKHAT Statistics                        r          rr          r

       N ×        N × x xiiii -  -  (x (xi+i+  ×  x  ×  x+i+i) ) ^         i=1       i=1^         i=1       i=1 k = ----------------------------------- k = -----------------------------------                                         rr            N           N22  -    -  (x (xii++  ×  x  ×  x+i+i) )                                     i=1 i=1

r - number of rows in the error matrixr - number of rows in the error matrix

xxiiii - number of obs in row i and column i (the diagonal cells) - number of obs in row i and column i (the diagonal cells)

xxi+i+ - total obs of row i - total obs of row i

xx+i+i - total obs of column i - total obs of column i

N - total of obs in the matrix N - total of obs in the matrix

Page 16: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

KHATKHAT

Page 17: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

KHAT StatisticsKHAT Statistics ► KHAT considers both omission and KHAT considers both omission and

commission errors commission errors

Page 18: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas
Page 19: Remote Sensing Classification Accuracy. 1. Select Test Areas ► Selecte test areas in an image to evaluate the accuracy of a classification ► Test areas

Readings Readings

► Chapter 7Chapter 7