TESTING FOR MARGINAL HOMOGENEITY OF REMOTE-SENSING CLASSIFICATION ERROR MATRICES WITH ORDERED CATEGORIES

The classification error matrix is frequently used for assessment of t he quality of land-use maps and statistical areal estimates derived fr om remotely sensed data. The difference between the row and column mar ginals within individual categories is a useful measure of systematic differences between the two classifications. The tabulated values of a n error matrix are the result of a sampling procedure, and it is there fore important to know the statistical significance of the differences . Statistical tests are presented for testing the significance of such differences of error matrices with ordered categories. One of the tes ts can be used to compare differences of several categories, as well a s a test for significance of differences for individual categories. A case study of photo interpretation of tree species according to an ord inal scale demonstrates the computation of the test statistics. Differ ent ways the tests might be used are considered, and some requirements related to the sampling model of the statistics are discussed.