Several indicators in the accuracy verification method of confusion matrix in image classification

ToolBox ->classification->postclassification->confusion matrix->using Ground Truth ROIs , we can get the following confusion matrix of classification precision verification.

To understand this accuracy verification result, we need to understand several evaluation indicators in the confusion matrix:

1, the overall classification accuracy ( Overall Accuracy )

equals the sum of the cells that are correctly classified divided by the total number of cells. The number of cells that are correctly classified follows the diagonal distribution of the confusion matrix, and the total number of cells equals the total number of cells for all true reference sources, such as Overall accuracy = (110230/125843) in this Precision classification accuracy table 87.5933% .

2. Kappa coefficients ( kappacoefficient )

it is by putting the total number of cells of all true references ( N ) multiplied by the confusion matrix diagonal ( Xkk ), minus the sum of the true reference cell number in a class and the total number of cells in the category, divided by the sum of the total number of cells of the cell minus the total number of true reference cells in a class and the total number of cells in the category, the result of summing all categories.

3, error ( Commission )

refers to the class which is divided into the user's interest, but actually belongs to another kind of cell, it shows in the confusion matrix inside. In this example, a total of 19210 cells are divided into woodland , which are correctly classified 16825,2385 The other category is divided into woodland (the sum of the other classes of woodland in the confusion matrix), then the error of 2385/19210= 12.42%is the wrong one.

4, leakage error ( omission )

refers to the true classification of the surface, when it is not divided by the classifier into the corresponding category of the number of cells. As in this example of the woodland class, there are real reference cell 16885 , of which 16825 is correctly classified, the remaining three One is divided into the remainder (the sum of other classes in a column of the arable class in the confusion matrix), the leakage error 60/16885=0.36%

5. Cartographic accuracy ( PROD.ACC )

is that the classifier correctly divides the entire image cell into A the number of cells of the class (diagonal value) and A total number of class true references (in the confusion matrix A the ratio of the sum of the class columns. As in this example, the woodland has 16885 Real reference cells, of which 16825 is correctly classified, so the cartographic accuracy of the woodland is 16825/16885=99.64% .

6. User accuracy ( USER.ACC )

is to correctly divide A the total number of cells (diagonal values) of the class and the classifier divides the entire image's cells into A the total number of cells for the class (in the confusion matrix A the sum of the class rows) ratio. As in this example, the woodland has a correct classification of 16825 , a total of the forest is divided into 19210, so the user accuracy of woodland is 16825/19210=87.58% .

Several indicators in the accuracy verification method of confusion matrix in image classification