The Determination of Optimal Threshold Levels for Change Detection Using Various Accuracy Indices

A thresholding technique is applied to identify the change and no-change categories from six transformed images produced by image differencing, image ratioing, and principal components analysis. The problems in using different accuracy indices, including overall accuracy, average and combined accuracy (based on both the user's and producer's accuracy approaches), and the Kappa coefficient of agreement in determining an optimal threshold level, are examined. Most indices are biased and affected by the ratio between the number of either the reference or classified samples of the change and the no-change categories. The Kappa coefficient is recommended because it takes into account all cells of the error matrices.