PERFORMANCE INDICATORS FOR THE STATISTICAL EVALUATION OF DIGITAL IMAGE CLASSIFICATIONS

A methodology is proposed for a post-processing assessment of the perf ormance of training set selections and classification procedures. This paper focuses on the necessity and usefulness of a critical validatio n of classification results. A statistical procedure is proposed to ev aluate the algorithms for the numerical classification of images. The approach is based on the derivation of performance indicators from mea surements of signature separability and thresholding analysis. Althoug h these measurements are not new in image processing techniques, they are used in this study in an original way for the comparison of output s resulting from different classification criteria. The theoretical de scription of the suggested method is followed by their practical appli cation to a case-study for mapping crop coefficients in an irrigation district.