Use of error probabilities to improve area estimates based on maximum likelihood classifications

Abstract

Maximum likelihood classification of multispectral remotely sensed data has been demonstrated to be extremely powerful and efficient in a great number of investigations; but it shows marked limitations when applied to highly heterogeneous surfaces, mainly because of distribution anomalies of the cover categories. In these cases, the extent of some cover classes tends to be overestimated with respect to that of other classes, so that the general utility of the entire process can be seriously decreased. In the present paper a method is presented for correcting this kind of misclassification based on the estimation of error probabilities from a preliminary classification and on the insertion of this information in a subsequent processing. The evaluation of error rates is performed by means of confusion matrices referring to the original training pixels, and the information obtained is then inserted in another discrimination process according to the theory of the insertion of prior probabilities in maximum likelihood classifiers. The whole procedure was tested in a semiarid environment with extremely heterogeneous cover types using Landsat-5 Thematic Mapper scenes and relying on representative ground references. The results show a marked improvement in classification performances mainly in terms of symmetry between the extension of the ground references and that of the surfaces obtained by the interpretation technique.