Removing ambiguities in a multi-spectral image classification

Abstract

This article deals with the processing of previously classified satellite sensor images according to land use in order to remove ambiguities in labels. The data are SPOT HRV XS images of the south-east of France. The date is the 18 August 1993. These images have already been classified with a maximum likelihood method but some labels are not correctly determined. For the elimination of ambiguities, we applied our method of determination of land use mixture within pixels. The processing used for the determination of mixture of spectra in pixels of multi-spectral satellite images has already been explained in other articles. We first briefly review our method of determination of land use mixture. We applied a linear analysis of spectra to the mask of the problematic class. Then we explain how we deal with ambiguities in labels of the maximum likelihood classification. Each problematic class is treated independently and we re-classify each pixel of the class according to their land use mixture extraction.

We finish with three examples of satellite images that have not been correctly classified. The first one is the vineyard case. The second is for bare soil and urban zones. These two kinds of land use are naturally not easy to separate because their spectra are very close. The last one is a forestry survey application, the determination of the density of planted pines. In all these examples the use of the determination of land use mixture has helped to correct the wrong classification.