Determination of proportions of land use blend in pixels of a multispectral satellite image

This article deals with the determination of proportions of land cover blend which define pixels in a multispectral satellite image. The processing is done in two steps. The authors first apply Q-mode factor analysis and a Varimax rotation to extract the principal factors of the pixels´ spectra. The spectra are then projected onto the plane of the two first factors. Blend proportions are calculated by comparing the position of the observed spectra with that of the reference spectra (i.e. pure constituent spectra) in the spectral space. The output of the above processing is a set of images giving the percentage of various types of land (forest, bare soil, ...) in each pixel. At the end of the article, the authors have compared the results with the ground truth. This comparison shows that the calculated proportions are close to the real land use proportions. They also obtain a classification image and an relative entropy value for each pixel.