Non-linear spectral mixture analysis of Landsat imagery by means of neural networks

Urban surfaces are highly inhomogeneous because of the high spatial and spectral diversity of man-made structures. Spectral unmixing techniques although developed to be used with hyperspectral data, are useful for assessing sub-pixel information on multispectral data as well. The large spectral variability imposes the use of multiple endmember spectral mixture analysis techniques, in which many possible mixture models are considered to produce the best fit. The use of many endmembers and mixture models result in prohibitive computational time. In this study, an artificial neural network is used to inverse the pixel spectral mixture in Landsat imagery. Endmember spectra, collected from the image were used to train the network and capture the spectral variability of man-made structures.