Nonlinear Spectral Unmixing of Landsat Imagery for Urban Surface Cover Mapping

The high spatial diversity of man-made structures, the spectral variability of urban materials, and the three-dimensional structure of the cities make the mapping of urban surfaces using Earth Observation data, one of the most challenging tasks in remote sensing field. Spectral unmixing techniques can be proven useful with medium spectral resolution data to assess urban surface cover information on a subpixel level. Due to the large spectral variability of urban materials and the multiple scattering of light between surfaces in urban areas, multiple endmembers should be used, and the nonlinearity of spectral mixture should be accounted for. In this study, these issues are addressed using an artificial neural network trained with endmember and nonlinearly mixed synthetic spectra to inverse the pixel spectral mixture in Landsat imagery. A spectral library is built, consisting of endmember spectra collected from the image and synthetic spectra, produced using a nonlinear model specifically developed for urban areas. The method was tested over a case study, and the validation against higher resolution products revealed an accuracy of around 90% for all abundance maps. The comparison performed between the linear and nonlinear implementation of the method proved the need for including the nonlinear term, especially for improving the built-up abundance map. The proposed method is easily transferable to any city and fast in terms of computations, which makes it ideal for the implementation of operational urban services.