Tree Density Detection using Spectral Unmixing without known Target Spectra

Detection of subpixel endmember proportions in the realm of land cover and land cover change mapping is imperative when considering new and innovative methods to improve local scale land over estimates. However, often these results are not accurate enough due to rare endmember spectra information. Several studies including spectral unmixing without known target spectra exist, however, the models used herein are based on the assumption that all endmembers extracted from imagery and employed are spectrally `pure'. But in medium resolution satellite data especially in semi arid regions for instance pure photosynthesis active vegetation pixels are rare, so these approaches seems to be unsuitable. In this paper, we present a new spectral unmixing model that does not need exact a priori knowledge about endmember spectra. A best fitting function between the endmember types, shadow' and, green vegetation' and the land cover type, tree' is calculated. As an experimental result, a comparison between tree density derived from this approach based on a 15 m resolution ASTER dataset and a 1 m resolution IKONOS classification is shown to illustrate the accuracy of the model.