Pixel-oriented selection of endmember combinations for linear stochastic spectral unmixing applied to urban environments

Urban areas are characterized by a high frequency of small sized changes in surface cover types whose spatial patterns strongly influence the environmental conditions in cities. Airborne hyperspectral data yield a new potential for their spectrally-based identification, but also raise new challenges in image analysis caused by high spatial and spectral variability of the data. In this context we present a new linear unmixing approach including a pixel-oriented selection of endmember combinations. This approach was especially developed for analyzing urban conditions using data of airborne DAIS 7915 scanner for the city of Dresden, Germany. Because of the big number of spectrally similar endmembers in the urban environment, application of standard unmixing techniques lead to strong local variations of different endmembers and confusion of surface cover types. In comparison to these standard techniques a new extended mathematical model is used which includes stochastic models for each endmember. Additionally, a procedure for pixel oriented selection of likely endmember candidates is developed based on the assumption that the number of endmembers is limited within a pixel. For this purpose, all possible combinations of different endmembers are defined and stored in a list which forms the basis for spatially and thematically constrained endmember selection. These combinations of endmembers are tested during the linear unmixing process. In the result, sensible endmember combinations could be identified during the unmixing process for the 10 km by 4.5 km study area in Dresden. In comparison with standard image classification techniques our approach shows advantages especially in areas dominated by mixed pixels. Thus, a spatially and thematically precise identification of urban surface cover types could be achieved.