Use of multispectral satellite imagery and hyperspectral endmember libraries for urban land cover mapping at the metropolitan scale

The value of characteristic reflectance features for mapping urban materials has been demonstrated in many experiments with airborne imaging spectrometry. Analysis of larger areas requires satellite-based multispectral imagery, which typically lacks the spatial and spectral detail of airborne data. Consequently the need arises to develop mapping methods that exploit the complementary strengths of both data sources. In this paper a workflow for sub-pixel quantification of Vegetation-Impervious-Soil urban land cover is presented, using medium resolution multispectral satellite imagery, hyperspectral endmember libraries and Support Vector Regression. A Landsat 8 Operational Land Imager surface reflectance image covering the greater metropolitan area of Brussels is selected for mapping. Two spectral libraries developed for the cities of Brussels and Berlin based on airborne hyperspectral APEX and HyMap data are used. First the combined endmember library is resampled to match the spectral response of the Landsat sensor. The library is then optimized to avoid spectral redundancy and confusion. Subsequently the spectra of the endmember library are synthetically mixed to produce training data for unmixing. Mapping is carried out using Support Vector Regression models trained with spectra selected through stratified sampling of the mixed library. Validation on building block level (mean size = 46.8 Landsat pixels) yields an overall good fit between reference data and estimation with Mean Absolute Errors of 0.06, 0.06 and 0.08 for vegetation, impervious and soil respectively. Findings of this work may contribute to the use of universal spectral libraries for regional scale land cover fraction mapping using regression approaches.