Towards The Development Of Universal Support Vector Regression Models For Quantifying Land Cover Across Cities

Remote sensing has become a valuable Earth observation tool for urban environmental research. Images acquired by various multispectral sensors have been extensively used for mapping urban land cover and urban land use, for detecting urban growth patterns, or for evaluating urban environmental conditions. With the emergence of the hyperspectral satellite missions Environmental Mapping and Analysis Program (EnMAP) and Hyperspectral Infrared Imager (HyspIRI), timely and globally sampled imaging spectrometer data will complement the rich data pool. Not least due to the high spectral information content and the resulting possibility for more accurate and detailed mapping, this new type of hyperspectral images appear promising for enhancing large scale urban remote sensing applications. Along with the advancement in sensor technology, the need for transferable and/or globally applicable methodological workflows for repeated and accurate mapping across space and time becomes evident. In this study, we illustrate an approach towards the development of universal support vector regression (SVR) models for quantifying urban land cover across multiple cities. Model universality is ensured by a comprehensive synthetically mixed training dataset, generated from spectral libraries that comprise relevant construction materials and natural surface types of the respective cities. The approach was tested based on two simulated EnMAP scenes covering the urban-rural gradients of Berlin, Germany, and Brussels, Belgium. Spectral libraries, developed from airborne hyperspectral acquisitions of both cities, were used to compile the synthetic training set for SVR model development. The mapping followed the vegetation-impervious-soil framework (VIS), a well-recognized mapping scheme in the urban context. Results demonstrate the universality of SVR models for VIS fraction mapping in Berlin and Brussels when the underlying synthetic training set includes library information from both cities (average MAE: 7.8% for Berlin, 14.5% for Brussels). In contrast, city-specific SVR models trained with synthetic mixtures from each spectral library separately only reveal similar performances when applied locally (average MAE: 8.0% for Berlin, 13.4% for Brussels), whereas their transfer to the other city leads to reduced accuracies (average MAE: 13.0% for Berlin, 20.0% for Brussels). These findings demonstrate the possibility to develop support vector regression models for detailed land cover mapping based on synthetic training information that can be transferred between cities. As a next step we will include additional cities with different spatial configurations and spectral compositions to establish the approach as a generalizable methodological workflow for urban mapping.status: publishe