Generalizing machine learning regression models using multi-site spectral libraries for mapping vegetation-impervious-soil fractions across multiple cities

Forthcoming spaceborne imaging spectrometers will provide novel opportunities for mapping urban composition globally. To move from case studies for single cities towards comparative and more operational analyses, generalized models that may be transferred throughout space are desired. In this study, we investigated how single regression models can be spatially generalized for vegetation-impervious-soil (VIS) mapping across multiple cities. The combination of support vector regression (SVR) with synthetically mixed training data generated from spectral libraries was used for fraction mapping. We developed three local models based on separate spectral libraries from Berlin (Germany), Brussels (Belgium), and Santa Barbara (U.S.), and a generalized model based on a combined multi-site spectral library. To examine the performance and transferability of the generalized model compared to local models, we first applied all model variants to simulated Environmental Mapping and Analysis Program (EnMAP) data from the three cities that were represented in the models, i.e., known sites. Next, we transferred the models to two unknown sites not represented in the models, San Francisco Bay Area (U.S.) and Munich (Germany). In the first mapping constellation, results demonstrated that the generalized model was capable of accurately mapping VIS fractions across all three known sites. Average mean absolute errors (AV-MAEs) were 8.5, 12.2, and 11.0% for Berlin, Brussels, and Santa Barbara. The performance of the generalized model was very similar to the local models, with Delta AV-MAEs falling within a range of +/- 0.7%. A detailed assessment of fraction maps and class-wise accuracies confirmed that modeling errors related to remaining limitations of urban mapping based on optical remote sensing data rather than to the choice between a local or generalized model. For the second mapping constellation, the generalized model proved to be useful for mapping vegetation and impervious fractions in the unknown sites. MAEs for both cover types were 5.4 and 10.9% for the San Francisco Bay Area, and 6.3 and 15.4% for Munich. In contrast, the three local models were only found to have similar accuracies as the generalized model for one of the two sites or for individual VIS categories. Despite the enhanced transferability of the generalized model to the unknown sites, deficiencies remained for accurate soil mapping. MAEs were 22.4 and 12.3%, and high over - and underestimations were observed at the low and high end of the fraction range. These shortcomings indicated possible limitations of the spectral libraries to account for the spectral characteristics of soils in the unknown sites. Overall, we conclude that the combination of SVR and synthetically mixed training data generated from multi-site libraries constitutes a flexible modeling approach for generalized urban mapping across multiple cities.