Mapping essential urban land use categories with open big data: Results for five metropolitan areas in the United States of America

Urban land-use maps outlining the distribution, pattern, and composition of various land use types are critically important for urban planning, environmental management, disaster control, health protection, and biodiversity conservation. Recent advances in remote sensing and social sensing data and methods have shown great potentials in mapping urban land use categories, but they are still constrained by mixed land uses, limited predictors, non-localized models, and often relatively low accuracies. To inform these issues, we proposed a robust and cost-effective framework for mapping urban land use categories using openly available multi-source geo-spatial "big data". With street blocks generated from OpenStreetMap (OSM) data as the minimum classification unit, we integrated an expansive set of multi-scale spatially explicit information on land surface, vertical height, socio-economic attributes, social media, demography, and topography. We further proposed to apply the automatic ensemble learning that leverages a bunch of machine learning algorithms in deriving optimal urban land use classification maps. Results of block-level urban land use classification in five metropolitan areas of the United States found the overall accuracies of major-class (Level-I) and minor-class (Level-II) classification could be high as 91% and 86%, respectively. A multi-model comparison revealed that for urban land use classification with high-dimensional features, the multi-layer stacking ensemble models achieved better performance than base models such as random forest, extremely randomized trees, LightGBM, CatBoost, and neural networks. We found without very-high-resolution National Agriculture Imagery Program imagery, the classification results derived from Sentinel-1, Sentinel-2, and other open big data based features could achieve plausible overall accuracies of Level-I and Level-II classification at 88% and 81%, respectively. We also found that model transferability depended highly on the heterogeneity in characteristics of different regions. The methods and findings in this study systematically elucidate the role of data sources, classification methods, and feature transferability in block-level land use classifications, which have important implications for mapping multi-scale essential urban land use categories.