Large-scale automated morphologic urban characterization with earth observation and Machine Learning Techniques

Kurzfassung

Through global urbanization and the growth of population, our planet undergoes radical changes. Through the last century, urban areas evolved from spare settled areas to agglomerations of millions of dwellers. Therefore a systematic characterization of these built habitats is an essential step forfundamental understanding of urban habitat. This knowledge, in turn, will contribute to large-scale, yet specific applications. For all these studies, an urban characterization is essential, and fundamental properties of urban morphology are informations of great value. Among the relevant morphologic features, built-up density is of high relevance due to its simplicity, and yet its high potential of information. To get a full morphological description of built-up areas, we also argue for the use of the built up height as a proxy of the vertical dimension of urban areas. Earth observation (EO) data proved to be a beneficial source of information to assess built-up height and density. In this aim, we used TanDEMX data as a support of our morphological characterisation. These data provides elevation information as DSM data of 0.4 arcseconds of resolution. Our study is part of the RIESGOS project, which covers Chile, Peru and Ecuador. In this frame, the work discussed here focus solely on Chile. Therefore, the area covered in this study is more consequent than the limitation of free TanDEM-X data per scientific project (100 000 km2). Additionally, the surface of our study site is to large to rely on VHR multispectral data, as this type of data is expensive. For this reason, we propose here an application of new methods enabling us to overcome the scarcity of TanDEM-X data by combining it with Sentinel-2 data. This technique is based on a regression approach between DSM data of the TDM mission and the optical data of Sentinel-2. The novelty of the regression technique discussed here resides its ensemble approach. We consider model outputs of advanced machine-learning regressors including Neural Networks (NN), Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Random Forest Regression (RFR). Those algorithms are trained with a different version of the AdaBoost.RT algorithm to make the approach more robust. To ensure diversity between the boosted regressors, we introduce a random feature subspace method, thereby we prune none-supportive boosted regressors and get the final prediction through a weighted mean function to achieve enhanced predictions. Finally, the outputs are combined to a single final prediction using a decision fusion strategy, namely stacked generalization. We see the interest of our work in two folds: 1) The combining approach of the ensemble regression allowed us to map built-up density and height of urban area even in case of missing TanDEM-X data, and this with great accuracy; 2) The heuristic approach of multiview pruning allowed us to reduce drastically computation time (1440 hours estimated on the middle part of the country to 760 hours) which provides evident interest in term of temporal survey of urban characteristics.