Boosted machine learning ensemble regression with decision fusing strategy for mapping built-up height and built-up density with openstreetmap data and SENTINEL-2 imagery

Abstract

Detailed characterization of urban environments for large areas is a challenging but crucial task for e.g., analyzing urbanization processes and assessing vulnerability and risks with respect to natural hazards, among others. Recent Earth Observation missions feature a notable tradeoff between a fairly high spatial resolution and large area coverage. In particular, ESA’s Sentinel-2 satellites provide superspectral imagery with a spatial resolution of 10 meters for the bands covering visible light and near infrared. Notably, Sentinel-2 data are provided free of charge to the public via a data hub, which is accessible online. In parallel, open-source geoinformation compiled by volunteers such as OpenStreetMap (OSM) can provide valuable information for characterization of urban environments. Numerous cities around the globe were already mapped in detail. Such data sets allow for instance the computation of built-up heights and built-up densities. However, spatially continuous and detailed OSM data are not available everywhere. To allow for truly large-area application, we combine OSM data and Sentinel-2 imagery and render the mapping of built-up heights and built-up densities as a supervised learning problem. Given the rational level of measurement of the two variables to be predicted, we generate a regression model based on Areas where data from both OSM and Sentinel-2 are available in order to estimate the target variables for areas where only Sentinel-2 data is available: The regression estimation problem is regarded as finding the mapping between an incoming vector (i.e., ubiquitously available image features from Sentinel-2) and an observable output from a given set of samples (i.e., automatically derived over spatially limited areas where data from both OSM and Sentinel-2 are available). From a methodological point of view, we introduce a novel regression approach to account for the No-free-Lunch-Theorem, which states that there is no algorithm that induces the most accurate learner in any domain, all the time. As such, we combine the model outputs of advanced machine learning-based Regression algorithms incorporating Support Vector Regression, Regression Trees, Gaussian Process Regression, and Neural Networks within a decision fusion framework. We extend this method within an ensemble learning approach, and learn individual models based on bootstrap Aggregation from the training data to further increase accuracy and robustness of predictions. Experimental results from the city of Cologne, Germany, underline the viability of the approach.