Christian Geiß, 1; Anne Schauß, 1,4; Torsten Riedlinger, 1; Stefan Dech, 1; Cecilia Zelaya, 2; Nicolás Guzmán, 2; Matías A. Hube, 3; Jamal Jokar Arsanjani, 4; Hannes Taubenböck, 1.
1 German Aerospace Center (DLR), German Remote Sensing Data Center (DFD), Münchner Straße 20, 82234 Oberpfaffenhofen-Weßling, Germany
2 Chilean Navy Hydrographic and Oceanographic Service (SHOA), Errázuriz Echaurren 254 Playa Ancha, Valparaiso, Chile
3 Pontificia Universidad Católica de Chile and National Research Center for Integrated Natural Disaster Management CONICYT/FONAP/15110017, Vicuña Mackenna, 4860 Santiago, Chile
4 Heidelberg University, GIScience Research Group, Berliner Straße 48, 69120 Heidelberg, Germany
The impact of natural hazards on mankind has increased dramatically over the past decades. Global urbanization processes and increasing spatial concentrations of exposed elements induce natural hazard risk at a uniquely high level. To mitigate affiliated perils requires detailed knowledge about elements at risk. Considering a high spatiotemporal variability of elements at risk, detailed information is costly in terms of both time and economic resources and therefore often incomplete, aggregated, or outdated. To alleviate these restrictions, the availability of very-high-resolution satellite images promotes accurate and detailed analysis of exposure over various spatial scales with large-area coverage. In the past, valuable approaches were proposed; however, the design of information extraction procedures with a high level of automatization remains challenging. In this paper, we uniquely combine remote sensing data and volunteered geographic information from the OpenStreetMap project (OSM) (i.e., freely accessible geospatial information compiled by volunteers) for a highly automated estimation of crucial exposure components (i.e., number of buildings and population) with a high level of spatial detail. To this purpose, we first obtain labeled training segments from the OSM data in conjunction with the satellite imagery. This allows for learning a supervised algorithmic model (i.e., rotation forest) in order to extract relevant thematic classes of land use/land cover (LULC) from the satellite imagery. Extracted information is jointly deployed with information from the OSM data to estimate the number of buildings with regression techniques (i.e., a multilinear model from ordinary least-square optimization and a nonlinear support vector regression model are considered). Analogously, urban LULC information is used in conjunction with OSM data to spatially disaggregate population information. Experimental results were obtained for the city of Valparaı ́so in Chile. Thereby, we demonstrate the relevance of the approaches by estimating number of affected buildings and population referring to a historical tsunami event.