The ongoing global transformation of human habitats from rural villages to ever growing urban agglomerations induces unprecedented seismic risks in earthquake prone regions. To mitigate affiliated perils requires the seismic assessment of built environments. Numerous studies emphasize thatremote sensing can play a valuable role in supporting the extraction of relevant features for preevent vulnerability analysis. However, the majority of approaches operate on building level. This induces the deployment of very high spatial resolution remote sensing data, which hampers, nowadays, utilization capabilities for larger areas due to data costs and processing requirements. In this paper, we alter the spatial scale of analysis and propose concepts and methods to estimate the seismic vulnerability level of homogeneous urban structures. A procedure is designed, which comprises four main steps dedicated to: 1) delineation of urban structures by means of a tailored unsupervised data segmentation procedure with scale optimization; 2) characterization of urban structures by a joint exploitation of multisensor data; 3) selection of most feasible features under consideration of in situ vulnerability information; and 4) estimation of seismic vulnerability levels of urban structures within a supervised learning framework. We render the prediction problem in three ways to address operational requirements that can evolve in real-life situations. 1) To discriminate two or more classes based on labeled samples of all classes present in the data under investigation, we use the framework of soft margin support vector machines (C-SVM). 2) To consider situations, where solely labeled samples are available for the class(es) of interest and not for all classes present in the data, we deploy ensembles of-one-class SVM (-OC-SVM). and 3) To fit data with a higher statistical level of measurement (interva- or ratio scale), we utilize a support vector regression (SVR) approach to estimate a regression function from the training samples.
Experimental results are obtained for the earthquake-prone mega city Istanbul, Turkey. We use multispectral data from the RapidEye constellation, elevation measurements from the TanDEM-X mission, and spatiotemporal analyses based on data from the Landsat archive to characterize the urban environment. In addition, different in situ data sets are incorporated for Istanbul’s district Zeytinburnu and the residual settlement area of Istanbul. When estimating damage grades for Zeytinburnu with SVR, best models are characterized by mean absolute percentage errors less than 11%, and fairly strong goodness of fit ( {mathbf{0}}.{mathbf{75}}$ ?>). When aiming to identify different types of urban structures for the remaining settlement area of Istanbul (i.e., urban structures determined by large industrial/commercial buildings and tall detached residential buildings, which can be considered here as highly and slightly vulnerable, respectively), results obtained with -SVM show a distinctive increase of accuracy compared to results obtained with ensembles of -OC-SVM. The latter were not able to exceed moderate agreements, with statistics slightly above 0.45. Instead, -SVM allowed obtaining statistics expressing substantial and even excellent agreements ( {mathbf{0}}.{mathbf{6}}$ ?>up to {mathbf{0}}.{mathbf{8}}$ ?>). Overall, analyzes provide very promising empirical evidence, which confirms the potential of remote sensing to support seismic vulnerability assessment.