Estimation of seismic building structural types using remote sensing and machine learning

Abstract

The current trend of urbanization leads to an increase of seismic vulnerability in earthquake prone regions. There is great demand for methods contributing to a comprehensive analysis of seismic vulnerability to face the urgent challenges of mitigation and catastrophe management. Remote sensing has high potential to contribute to an area-wide and up-to-date assessment of seismic vulnerability. For an estimation of building stock damage the built-inventory is generally categorized into different seismic building structural types, representing a construction's seismical behavior. This study reveals indirect correlations between remotely sensed data and seismic building structural types, which enable a supervised classification. Site of research is the City of Padang, Indonesia, whose urban environment is characterized by 145 features calculated by means of high resolution optical imagery, height information from a normalized digital surface model and multi-temporal medium resolution optical data. In-situ building information is given through survey data collected after the earthquake event of September 2009. Using Machine Learning techniques a work flow is presented to classify seismic building structural types. A feature selection analysis is carried out, and the features most explanatory for the determination of seismic building structural types are identified. Coping with large amounts of features and in-situ data scarcity, plausible classification results are achieved and dependencies between remotely sensed data and building stability are verified.