Multi-source data and machine learning supporting high-resolution building exposure extraction

Recent urban earthquakes and rapid urbanization have intensified the demand for fine-scale building exposure information in disaster risk assessment. However, existing approaches for high-resolution building exposure extraction often suffer from limited data completeness, insufficient semantic detail, and weak update capability, particularly at detailed spatial scales. Moreover, traditional methods relying on homogeneous data sources and static classifications struggle to represent the heterogeneity of urban building exposure.

To address these limitations, we propose a multi-source data-driven framework combined with machine learning to extract high-resolution building exposure information, focusing on building function and building height. Building function types are inferred by integrating OpenStreetMap building footprints with time-series mobile signaling data, exploiting differences in population activity patterns across day-night and workday-non-workday periods. Machine learning techniques are then applied to identify clusters of buildings with similar population dynamic characteristics, enabling the inference of building function types. Building height is extracted from bi-temporal Sentinel-2 imagery by capturing variations in image brightness induced by seasonal differences in building shadow length, and a random forest model is employed to learn the nonlinear relationship between image features and building height, thereby reducing reliance on very high-resolution imagery and manual interpretation.

Case studies in representative Chinese cities indicate that the integration of multi-source data and machine learning enables more effective use of data for different building exposure attributes, resulting in improvements in spatial detail, attribute completeness, and data timeliness. Population-dynamic-based building function identification provides an activity-oriented characterization of building use, while building height estimation based on freely available Sentinel-2 imagery offers a cost-efficient and scalable approach. Overall, these findings suggest that multi-source data integration and machine learning can support large-scale, high-resolution urban building exposure mapping.