Structural Type Identification of City-Scale Building Groups Based on Deep Learning

Accurate characterization of structural types within building groups holds critical significance for urban planning strategies and disaster resilience frameworks. Current methodologies relying on field investigations present substantial challenges in labor allocation and temporal efficiency. This study develops a deep learning–based method for automated structural type classification using building footprint geometry and height data. The methodology generates contextual feature imagery encompassing both target structures and their spatial surroundings to establish a training data set. To address inherent class imbalance issues, this study implements optimized oversampling techniques during model development. An iterative training procedure incorporating prior classification outcomes progressively enhances prediction accuracy. The proposed method features encouraging accuracy and efficiency. Validation experiments were carried out in several cities in China. In Tangshan City (213,020 buildings) the method achieved 98% classification accuracy. Its application in Tianjin City (262,305 buildings) demonstrated processing capabilities of 100,000 buildings within 8 min. This computationally efficient approach establishes a paradigm for city-scale infrastructure analysis while maintaining high precision standards, offering actionable insights for urban renewal and disaster risk reduction.