Artificial intelligence-based assessment of large-scale building safety risk in non-disaster scenarios by integrating InSAR-derived subsidence and environmental remote sensing data

Conventional building safety assessment methods, such as civil engineering field surveys and IoT-based structural health monitoring, face critical limitations, including restricted applicability to a multitude of buildings, reliance on costly specialized equipment, and low operational efficiency. Recent approaches leveraging synthetic aperture radar (SAR) imagery, light detection and ranging (LiDAR) or building information modeling (BIM), have primarily focused on detecting damaged buildings in post-disaster scenarios. However, proactive building safety risk assessment in non-disaster settings is essential for early identification of potentially unsafe buildings, while integrated AI based large-scale assessment combining building attributes and remote sensing data remains limited. To address these challenges, a comprehensive index system integrating multi-source data— building census information, subsidence and tilt measurements derived from persistent scatterer interferometric synthetic aperture radar (PS-InSAR) technology, and environmental factors derived from remote sensing data—was developed. Furthermore, this study proposes a novel framework integrating remote sensing, machine learning (ML), and deep learning (DL) for large-scale building safety assessment in non-disaster settings. Feature engineering employed Pearson correlation analysis for initial screening and random forest for robust feature selection. The performance of multiple ML (SVM, RF, LightGBM) and DL (MLP, CRNN, TabNet) models was evaluated and compared. Results indicated that LightGBM achieved high accuracies (0.91, 0.81, 0. 90) and recalls (0.71, 0.72, 1.00) in identifying cracked buildings across the entire region, urban-rural fringe, and urban area datasets, while TabNet exhibited the best performance in rural area datasets (accuracy: 0.86, recall: 0.71). On-site inspections confirmed the model's effectiveness, with detected cracks consistent with ground observations.