A new machine learning based framework for seismic damage ratio prediction of building clusters using satellite remote sensing data

Seismic-induced building damage prediction is essential for disaster mitigation and risk assessment. However, traditional methods often require detailed structural parameters that are labor-intensive to obtain, while post-earthquake remote sensing is frequently hindered by data latency. To address these limitations, we propose a novel framework utilizing pre-event multi-source satellite data to predict damage, employing building height and construction period as proxies for structural types. Taking the 2023 Türkiye Mw 7.8 earthquake as a case study, we modeled the weighted building cluster damage ratio (WBCDR) using 26 predictors derived from remotely sensed building key attributes, seismic parameters, topography, and spectral-textural features. Random Forest (RF) and Support Vector Regression (SVR) models were evaluated at spatial scales of 300 m, 500 m, and 1 km. RF outperformed SVR, with the best mean test performance at 500 m across ten splits (R2 = 0.8248, Mean Absolute Error (MAE) = 0.0482, Root Mean Squared Error (RMSE) = 0.0916). Notably, Pseudo-Spectral Acceleration at 0.3 s (PSA03), Peak Ground Acceleration (PGA), and textural Dissimilarity were identified as dominant influencing factors. The proposed framework enables rapid and effective large-scale seismic building damage prediction in areas lacking detailed structural information, providing valuable support for seismic risk assessment and emergency response.