Data-Driven Earthquake Multi-impact Modeling: A Comparison of Models

Abstract

AbstractIn this study, a broad range of supervised machine learning and parametric statistical, geospatial, and non-geospatial models were applied to model both aggregated observed impact estimate data and satellite image-derived geolocated building damage data for earthquakes, via regression- and classification-based models, respectively. For the aggregated observational data, models were ranked via predictive performance of mortality, population displacement, building damage, and building destruction for 375 observations across 161 earthquakes in 61 countries. For the satellite image-derived data, models were ranked via classification performance (damaged/unaffected) of 369,813 geolocated buildings for 26 earthquakes in 15 countries. Grouped k-fold, 3-repeat cross validation was used to ensure out-of-sample predictive performance. Feature importance of several variables used as proxies for vulnerability to disasters indicates covariate utility. The 2023 Türkiye–Syria earthquake event was used to explore model limitations for extreme events. However, applying the AdaBoost model on the 27,032 held-out buildings of the 2023 Türkiye–Syria earthquake event, predictions had an AUC of 0.93. Therefore, without any geospatial, building-specific, or direct satellite image information, this model accurately classified building damage, with significantly improved performance over satellite image trained models found in the literature.