Eigenvector spatial filtering enhancing natural hazards vulnerability assessment in a susceptible urban environment: A case study of Izmir earthquake in Turkey

Abstract

The increasing risk of earthquakes in urban areas has made it crucial to develop accurate vulnerability models for city infrastructure and systems. We aimed to assess and compare the effectiveness of different models and vulnerability analysis techniques in predicting earthquake vulnerability in the specific context of Izmir, Turkey. One central hypothesis in this research aimed to determine whether integrating Eigenvector Spatial Filtering (ESF) into both regression models and machine learning algorithms would yield a comparable enhancement in model performance. We performed earthquake vulnerability modeling (EVM) by considering (ⅰ) only seismic-related variables (SRV) and (ⅱ) integrating ESF by using Moran's eigenvector maps (MEMs). For each approach, we evaluated the predictive performance of two simple regression-based models; generalized linear model (GLM) and generalized additive model (GAM), and two complex machine learning ones; generalized boosting model (GBM), and random forest (RF). The study utilized five primary indicators encompassing geotechnical, physical, structural, social, and facilities data. The predictive performance of the models was assessed using evaluation metrics including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and adjusted R2. The results indicated that the optimal candidate model consisted of five key variables: altitude, building height, distance to safety gathering places, Peak Ground Acceleration (PGA), and population density. We found that decision-tree-based methods performed better than regression-based methods for both modeling schemes. RF exhibited the highest predictive performance for the training data (RMSE = 0.59, adjusted R2 = 0.71), while GBM outperformed other models for the test data (RMSE = 0.79, adjusted R2 = 0.78). However, incorporating ESF to the EVM analysis revealed that regression-based methods, particularly the GLM, obtained highest improvement in accuracy (RMSE 0.94 vs 0.76 and adjusted R2 0.56 vs 0.71 for the SRV and SRV + MEMs modeling approach). Significant differences were observed between GLM-GBM and GLM-RF comparisons, as well as GAM-GBM and GAM-RF comparisons. The findings of this research are expected to be helpful for informed decision-making, targeted risk reduction, and the development of effective policies and strategies to enhance preparedness and resilience in the face of seismic events in highly susceptible urban systems.