A Hybrid GeoAI Framework Integrating Smile Random Forest for Comprehensive Seismic Vulnerability Assessment: A Case Study of Mersin, Türkiye
Abstract
This study addresses the critical issue of earthquake vulnerability in Mersin, Türkiye, given that Mersin city and its surrounding districts are highly susceptible to seismic threats due to high population and building density, substandard and low-durability structures, narrow road networks, and severe urban congestion. The research employs a comprehensive approach, utilizing a multi-criteria evaluation model and a novel hybrid random forest model to estimate the city's vulnerability proportionally. Spatial data encompassing physical, population, building quality, accessibility, relief, and hazard facilities were incorporated into the assessment. Weights for these components were determined through the Analytic Network Process (ANP) model, and a hybrid approach using Linear, Small, and Large functions calculated distances between options with fuzzy-fication. Resampling 10 m × 10 m maps addressed variations in spatial resolutions, while 80% training set and 20% test set mitigated overfitting concerns. Expert opinions were pivotal in establishing criteria and sub-criteria for determining safe areas for temporary accommodation, rescue centers, and a Seismic vulnerability assessment (SVA). The Smile Random Forest hybrid model was instrumental in generating these outcomes. Notably, the vulnerability map indicated that 24% of Mersin's areas fall within the high and very high vulnerability range. Key contributors to vulnerability included Geology factors (26.4), Land use (16.1), Epicenters (13.1), and slope and DEM (6.4). These findings underscore the imperative for strategic planning and interventions to minimize earthquake-induced damage in Mersin.