Enhancing evacuation shelter suitability in compound hazard-prone regions with a Bayesian optimized convolutional neural network approach

Abstract

Effective evacuation planning in hazard-prone regions, particularly during concurrent floods and landslides, remains a significant challenge. This study proposes an innovative methodology to enhance evacuation shelter suitability in response to geo-hydrological hazard scenarios, focusing on vulnerable regions like India's Western Ghats. Our approach integrates susceptibility assessments for floods and landslides, employing Variable Inflation Factor (VIF) analysis to ensure predictor variable independence. Leveraging Convolutional Neural Networks (CNN) augmented with an attention mechanism and refined through Bayesian optimization, our model accurately predicts landslide and flood susceptibility, facilitating robust evacuation planning. Additionally, our methodology incorporates constraint factors and a Fuzzy Analytical Hierarchy Process (F-AHP) to address multi-hazard scenarios, providing a comprehensive framework for emergency response strategies. Implemented in the complex terrain of the Western Ghats, our model demonstrates exceptional accuracy metrics, with optimized CNN models achieving over 91 % accuracy, precision, and recognition rate. Evacuation zone mapping identifies highly suitable areas (18 %), unsuitable areas (47 %), and moderate to low suitability zones (14 % and 21 % respectively), offering valuable insights for resilience-building efforts in hazard-prone regions. This integrative approach not only precisely identifies hazard-prone areas, but also provides a robust framework for refining emergency response strategies, contributing to sustainable development and resilience in vulnerable environments.