Explainable seismic damage prediction model based on CELS-WOA-Stacking

Rapid and accurate assessment of post-earthquake building damage is crucial for effective emergency response and recovery. This study aims to enable knowledge-guided decision support in seismic disaster scenarios and to examine the relationships between influencing factors and seismic damage. Based on the Nepal Earthquake Damage Database (NBDP), we developed a hybrid earthquake damage prediction model that combines an integrated stacking framework with an improved whale optimization algorithm (Chaotic-Lévy Elite-enhanced Whale Optimization Algorithm, CELS-WOA). To address model complexity in resource-constrained post-disaster environments, we employed committee voting for feature selection and quantified the contribution of each variable using Shapley Additive Explanations (SHAP). The results demonstrate that our model outperforms single models and traditional ensemble approaches under various optimization strategies, achieving a prediction accuracy of 0.904. Additionally, we analyzed the SHAP values of regional variables in conjunction with seismic intensity information, uncovering latent mechanisms of seismic intensity encoding through geospatial attributes. Key indicators affecting seismic damage include pre/post-earthquake building height/floors, geographic location, construction age, and mud-stone material usage. Specifically, building height/floors directly reflect damage, while geographic location influences damage through hidden earthquake intensity. Construction age and mud-stone material influence the damage of the building by reflecting their own ability to resist vibration. This study clarifies the impact of influencing factors on the earthquake loss of buildings based on machine learning, which provides a basis for intelligent loss assessment for earthquake disaster scenarios.