Robustness study of seismic hazard assessment models based on multi-source data

Traditional seismic hazard assessment models mainly focus on seismic conditions, often neglecting topography, geology, and environmental factors. The limited robustness and “black-box” nature of single machine learning algorithms pose challenges in reliably assessing seismic hazard. This study leverages multi-source data and five hyperparameter optimization (HPO) algorithms to enhance the robustness of seismic hazard mapping (SHM) models in Jiuzhaigou County, China, by fine-tuning the traditional random forest (RF) model. The SHapley Additive exPlanations (SHAP) method was used to reveal decision-making mechanisms in seismic hazard models, supported by local interpretations and on-site investigations. This study considered 23 factors, including seismic factors, terrain morphology, geological conditions, and environmental conditions. We evaluated the performance of each hybrid model using metrics such as Coefficient of Determination (R² score) and various validation methods. The results show that the Random Forest-Tree-structured Parzen Estimators (RF-TPE) model achieves the highest R² score (0.9877). The RF-TPE outperforms other models in terms of robustness, making the SHM constructed with this model more reliable. The study also found that, in modeling SHM in the region, factors such as Peak Ground Velocity (PGV), hypocentral distance, Peak Spectral Acceleration (PSA), distance to fault, slope, and elevation are the most influential.