Earthquake fatality prediction based on hybrid feature importance assessment: a case study in Yunnan Province, China

Reliable earthquake fatality prediction is an important reference for post-earthquake emergency response efforts. Seismic data are the basis for constructing earthquake casualty prediction models, but the selection and evaluation of earthquake features are more critical due to the scarcity of destructive earthquake samples. In order to make full use of the high-dimensional survey data of destructive earthquake disasters in the Earthquake Reports in Yunnan Province since 1992, and effectively use it to improve the ability to predict the number of earthquake casualties, this paper proposes a hybrid feature importance evaluation method based on four conventional feature contribution methods (IG, PPMCC, SRCC and MDI), ranking the importance of 63 features that affect the number of earthquake casualties in Yunnan Province, and reducing the feature dimension accordingly. Then, cross-validation is used to compare the accuracy of the four machine models before and after dimensionality reduction. We found that (1) among the 10 features with the highest hybrid importance, there were 8 population distribution features, 1 geological hazard feature (number of landslides) and 1 damage degree feature (highest intensity of earthquakes); (2) the feature dimensionality reduction based on the importance of hybrid features can effectively improve the prediction accuracy of machine learning models; and (3) in the comparison of several methods, the Particle Swarm Optimized Support Vector Machine model had the highest prediction accuracy, with an R2 over 0.934. The research results showed that this method can significantly improve the prediction accuracy of the machine learning model and has some reference value for earthquake emergency rescue and post-disaster reconstruction work.