Objective  To employ machine learning methods to analyze various emergency-related information following an earthquake, taking Yunnan as an example, in order to predict emergency response levels and support decision-making by emergency rescue agencies and experts, thereby improving response efficiency and effectiveness.

Methods  A total of 226earthquake events in Yunnan Province from 2014 to 2023 recorded in the Global Natural Disaster Repository of the National Cryosphere Desert Data Center(China) were used as the dataset. Model input features included seismic intensity, focal depth, soil moisture content, population density from the previous year, regional GDP from the previous year, economic losses caused by the earthquake, and affected area. The output was the classification into one of four emergency response levels. A support vector machine(SVM) classification model optimized using Bayesian optimization(BO) was constructed to predict earthquake emergency response levels. The linear, quadratic, cubic, and Gaussian kernel functions were used for training to identify the most suitable kernel for predicting earthquake emergency response levels.

Results  The SVM model with a quadratic kernel demonstrated superior uniformity and overall performance, achieving the best results after Bayesian optimization. The classification model accurately predicted emergency response levels in Yunnan Province. The AUC values for each level were as follows: Level I was 0.863 8, Level Ⅱ was 0.862 8, Level Ⅲ was 0.810 7, and Level Ⅳ was 0.874 0. On the test set, the PR-AUC values were 0.305 5 for Level I, 0.791 6 for Level Ⅱ, 0.830 4 for Level Ⅲ, and 0.897 2 for Level Ⅳ, with a macro-average F1-score of 0.796 7 at the model's operating point.

Conclusion  The emergency response level was governed by a nonlinear function. The Bayesian optimized-SVM approach with a quadratic kernel showed strong potential for application in predicting earthquake emergency response levels and can effectively support emergency preparedness plans. This approach provides a promising method for developing intelligent disaster response decision support systems. Future research should explore data augmentation techniques and expanded data sources to enhance the comprehensiveness of input features and improve the prediction accuracy for higher-level emergency responses.