Public safety and health cannot be secured without the comprehensive recognition of characteristics and reliable emergency response schemes under the disaster chain. Distinct from emergency resource allocation that focuses primarily on a single disaster, dynamic response, periodic supply, and assisted decision-making are necessary. Therefore, we propose a multiobjective emergency resource allocation model considering uncertainty under the natural disaster chain. Resource allocation was creatively combined with path planning through the proposed multiobjective cellular genetic algorithm (MOCGA) and the improved A* algorithm with avoidance of unexpected road elements. Furthermore, timeliness, efficiency, and fairness in actual rescue were optimized by MOCGA. The visualization of emergency trips and intelligent avoidance of risk areas were achieved by the improved A* algorithm. The effects of logistics performance, coupling of disaster factors, and government regulation on emergency resource allocation were discussed based on different disaster chain scenarios. The results show that disruption in infrastructure support, cascading effect of disasters, and time urgency are additional environmental challenges. The proposed model and algorithm work in obtaining the optimal solution for potential regional coordination and resilient supply, with a 22.2% increase in the total supply rate. Cooperative allocation complemented by political regulation can be a positive action for successfully responding to disaster chains.

Traffic accidents occur frequently, causing significant losses to people’s lives and property safety. Accurately predicting the severity level of traffic accidents is of great significance. Based on traffic accident data, this study comprehensively considers various influencing factors such as the geographical location, road conditions, and environment. The data are divided into accident-related categories, weather-related categories, and road- and environment-related categories. The machine learning method is improved through integration for the accident level prediction. In the experiment, effective preprocessing measures were taken for problems such as data imbalance, missing values, the encoding of categorical variables, and the standardization of numerical features. The unbalanced distribution of “Severity” was improved through under-sampling and over-sampling techniques. Firstly, we adopted a multi-stage fusion strategy. A multi-layer perceptron (MLP) was used for the preliminary prediction, and then its result was combined with the original features to form a new feature. Decision tree, XGBoost, and random forest algorithms, respectively, were applied for the secondary prediction. The analysis results show that the improved machine learning model is significantly superior to a single model in the overall performance. The “MLP + random forest” model performs well in evaluation indicators such as the accuracy, recall rate, and F1 value. The accuracy rate is as high as 94%. In the prediction of different traffic accident severity levels (minor, moderate, and severe), the improved machine learning model also generally shows better performance and stability. The research results of this study have broad prospects in the field of intelligent driving. It can realize real-time accident prediction and early warnings, and provide decision support for drivers and autonomous driving systems. The research also provides a scientific basis for traffic planning and management departments to improve driving conditions and reduce the probability and losses of traffic accidents.