Abstract: In recent years, the increasing frequency of extreme natural disasters has made efficient emergency rescue a critical component of public safety. Traditional rescue team dispatching is often hindered by the high uncertainty of road travel time and on-site rescue time, making rapid and precise allocation difficult. Leveraging multi-source historical and terrain data, this study proposes a prediction–decision integrated optimization method based on pre-trained random forest structural constraints. The method mines decision patterns from historical rescue data, explicitly linearizes the nonlinear tree structure of the random forest, and embeds it into a mixed-integer programming model to construct a data-driven uncertainty optimization framework that balances predictive accuracy, interpretability, and computational efficiency. To address the large-scale and multi-layered structure of pre-trained trees, a truncated embedding strategy based on a general magnitude metric is designed, which significantly reduces solution time while maintaining near-optimal performance. Numerical experiments show that the proposed model substantially outperforms traditional methods in both rescue cost control and response time reduction, demonstrating strong robustness and generalizability. This research provides a new paradigm and practical support for emergency dispatch optimization under uncertainty.

Key words: Emergency rescue dispatch, Prediction–decision integration, Random forest, Constraint embedding, Mixed-integer programming