Real-time early warning of landslide disaster risks on major highways in Ganzhou City, China

Extreme weather events like heavy rainfall have become more frequent recently, increasing the occurrence of landslides and slope instability along mountainous highways and threatening transportation safety. This research aims to develop an effective real-time early warning system for highway landslides triggered by extreme weather. Using landslides along Ganzhou’s major highways as a case study, a 250-m buffer zone was established along the roads, within which 88,497 slope units were identified using multi-scale segmentation. Subsequently, 1547 landslide samples and 18 conditioning factors were collected for landslide susceptibility prediction (LSP) based on random forest (RF), C5.0 decision tree (DT), and support vector machine (SVM) models. Model performance was evaluated using receiver operating characteristic (ROC) curves, the distribution characteristics of the landslide susceptibility index (LSI), and a confusion matrix. A continuous probability rainfall threshold equation was then fitted using data from the rainfall station. Subsequently, the analytic hierarchy process (AHP) method was employed to assess highway vulnerability. Finally, by integrating LSI, rainfall thresholds, vulnerability, and disaster-bearing entities, effective early warning was achieved for two typical landslide cases. Results indicate that the RF model yielded the best LSP outcomes, with an of 0.958, an RMSE of 0.069, and a sum of squared residuals of 0.331 for the continuous probability equation. The hazard assessment reached 90.4% accuracy, with hazard values exceeding 0.8 in both typical cases. AHP analysis, validated by expert experience and consistency tests, identified slope, road density, and road grade as key vulnerability factors. Ultimately, real-time risk early warning for typical landslide events was achieved by incorporating population distribution and economic value.