Abstract: Rainfall is the main cause of geological disasters, and frequent rainfall in Yunnan Province leads to frequent geological disasters, posing a serious threat to the safety of people’s lives and property. Geological disaster meteorological risk forecasting is an effective means of disaster prevention and reduction. This study takes the high mountain canyon area of Nujiang Prefecture as an example, constructs information thresholds based on information value models, optimizes samples based on information thresholds, evaluates the susceptibility of comprehensive geological disasters in Nujiang Prefecture using machine learning models, calculates the effective rainfall coefficient of Nujiang Prefecture to establish a meteorological risk forecasting model, and verifies the accuracy of the model using historical disaster points. The results showed that the AUC values of landslide and debris flow disaster evaluation models optimized by information threshold samples were 0.97 and 0.99, respectively, with prediction accuracies of 0.93 and 0.98. The comprehensive geological disaster risk in Nujiang Prefecture was extremely high, and high-risk areas were mainly distributed along rivers and roads in the canyon. The meteorological risk warning model had a forecast hit rate of 90.91%, a missed report rate of 0, and a false alarm rate of 22.22%. The area of high-risk areas at the end of rainfall was 472.24 km². Optimizing samples with information thresholds has greatly improved the predictive and generalization capabilities of machine learning models, and the meteorological forecast model with a decay coefficient of 0.5 has improved the accuracy of geological disaster meteorological risk forecasting. The overall research results can provide substantial support for disaster prevention and reduction work in Nujiang Prefecture and similar areas.

Key words: geological disasters, information threshold, optimized samples, machine learning, rainfall attenuation coefficient, meteorological risk forecasting