Optimal factor combinations selection in typhoon-induced landslides susceptibility mapping using machine learning interpretability

Typhoon is a key dynamic factor triggering landslides. In view of the fact that the previous susceptibility evaluation models rarely consider the interaction between typhoon and static factors, carry out research on the optimal dynamic and static factors combination of typhoon-induced landslides susceptibility. Using the interpretability of machine learning, the importance ranking of dynamic and static factors is carried out to identify key impact factors. On this basis, the importance of static factors under the influence of typhoon is compared, and the interaction between typhoon and static factors is analyzed. Finally, the optimal combination of dynamic and static factors is proposed by using k-fold cross-validation method and taking the average descent accuracy as the index. The results show that the importance of the key influencing factors of typhoon-induced landslide from high to low mainly includes: elevation, NDVI, road and other factors; the addition of typhoon and rainstorm factors significantly increased the importance of factors susceptible to typhoon, such as water system and vegetation, with an increase rate of 24.8-151.7 %. The optimal dynamic and static factors combination of typhoon rainstorm landslide includes all key static factors and four dynamic factors, among which the dynamic factors are: maximum sustained wind speed, rainfall, distance from typhoon center and near gale wind circle radius. The results of ROC curve verification show that the selection of the optimal factor combination can increase the accuracy of the evaluation model by 1.5%-3.5%, which can significantly improve the accuracy and rationality of the susceptibility mapping of typhoon-induced landslides. Keywords: Impact factor, Typhoon, Landslides susceptibility, Interpretability of machine learning.