An improved method to predict man-made slope failure using machine learning tools

Landslide hazards associated with man-made slopes are increasing due to ageing and extreme weather conditions under a changing climate. To effectively mitigate landslide risks, the implementation of regional landslide early warning systems is desirable, albeit challenging, if not impossible, due to the scarcity of reliable landslide data and suitable predictive tools. In this paper, a thorough analysis has been conducted on reasonably reliable and substantial amounts of historical rainfall data, slope features, and landslide inventory of man-made slope failures in Hong Kong. Four different machine learning methods, namely logistic regression (LR), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBoost), have been employed. The predicted number of landslides from the machine learning methods is compared with the predictions made by the current Landslip Warning System in Hong Kong. The effects of rainfall parameters and slope features on model performance are also investigated. The analysed results show that dynamic rainfall conditions are identified as the most influential factors for predicting man-made slope landslide. A combination of 1 and 12 h maximal rolling rainfall (MRR) demonstrates superior performance compared to relying solely on the 24 h MRR. Therefore, this combination is recommended for predicting man-made slope failures in Hong Kong.