Rainfall-induced landslides are common natural disasters in many mountainous cities and pose great threats to life and property. To mitigate landslide risk, building a prediction model to provide both spatial and temporal probabilities of landslide occurrence is essential but challenging. The proper assessment of landslide risk requires both good landslide records and suitable modeling tools. However, the critical data of spatial location and failure time of landslides in developing a spatiotemporal prediction model is often unavailable. Machine learning (ML) has recently emerged as a powerful data-driven tool in landslide science, especially in landslide susceptibility mapping, and been proven promising to correlate landslide occurrence with multiple controlling factors. Additionally,...[ Read more ]

Rainfall-induced landslides are common natural disasters in many mountainous cities and pose great threats to life and property. To mitigate landslide risk, building a prediction model to provide both spatial and temporal probabilities of landslide occurrence is essential but challenging. The proper assessment of landslide risk requires both good landslide records and suitable modeling tools. However, the critical data of spatial location and failure time of landslides in developing a spatiotemporal prediction model is often unavailable. Machine learning (ML) has recently emerged as a powerful data-driven tool in landslide science, especially in landslide susceptibility mapping, and been proven promising to correlate landslide occurrence with multiple controlling factors. Additionally, data scarcity and imbalance issues are the frequent issues encountered when dealing with landslide prediction. The application of machine learning techniques to mitigate these issues needs to be explored.

Comprehensive databases of rainstorms, landslides and various slope features in Hong Kong are compiled and integrated storm by storm to link slope failures to their triggering rainstorms. Using this unique dataset, a spatiotemporal prediction model for natural terrain landslides is firstly developed and its predictive ability is validated against two historical rainstorms. To further improve the prediction performance for natural terrain landslides, a transfer learning model is established based on neural network by incorporating another complete landslide dataset (i.e. Enhanced Natural Terrain Landslide Inventory). For man-made slopes, a frequency-based landslide prediction model using weighted neural network is developed and compared with Geotechnical Engineering Office (GEO)'s landslip warning system. Furthermore, a probabilistic prediction model for man-made slopes is developed from highly imbalanced dataset using data sampling and its prediction performance is validated through a case study.

The proposed ML-based prediction model for natural terrain landslides incorporates real-time rainfall conditions for landslide forecasting successfully. Dynamic susceptibility maps are produced and proven satisfactory to capture distinct rainstorm characteristics (i.e., rainfall spatial distribution and intensity). The application of transfer learning has further improved the prediction performance demonstrated through the loss and accuracy curves during training and four performance metrics during testing. To gain a significant performance improvement, a source dataset with at least eight times larger than the target dataset is suggested for model building and transferring. For man-made slopes, weighted neural network accounts for statistical uncertainty of landslide frequency and establishes reasonable correlations between failure frequency and multiple controlling factors. The prediction model outperforms GEO's landslip warning system (i.e., the R² score improves from 0.79 to 0.90). Data sampling methods improves the accuracy of machine learning methods for highly imbalanced landslide dataset. SMOTEBoost using over-sampling could achieve a high precision of 65% and recall of 45%. Failure probability of man-made slopes is estimated on slope-level and the spatial distributions of observed and predicted landslide locations show high consistency. Slope angle, 4-h and 12-h maximum rolling rainfall factors are proven to be the most critical predicting factors for both natural terrain and man-made slopes. Slope type is another important factor need to be considered when deal with man-made slope failure prediction.