Machine Learning-based Natural Terrain Landslide Susceptibility Analysis – A Pilot Study
Recently, the Geotechnical Engineering Office has initiated a pilot study on data-driven landslide susceptibility analysis (LSA) using a machine learning (ML) approach. A study area covering about one-fifth of the total natural hillside area of Hong Kong on and around the Lantau Island was considered. Three common tree-type ML classifiers: Decision Tree, Random Forest and XGBoost have been used. Conditioning factors (or features) including rainfall, geological and topography-related features were considered. In the study, the domain knowledge on natural terrain landslides in Hong Kong were critically incorporated into the susceptibility models through feature engineering to ensure that the resulted models are physically meaningful. In addition, an approach proposed to resolve the serious data imbalance problem, which is common in LSA, will be highlighted. Under this approach, the predicted probabilities of the positive class (i.e., landslide) can also be taken as a proxy to the landslide probability. This paper reports the methodology and key findings of this pilot study. The approach can be extended to cover other ML algorithms and features, and to a territory-wide scale with a view to enhancing the resolution and accuracy of the current susceptibility model of natural hillsides in Hong Kong.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.