Unprecedented Breakthrough of Landslip Warning System in Hong Kong: Real-Time, Data-Driven and Performance-Based

In managing the landslide risk in Hong Kong, the Geotechnical Engineering Office (GEO) of the Civil Engineering and Development Department has progressively established and maintained high quality inventories of territory-wide landslide-related datasets since the 1980s. Over the past forty-five years, the GEO has used these high-resolution spatio-temporal data to support the technical development of a landslide prediction model as part of the Landslip Warning System in Hong Kong. Amongst others, data-driven analyses using a conventional statistical approach have been pursued to establish rainfall-landslide correlations for man-made slopes. Recently, the GEO has explored the potential application of machine learning (ML) and big data analytics, using datasets from 1996 to 2023, for landslide prediction in Hong Kong.Several common ML algorithms such as XGBoost, Logistic Regression, and Neural Network, are being tested to establish the multivariate and non-linear correlation among a wide range ofpertinent features and the occurrence of landslides on man-made slopes. Domain knowledge of geotechnical and geological engineering was incorporated in the course of developing the ML model. This paper presents the modelling approach and workflow using the XGBoost algorithm through data pre-processing, algorithm selection, feature selection, model training and evaluation. The results indicate apromising predictive performance of the XGBoost model against various evaluation metrics compared with the conventional statistical model, and draw insight into contributing factors of landslide occurrence in Hong Kong.

Keywords: Machine learning, Landslide prediction, Rainfall-landslide correlation, Landslip warning.