Data-driven landslide forecasting: Methods, data completeness, and real-time warning

• Published in: Engineering geology 2023-05, Vol.317, p.107068, Article 107068
• Main Authors: Xiao, Te, Zhang, Li-Min
• Format: Article
• Language: English
• Subjects: China; data collection; Data-driven methods; decision support systems; Landslide forecasting; Landslide risk; landslides; Machine learning; prediction; Rain-induced landslides; risk; risk management
• ISSN: 0013-7952; 1872-6917
• DOI: 10.1016/j.enggeo.2023.107068

Various data-driven methods, including empirical, statistical, and machine learning methods, have been developed to promptly forecast rain-induced landslides. Their abilities differ considerably in spatio-temporal landslide prediction and in handling datasets of varying qualities. A challenging issue that significantly hinders the applications of data-driven methods is the data incompleteness in most landslide inventories, particularly the lack of accurate landslide time that is a vital link between each landslide and its triggering rainstorm. This study systematically compares the performances of three categories of data-driven methods for landslide prediction and proposes a novel machine learning model featured by probabilistic landslide modelling for spatio-temporal landslide prediction. The integrated machine learning model can be developed on a realistic landslide database, regardless of whether the landslide timing information is known or not. It not only promptly predicts the spatio-temporal evolution of landslides during a rainstorm but also reliably characterises the factual landslide risk, which provides a powerful real-time decision-making tool for landslide early warning and risk management. The model is validated against the landslide incidents in Hong Kong in the past 35 years both spatially and temporally, and outperforms other data-driven models in both prediction ability and accuracy. •The performances of three categories of data-driven methods for landslide prediction are evaluated.•A novel machine learning model featured by probabilistic landslide modelling is proposed.•The ‘at least one failure’ strategy enables the use of incomplete data in landslide forecasting.•Machine learning-empowered spatio-temporal prediction enables real-time landslide warning.•The machine learning method outperforms other data-driven methods in prediction accuracy and ability.