Predicting landslide runout paths using terrain matching-targeted machine learning

• Published in: Engineering geology Vol. 311; p. 106902
• Main Authors: Ju, Lu-Yu, Xiao, Te, He, Jian, Wang, Hao-Jie, Zhang, Li-Min
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 20-12-2022
• Subjects: Landslide; Landslide risk; Machine learning; Runout path; Terrain matching; Landslide risk; Landslide; Runout path; Terrain matching; Machine learning
• ISSN: 0013-7952; 1872-6917
• DOI: 10.1016/j.enggeo.2022.106902

Landslide debris will travel certain distances and threaten people and properties along its runout path, highlighting the importance of runout path prediction in landslide risk management. Conventional landslide runout models, either statistical or machine learning-based, only consider the geographic characteristics at the source without fitting the terrain features along the path. A novel terrain matching-targeted machine learning model is proposed for predicting landslide runout paths, in which a consistent terrain matching strategy is introduced for both training and prediction. The model first forward predicts multiple travel distances based on geographic characteristics of all cells along a possible runout path, and then determines the termination cell whose predicted travel distance fits the terrain features best to backward estimate model parameters. Such a terrain matching process not only accounts for geographic characteristics along the paths but also enables the incorporation of three-dimensional terrain reality into model training. A case study of natural terrain landslides in Hong Kong is conducted to validate the proposed machine learning model. Results indicate that the terrain matching-targeted machine learning models significantly outperform conventional statistical models in terms of prediction accuracy. The fall height and landslide scale are the most critical physical factors affecting travel distances of channelized landslides and open hillslope landslides, respectively. The landslide runout model is applied to the Mid-Levels at the foot of Victoria Peak to identify high-risk urban areas vulnerable to landslides, which provides guidelines for designing landslide prevention and mitigation measures.
•A novel machine learning model is proposed for predicting landslide runout paths.•Internal terrain matching is introduced to consider terrain reality and trace features.•Machine learning models have much higher accuracy than statistical models.•The most critical factors affecting landslide travel distances are identified.•The landslide runout model is applied to identify areas vulnerable to landslides.