Back to search

McMaster University

An Integrated Approach Toward Reliable and Adaptive Landslide Early Warning Systems

Abstract

dc:description.abstract

Landslides pose significant threats to human life, infrastructure, transportation systems, and socio-economic development. For creeping-type landslides, progressive deformation provides an opportunity for monitoring-based early warning. However, reliable warning remains challenging because landslide displacement is nonlinear, uncertain, site-specific, and strongly affected by monitoring data quality, monitoring frequency, triggering conditions, and interpretation criteria. Conventional warning approaches often rely on fixed empirical thresholds, deterministic displacement predictions, or expert-based interpretation. These approaches may be insufficient for adaptive warning under various deformation conditions, especially when uncertainty quantification, abnormal deformation transitions, and final failure time prediction must be considered together. This doctoral thesis develops an integrated data-driven framework to enhance the reliability, adaptability, and interpretability of landslide early warning systems. The thesis consists of five interconnected research studies. First, machine-learning-based geotechnical deformation prediction is investigated using tunneling-induced ground settlement as a representative deformation problem. Four commonly used machine learning models (Support Vector Machine, Random Forest, Back-Propagation Neural Network, and Deep Neural Network) are compared using datasets with different sizes and feature characteristics. This study examines the feasibility of applying machine learning to nonlinear geotechnical deformation prediction tasks and highlights the importance of data quality, feature representation, and model generalizability, thereby providing a solid methodological foundation for subsequent landslide studies. Second, an uncertainty-aware landslide displacement prediction framework is developed using prediction intervals. Bootstrap and Lower Upper Bound Estimation models are ensembled, and a prediction interval-based cost function is introduced to improve the balance between interval coverage and interval width. The resulting framework predicts future displacement with a probable range rather than a single deterministic value, offering more informative support for risk-informed landslide early warning. Third, an adaptive unsupervised anomaly detection framework is developed for landslide early warning by combining single-centroid K-Means clustering with Extreme Value Theory. Multidimensional monitoring features are used to characterize the evolving normal deformation state, while probabilistic thresholding enables dynamic vigilance and alarm levels. Principal Component Analysis is used to improve visualization and interpretability. This framework reduces reliance on fixed empirical thresholds and supports adaptive warning as monitoring data evolve. Fourth, a hybrid landslide failure time prediction framework is developed by integrating residual-based anomaly detection with the inverse velocity method. Long Short-Term Memory models are trained to learn stable pre-acceleration deformation behavior, and residual deviations between observed and predicted displacements are used to identify the trend update point for initiating the inverse velocity analysis. An adaptive inverse velocity strategy further balances short-term responsiveness and long-term stability, improving the practical reliability of remaining-time prediction. Finally, a multivariate sequential anomaly detection framework is proposed for onset-of-acceleration identification in creeping landslides. Displacement-derived indicators are combined into two-dimensional feature spaces, and squared Mahalanobis distance is used to quantify deviations from the baseline deformation state. Dynamic upper-tail thresholding, exponential moving average smoothing, and a two-consecutive-exceedance rule are adopted to reduce false warnings caused by isolated fluctuations. The framework is demonstrated using the Ohto landslide and further evaluated using eight additional landslide cases, with monitoring frequency sensitivity examined through denser and sparser resampling scenarios. Overall, this thesis contributes a coherent set of methods for deformation prediction, uncertainty quantification, adaptive anomaly detection, adaptive failure-time forecasting, and onset-of-acceleration identification. The results indicate that carefully designed data-driven frameworks can transform landslide monitoring records into more reliable, adaptive, and interpretable warning information under the investigated case-study conditions. Collectively, the developed frameworks provide methodological support for operational landslide early warning systems, although further validation using larger independent datasets and prospective field monitoring is still required before broad real-time deployment.

Degree

thesis:*
Department dc:contributor.department
Civil Engineering
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Tang, Libin
Advisor dc:contributor.advisor
  • El-Dakhakhni, Wael

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
McMaster University
Base URL
macsphere.mcmaster.ca/server/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Tang, Libin. An Integrated Approach Toward Reliable and Adaptive Landslide Early Warning Systems. 2026. https://hdl.handle.net/11375/33362