A multi-tiered AI-enhanced approach for operational landslide prediction and stability assessment
Type: dissertation
Abstract: Hong Kong, with 60% of its terrain on natural slopes, faces elevated landslide risks from extreme precipitation and typhoons, making it a critical case study for this hazard. Within this context, contemporary landslide early warning systems are constrained by a critical efficacy-efficiency trade-off. Simplistic regional models (e.g., Intensity-Duration thresholds) generate high false-alarm rates by neglecting spatial variability and rainfall randomness, while sophisticated physical models are computationally challenging for real-time use. Persistent gaps include a focus on surface displacement over internal hydro-mechanical processes, vulnerability to data loss during extreme weather, and a lack of automation in stability-analysis workflows. To address these challenges, the primary objective of this research is to develop an operational, three-tiered framework that bridges the gap between data-driven and deterministic landslide models through: (1) a foundational critical review that synthesizes fragmented knowledge on monitoring technologies, prediction models and failure mechanisms; (2) a data-driven regional screening model to capture spatiotemporal trigger randomness; (3) a deep learning-enhanced subsystem for forecasting critical subsurface precursors; and (4) an automated stability tool that resolves the parameterization bottleneck and delivers end-to-end Factor of Safety analysis. The pursuit of these objectives is operationalized through a tripartite framework founded on a comprehensive review of monitoring and prediction paradigms: a) Tier 1 establishes a multi-data-driven classifier that moves beyond binary decisions to probabilistically categorize slopes as low-risk, high-risk, or ambiguous “warning” cases requiring deeper analysis. The model was developed from a uniquely curated 1984–2022 regional dataset of Hong Kong natural hillside incidents, combining antecedent rainfall sequences and spatial features. Methodologically, it addresses real-world spatial–temporal imbalance with tailored preprocessing and applies a dual-threshold strategy for nuanced risk gradation; b) Ambiguous cases identified by the probability model activate the deep learning-enhanced subsurface tier 2. Two methodological innovations are central: (1) multi-parallel LSTM networks with a novel independence-oriented time-series windowing for operationality; and (2) a multi-node prediction framework technique designed to handle data loss and forecast critical precursors (volumetric water content, matric suction, earth pressure). This tier is operationalized as OWHK (Operational VWC Forecasting Model for Shallow Layers in Hong Kong), which fuses LSTM-based forecasting; and c) Tier 3 introduces the Automated Parameterization Tool (APHs), which uses multi-LSTM networks to derive advanced soil parameters directly from standard lab data. Building on APHs, a hybrid multi-deep learning pathway predicts a regional Soil Saturation Index (SSI) from real antecedent rainfall sequences and terrain factors and computes the infinite-slope Factor of Safety, directly linking hydrologic response to geotechnical mechanics. The efficacy of this framework is demonstrated through rigorous testing and validation. The Tier 1 classifier achieved an Area Under the ROC Curve of 0.92, successfully categorizing slopes while reducing false alarms. The Tier 2 OWHK model forecasted VWC with high accuracy (MAE < 0.6%, R² > 0.92) across 15 sensors at three instrumented sites. Complementary LSTMs predicted earth pressure and matric suction with RMSEs of 1.62 kPa and 3.18 kPa, respectively. An extended VWC framework provides both point and interval predictions (MAE 0.35–1.22%) and propagates VWC into FOS calculations (R² 0.89). The APHs workflow achieved a parameterization accuracy of 99.4% through seven novel multi-parallel deep LSTM networks. This accuracy was validated against a conventional reference case (99.6%) and an experimental laboratory study, both showing strong agreement between simulated and measured results. The tool significantly accelerates parameterization, delivering precise results in seconds. The Tier 3 hybrid SSI-FOS pathway was externally validated against the deterministic GeoStudio model for the 2005 Fei Ngo Shan landslide, showing excellent agreement (FOS MAE 0.06–0.07). It also correctly assessed the stability conditions of the independent 2023 Yiu Hing Road landslide. Collectively, these results lead to several key conclusions: (a) A tiered structure integrating data-driven AI with physics-based modeling is essential for achieving both regional scalability and geotechnical accuracy in operational early warning systems; (b) The 1D-CNN architecture demonstrates superior performance for prediction, offering optimal balance between precision and computational efficiency compared to LSTM and GRU alternatives; (c) For subsurface forecasting, a multi-parallel LSTM structure with specialized networks targeting distinct hydrological regimes proves more effective than single-model approaches; (d) Volumetric Water Content emerges as the most critical and physically consistent predictive feature addressing rainfall spatiotemporal uncertainty; (e) The APHs tool represents a breakthrough in automation with accuracy surpassing traditional empirical methods; (f) A hybrid pathway fusing data-driven Soil Saturation Index with physics-based Factor of Safety calculations enables successful regional-scale, time-aware stability assessment; (g) Hillslope geometry constitutes a first-order control mechanism, with convergent slopes showing particular vulnerability to rapid moisture buildup and antecedent-condition-driven failure timing; (h) The "effective non-rainfall duration" provides a crucial metric for quantifying slope recovery time. In summary, the significance of this work is threefold, directly addressing the initial research gaps: a) Theoretical: It establishes a direct, learned link between rainfall triggers, subsurface hydrology, and geotechnical stability mechanics; b) Methodological: It pioneers novel techniques including a dual-threshold risk grader, independence-oriented windowing for operational forecasting, multi-parallel LSTM architectures, and a hybrid SSI-to-FOS coupling; c) Practical: It delivers a suite of operational tools: a multi-model, multi-event landslide forecaster, OWHK, a subsurface response predictor, APHs, and an SSI-FOS automated stability tool.
Author: Ebrahim, Kyrillos Magdy Philipe
Language: English
Subfield: Management, Monitoring, Policy and Law
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
Funder: Hong Kong Polytechnic University
OpenAlex ID: W7216819657