Bayesian-Inspired Dynamic-Lag Causal Graphs and Role-Aware Transformers for Landslide Displacement Forecasting

Area of Science:

• Geosciences
• Artificial Intelligence
• Earthquake Engineering

Background:

• Intense rainfall is increasing landslide frequency and risk, particularly in regions with steep slopes and thin soils like southern China.
• Existing landslide displacement prediction methods often struggle with cross-domain transfer and adaptability due to complex, multi-stage pipelines.
• Accurate prediction across diverse deformation regimes is crucial for effective early warning systems.

Purpose of the Study:

• To develop a novel, adaptable, and accurate landslide displacement prediction method for early warning systems.
• To address the limitations of existing approaches in cross-domain transfer and adaptability.
• To leverage causal discovery for improved prediction of rainfall-induced landslides.

Main Methods:

• Developed CRAFormer, a causal role-aware Transformer model guided by a dynamic-lag Bayesian network-style causal graph.
• Utilized a discovered directed acyclic graph (DAG) to partition drivers into causal roles and create role-specific, non-anticipative masks for encoders.
• Implemented a context-aware Top-2 gate for sparse fusion of branch outputs and incorporated a leakage-free ICS tail for exogenous rainfall forecasts.

Main Results:

• CRAFormer reduced Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by 59-79% across stations compared to the strongest baseline.
• The model demonstrated improved accuracy near turning points and step events, characteristic of landslides.
• Performance was robust across two contrasting landslide sites (LaMenTun and BaYiTun) in Guangxi using curated GNSS datasets.

Conclusions:

• Causal discovery combined with neural prediction offers a practical approach for rainfall-induced landslide forecasting.
• CRAFormer shows significant potential for enhancing early warning systems through accurate and adaptable displacement prediction.
• The study validates the contributions of causal masks, leakage-free ICS tails, and monotonicity priors in improving prediction accuracy.