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

TL;DR

This paper introduces CRAFormer, a new model that improves landslide displacement forecasting by using causal graphs and role-aware Transformers, especially in regions with frequent landslides.

Contribution

The novel contribution is CRAFormer, a causal role-aware Transformer that uses dynamic-lag Bayesian networks to improve landslide displacement prediction across different sites.

Findings

CRAFormer reduces MAE and RMSE by 59–79% across stations compared to the strongest baseline.

The model performs well in capturing displacement patterns during extreme rainfall events.

Ablation studies confirm the effectiveness of causal masks, leakage-free ICS tail, and monotonicity prior.

Abstract

Increasingly frequent intense rainfall is increasing landslide occurrence and risk. In southern China in particular, steep slopes and thin residual soils produce frequent landslide events with pronounced spatial heterogeneity. Therefore, displacement prediction methods that function across sites and deformation regimes in similar settings are essential for early warning. Most existing approaches adopt a multistage pipeline that decomposes, predicts, and recombines, often leading to complex architectures with weak cross-domain transfer and limited adaptability. To address these limitations, we present CRAFormer, a causal role-aware Transformer guided by a dynamic-lag Bayesian network-style causal graph learned from historical observations. In our system, the discovered directed acyclic graph (DAG) partitions drivers into five causal roles and induces role-specific, non-anticipative masks for lightweight branch encoders, while a context-aware Top-2 gate sparsely fuses the branch outputs, yielding sample-wise attributions. To safely exploit exogenous rainfall forecasts, next-day rainfall is entered exclusively through an ICS tail with a leakage-free block mask, a non-negative readout, and a rainfall monotonicity regularizer. In this study, we curate two long-term GNSS datasets from Guangxi (LaMenTun and BaYiTun) that capture slow creep and step-like motions during extreme rainfall. Under identical inputs and a unified protocol, CRAFormer reduces the MAE and RMSE by 59–79% across stations relative to the strongest baseline, and it lowers magnitude errors near turning points and step events, demonstrating robust performance for two contrasting landslides within a shared regional setting. Ablations confirm the contributions of the DBN-style causal masks, the leakage-free ICS tail, and the monotonicity prior. These results highlight a practical path from causal discovery to forecast-compatible neural predictors for rainfall-induced landslides.

Figures