A kinematics-informed dual-stream deep learning framework for multi-timescale landslide displacement forecasting and risk-informed decision support

Landslide hazard mitigation requires forecasts that support both emergency response and longer-term engineering planning. Existing data-driven models often target a single horizon and rely mainly on displacement sequences, limiting their representation of deformation stages and long-horizon stability. We propose a kinematics-informed dual-stream framework for multi-timescale displacement forecasting and risk-informed decision support. A TCN-BiLSTM temporal stream is fused through co-attention with an XGBoost stream encoding displacement-derived velocity, acceleration, and improved tangential angle. The latter provides interpretable deformation-state information without imposing governing mechanical constraints. To prevent information leakage, EEMD is applied within causal rolling windows; 30-day and 90-day forecasts use scenario-based rainfall rather than observed future rainfall. Predicted trajectories are converted into three engineering-oriented risk indicators and integrated into a digital-twin-assisted decision-support framework. Evaluation on the Zaoshuwa landslide used G2-G5 for model development and G1 as an unseen spatial target. Across 1-day, 7-day, 30-day, and 90-day horizons, the framework improved prediction accuracy and trend stability relative to benchmark models. These results show that combining temporal and kinematic representations supports multi-timescale forecasting and risk assessment. The framework is intended for decision support; warning thresholds and spatial transferability require site-specific calibration and further validation.