Landslide Deformation Uncertainty Quantification Using Conformalized Graph Neural Networks: A Case Study in Sichuan Province, China

Landslide deformation prediction plays a crucial role in geohazard risk management, yet existing methods often struggle to quantify prediction uncertainties effectively. This paper presents GNN-CF, a novel framework that integrates Graph Neural Networks (GNN) with Conformal Prediction for reliable interval forecasting of landslide deformation. The proposed approach addresses two significant challenges: ensuring valid marginal coverage despite dependencies between calibration (training) and test data, and reducing prediction interval inefficiency while maintaining coverage guarantees. The framework incorporates a topology-aware correction model and employs graph-based regularization terms to enhance prediction accuracy. The methodology is validated using field monitoring data from the Wutong Village landslide in Sichuan Province, China, where six GNSS monitoring points track surface deformation patterns. Experimental results demonstrate that GNN-CF consistently outperforms state-of-the-art (SOTA) methods, including GRU-LUBE, DNLSSM, and SA-DeepAR, achieving superior Prediction Interval Coverage Probability (PICP) while maintaining smaller Prediction Interval Normalized Average Width (PINAW). The framework’s effectiveness in balancing coverage and computational complexity makes it a feasible solution for landslide monitoring applications requiring reliable uncertainty quantification.