Landslide Prediction Research Based on the Integration of InSAR Technology and the CEEMD-TCN-BiGRU-Attention Model

To address the limitations of traditional landslide monitoring methods and the insufficient capability of existing prediction models in capturing nonlinear temporal features, this study proposes a landslide prediction method that integrates InSAR technology with a CEEMD-TCN-BiGRU-Attention model. Taking the Zhamunongba landslide in southeastern Tibet as the research object, a large-scale deformation field was first obtained using SBAS-InSAR technology, from which time-series deformation data of monitoring points were extracted. The original data were then decomposed into trend and periodic deformation components using Complementary Ensemble Empirical Mode Decomposition (CEEMD). Subsequently, the Temporal Convolutional Network (TCN) with causal convolution and residual connections was employed to enhance feature extraction, followed by a Bidirectional Gated Recurrent Unit (BiGRU) to capture complex spatiotemporal dependencies and an attention mechanism to dynamically weight key features, enabling precise prediction of deformation components with distinct characteristics. Notably, for the periodic deformation component, a multi-factor driven prediction mechanism was constructed by integrating cumulative rainfall and average soil moisture at multiple time scales, achieving high-precision periodic deformation forecasting. Ablation experiments demonstrate that the CEEMD-TCN-BiGRU-Attention model excels at capturing temporal correlations and nonlinear features of landslide deformation data, achieving an MAE of 3.58 mm, RMSE of 4.16 mm, and R² of 96.26%. This model outperforms the TCN-BiGRU-Attention, TCN-Attention and single TCN models in prediction accuracy. The proposed landslide prediction method realizes high‑precision forecasting and holds important value for disaster early warning.