DS-InSAR combined with dual-attention enhanced CNN-BiGRU for deformation prediction of typical landslides in the Jinsha River Basin

Accurate prediction of step-like abrupt landslide deformation is critical for disaster early warning but remains hindered by the over-smoothing bias of conventional models. This study proposes an end-to-end prediction framework integrating Distributed Scatterer InSAR (DS-InSAR) monitoring data with a dual-attention enhanced CNN-BiGRU network. Two architectures—direct-output and Sequence-to-Sequence (Seq2Seq)—were validated on the Guoba Landslide, a large landslide in the Jinsha River Basin. The dual attention mechanism adaptively isolates critical precursor features, enhancing both precision and interpretability. A global evaluation across all measurement points delineates their domains of applicability. The direct-output DA-CNN-BiGRU acts as a robust spatial filter, achieving a global R2 of 0.92 for ubiquitous continuous creep. Conversely, the Seq2Seq-DA-CNN-BiGRU is specifically optimised to mitigate this over-smoothing bias and capture severe step-like deformations, achieving an R2 of 0.636 at a representative critical point where baselines fail. Across representative kinematic points, it reduces average RMSE and MAE by 24.02% and 30.04%, respectively, compared to standard CNN-BiGRU. This study provides a targeted, interpretable early-warning scheme for complex landslides in high-altitude canyon regions.