Accurate prediction of landslide displacement is crucial for evaluating stability. However, factors such as material variability, progressive failure at soil-rock interfaces, and nonlinear hydro-mechanical coupling induced by rainfall contribute to complex deformation patterns in accumulation landslides, making reliable predictions difficult. Most existing models assume uniform landslide materials and overlook differences in hydrological responses across areas, thereby limiting their accuracy(Zhou et al, 2025). To tackle this issue, a new prediction framework has been developed that focuses on a typical soil-rock dual-structure accumulation landslide in the Yellow River Basin.
A landslide displacement time series was obtained from 168 Sentinel-1 images collected between January 2020 and August 2025. Grey relational analysis (GRA) assessed the relationships between surface displacement and hydrological factors, including precipitation over various time scales and soil water content at different depths. This led to the creation of a displacement-hydrological grey relational index. Global Moran’s I and optimized hotspot analysis were then used to explore the spatial heterogeneity of displacement-hydrological interactions, allowing for the derivation of spatial weighting factors. The displacement time series was decomposed using variational mode decomposition (VMD), with trend components predicted by a multilayer perceptron (MLP) and periodic components predicted by a CNN-BiGRU model. Finally, the spatial weighting factors were integrated into the loss function to enhance landslide displacement predictions.
The results show that: (1) The hydrological responses of different landslide parts exhibit significant spatial variability, with stronger displacement-hydrology relationships in the central and flank areas compared to the toe. (2) The grey relational index shows spatial clustering, identifying high-value hotspots in the central and flank zones and low-value cold spots in the toe. (3) The proposed framework outperformed ten benchmark models, achieving RMSE reductions of 22.74% to 58.44%. Compared to the conventional CNN-BiGRU model, RMSE and MAE were further reduced by 18.7% and 16.4%, respectively. These results highlight the importance of incorporating spatial variability into deep learning frameworks for improved landslide displacement predictions and early warning systems.
Acknowledgments
This work was supported by the Natural Science Foundation of Henan Province for Young Scientists (Grant No. 252300423262).
References
Zhou C, Ye M, Xia Z, et al. An interpretable attention-based deep learning method for landslide prediction based on multi-temporal InSAR time series: A case study of Xinpu landslide in the TGRA[J]. Remote Sensing of Environment, 2025, 318: 114580.
 

Accumulation landslide,InSAR deformation monitoring,Hydrological response heterogeneity,CNN–BiGRU