Landslides are common geological disasters, characterized by their sudden occurrence and destructive impact. Modeling and predicting landsli
Landslides are common geological disasters, characterized by their sudden occurrence and destructive impact. Modeling and predicting landslide displacement behavior can help identify potential risks in advance. This study proposes a Cross-Attention Stacked Transformer landslide displacement prediction model (CA-Stacked Transformer). Through the introduction of a cross-attention mechanism, the model adaptively fuses exogenous and endogenous variables. A new prediction framework is developed based on the Transformer model, combining external environmental features with historical landslide displacement data from the past year. This framework effectively captures the temporal dynamics and potential nonlinear relationships in landslide displacement while streamlining the prediction process by eliminating the need to decompose displacement into periodic and trend components, as required by most methods. Using two monitoring points of the Baishuihe landslide in the Three Gorges Reservoir area as examples, the CA-Stacked Transformer model outperforms traditional forecasting methods across multiple evaluation metrics. For instance, at the XD01 monitoring point, the model achieves RMSE improvements of 65.2 %, 79.2 %, 37.5 %, 59.9 %, 71.1 %, and 46.2 % compared to the CA-LSTM, CA-RNN, CA-GRU, CA-CNN-LSTM, CA-TCN, and STGCN models, respectively; similarly, at the ZG118 monitoring point, RMSE improvements of 68.7 %, 74.7 %, 73.1 %, 53.3 %, 68.4 %, and 59.6 % are observed. These results demonstrate that the CA-Stacked Transformer model possesses strong generalization ability and predictive accuracy, significantly enhancing landslide forecasting performance.
