landslide displacement prediction, intervals prediction, large pre-trained model, transfer learning, transformer, Engineering (General). Civil engineering (General), TA1-2040, Chemical engineering, TP155-156, Physics, QC1-999

Assessing the uncertainty in landslide displacement prediction is essential for improving the dependability of landslide early warning systems. In this study, we employ advanced Transformer-based models to develop robust probabilistic forecasting methods, which are used to construct reliable prediction intervals for landslide displacement. Furthermore, to overcome the challenge of limited data availability, a common issue due to the high costs associated with monitoring landslide displacement, we leverage Lag-LLama, a large pretrained model initially trained on extensive cross-domain time series data, to enhance the probabilistic prediction of landslide displacement. Our experiments, conducted on six real-world landslide displacement datasets from China, demonstrate that the Lag-LLama-based approach significantly outperforms state-of-the-art time series forecasting models in the context of landslide displacement interval prediction. These results highlight the potential of large pre-trained models in addressing data scarcity and improving predictive accuracy in geohazard monitoring applications.