A dataset-enhanced Linformer model for geo-hazards susceptibility assessment: a case study of the railway in Southwest China

The assessment of railway safety hazards is a long-term and difficult task, and developing efficient assessment models has made significant contributions to preventing and controlling these hazards. This paper aims to demonstrate that the Linformer model based on self-attention performs better than other typical qualitative, machine learning, and deep learning models in hazard susceptibility assessment. The models are used to evaluate the susceptibility of the southwest railway. And 13 conditional factors that affect geological hazards are selected based on the geological features, geomorphometric features, hydrological features, and environmental features of the southwest railway hazard research area. A data set of 80% of the 1474 hazard and non-hazard points in the study area is randomly selected to train the susceptibility model, and the remaining 20% is used to verify the newly established model. The LR, SVM, LSTM, and Linformer models are tested using the accuracy, precision, recall rate, F1 score calculation, and AUC assessment indicators to confirm different variability. Based on the Jenks method, the susceptibility evaluation results of the southwest railway in the study area are divided into five groups from high to low. The geo-hazard susceptibility map shows that while the central and northeastern railways are less impacted by hazards, the susceptibility is higher in the southwest of the study area.