PSFC-LSTM: A Novel Method for LSTM Prediction of Surface Deformation in High-Speed Railway Based on Clustering of PS Point Metadata Features

To meet the urgent demand for accurate deformation prediction in safe operation of high-speed railway(HSR), a novel LSTM prediction method based on clustering of PS point metadata features (PSFC-LSTM) is proposed in this study, aiming to overcome the limitation of traditional InSAR monitoring time series prediction methods that fail to make full use of PS point information, and improve the prediction accuracy of surface subsidence of HSR. The method begins by extracting key features from PS point metadata, which are subsequently clustered into three groups using the K-Means algorithm. Concurrently, deformation trends of the PS points are analyzed and classified into four distinct types: sinusoidal, linear increasing, linear stable, and linear decreasing. By integrating the clustering results with the deformation trend classifications, the study area is systematically divided into 12 sub-regions. In each sub-region, an independent LSTM model is employed to predict cumulative deformation variables. By incorporating both PS point metadata features and deformation trends, this method effectively enhances predictive accuracy. To validate the proposed method, we utilized 28 Sentinel-1 SAR acquisitions spanning from October 17, 2020, to October 24, 2021, to conduct PS-InSAR monitoring and cumulative deformation predictions for the Jincheng section of the Tai-Jiao High-Speed Railway (Tai-Jiao HSR) in China. The results demonstrate that, for cumulative deformation predictions in the final four time periods of the study timeframe, the PSFC-LSTM method reduced the average RMSE by 13.8%, lowered the average MAE by 26.0%, and improved the average R² by 4.4% compared to conventional LSTM models. These findings confirm that the PSFC-LSTM method effectively harnesses PS point information to significantly enhance the predictive accuracy of surface deformation, offering valuable theoretical insights and data support for deformation monitoring in linear infrastructure projects such as HSR.