Time-series InSAR landslide three-dimensional deformation prediction method considering meteorological time-delay effects

Abstract

Landslide deformation prediction is a critical component of disaster early warning systems and significantly contributes to disaster prevention and mitigation. However, the high cost of traditional deformation monitoring equipment limits its extensive application across large areas. Furthermore, existing landslide deformation prediction methods often overlook the nonlinear influence of meteorological conditions. This study introduces a novel framework for predicting three-dimensional landslide deformations by employing time-series interferometric synthetic aperture radar (InSAR), which accounts for the time-delay effects of meteorological factors. First, the framework leverages ascending and descending orbit time-series InSAR technology to generate three-dimensional deformation data for landslides. Subsequently, the deformation data were decomposed into trend and periodic components using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method. An autoregressive integrated moving average (ARIMA) model was developed for trend prediction, and an improved unscented Kalman filtering (UKF) model was developed for the periodic component. The prediction of the periodic component uses a second-order Taylor series and adaptive adjustments of observation and process noise covariance matrices to create a UKF deformation prediction model that incorporates meteorological time-delay effects. The method's effectiveness was validated through experiments on the Xiongba and Sela landslides in Gongjue County, Tibet. Results demonstrated that ascending and descending orbit time-series InSAR technology can accurately detect three-dimensional deformation, revealing maximum deformation rates of 43.81 cm/year for the Xiongba landslide and 31.83 cm/year for the Sela landslide. For the trend component, the ARIMA model achieved a correlation coefficient (R²) exceeding 0.9 and root mean square error (RMSE) below 1.0 cm across multiple monitoring points. Regarding the periodic component, the framework's reliability was first confirmed through simulation experiments. Further analyses of the Xiongba and Sela landslides revealed that rainfall and temperature exhibit distinct time-delay effects on deformation. Incorporating meteorological data into the improved UKF model significantly enhanced the prediction accuracy, yielding R² values between 0.8 and 0.9, with the RMSE and mean square error (MAE) outperforming those of the long short-term memory (LSTM) and recurrent neural network (RNN) comparison models. Overall, the proposed framework offers vital technical support for risk prediction and early warning of large-scale landslide disasters, facilitating more accurate landslide forecasting.