An interpretable attention-based deep learning method for landslide prediction based on multi-temporal InSAR time series: A case study of Xinpu landslide in the TGRA

Abstract

The prediction of landslide deformation is crucial for early warning systems. While conventional geotechnical in-situ monitoring is restricted due to its high cost and spatial limitations over large regions, deep learning-based methodologies with remote sensing data have become increasingly prevalent in contemporary predictive research, yet this frequently engenders the enigmatic “black box” issue. To address this, we improve the landslide displacement prediction framework by combining interpretable deep learning based on an attention mechanism and Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) techniques. MT-InSAR is first used to extract a landslide displacement time series from Copernicus Sentinel-1 SAR images. Then Variational Mode Decomposition (VMD) is employed to separate the nonlinear displacement time series into trend, seasonal, and noise components. The Auto-Regressive Integrated Moving Average (ARIMA) model and Bidirectional Gated Recurrent Unit (BiGRU) are applied to predict trend and seasonal displacements, respectively. The inputs for these predictions are determined by analyzing landslide influencing factors. This study uses the Xinpu landslide in the Three Gorges Reservoir Area of China to evaluate the proposed method and compare its performance with existing models. The CNN-Attention-BiGRU algorithm effectively captures the nonlinear relationship between landslide deformation and its triggering factors, outperforming conventional deep learning models such as BiLSTM, BiGRU, and CNN-BiGRU, achieving improvements in Root Mean Square Errors (RMSEs) by 21%—55% and Mean Absolute Errors (MAEs) by 23%—56%. By applying deep learning with an attention mechanism, our proposed method considers the underlying principles of landslide deformation, and factors with higher relative importance for prediction modeling are interpreted to be concentrated annually between April and August, enabling a more effective and more accurate prediction of large-scale landslide kinematics for the studied reservoir region.