A hybrid deep learning approach for highway landslide susceptibility assessment based on InSAR data

Landslides, as a common geological hazard, are characterized by their sudden onset, widespread distribution, and severe consequences, posing a serious threat to the safe operation of mountainous roads. Existing landslide susceptibility studies primarily rely on historical landslide points, with limited consideration of potential landslide areas. Additionally, deep learning models face challenges in handling long-sequence data related to landslide conditions, such as insufficient contextual modelling capabilities. For this purpose, this paper takes the China-Pakistan Highway as the study area, constructs a landslide dataset integrating Line of Sight (LOS) vertical deformation rates based on Interferometric Synthetic Aperture Radar (InSAR) data, extracts 21 multi-source conditioning factors, and builds a hybrid model combining a convolutional neural network and a bidirectional recurrent neural network (CNN-BiRNN) for landslide susceptibility mapping. The CNN-BiGRU model demonstrated superior performance compared to other models (Area Under the Curve (AUC) = 97.2%, Accuracy (ACC) = 90.7%, F1-score (F1) = 90.8%), indicating its strong potential in landslide susceptibility identification. This provides a reference basis for subsequent studies on the precise identification and prevention of landslide disaster risk zones.