A hybrid deep learning method for landslide susceptibility analysis with the application of InSAR data

Landslides pose severe threats to human life and property, necessitating significant prevention methods. Landslide susceptibility analysis is an effective approach for landslide prevention. However, the landslide inventory databases are used when applying current approaches neither contains the potential landslide points in study area nor do they consider ground deformation features in the existing landslide predisposing factors. Therefore, InSAR data were used in this study to substitute the existing landslide databases with 144 historical landside locations in Shuicheng County which could not include potential landslide points. Meanwhile, this study proposed a hybrid deep learning method for landslide susceptibility analysis with the application of InSAR data, involving two parts: (i) The first part concerns the extraction of landslide predisposing factors: in addition to multi-source data such as topography, geomorphology, and hydrology, InSAR data were introduced, and 36 types of landslide predisposing factors with ground deformation features were extracted based on InSAR locations. Multicollinearity test and importance analysis by information gain (IG) for these factors were carried out to obtain landslide predisposing sequence factors. (ii) The second part of the approach followed consists in the construction of prediction model: the hybrid deep learning method integrates convolutional neural networks (CNN) and three recurrent neural network (RNN) variants, namely CNN-long short-term memory (CNN-LSTM), CNN-gated recurrent unit (CNN-GRU), and CNN-simple recurrent unit (CNN-SRU). First, the CNN was employed to obtain main affected landslide predisposing sequence factors to reduce the non-influencing features. The landslide prediction model was built with long short-term memory (LSTM), gated recurrent unit (GRU), and simple recurrent unit (SRU) to quantitatively predict landslide susceptibility for generation of landslide susceptibility maps. The area under the curve (AUC), accuracy (ACC), kappa coefficient (KAPPA), and the Matthews correlation coefficient (MCC) were used for model performance evaluation. Compared with the existing methods of CNN-support vector machine (CNN-SVM), CNN-random forest (CNN-RF), and CNN-logistic regression (CNN-LR), the results of the method in this study have higher performance. The CNN-GRU model has the highest precision with an AUC value of 98.4%, an ACC value of 93.7%, a KAPPA value, and a MCC value of 87.4% and 87.5%, respectively, indicating the excellent validity and feasibility of the method in this study.