Deep transfer learning with Bayesian optimization for evolutionary-stage prediction of step-like landslides

The landslide displacement in the Three Gorges Reservoir Area (TGRA) follows a step-like pattern, making the evolutionary stage difficult to predict. An optimized transfer learning model integrating a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) is proposed for predicting the evolutionary stage of displacement. The Bayesian algorithm is used to optimize hyperparameters of the models. The CNN-BiLSTM-Bayesian model first trains a deep learning model based on the source domain (the Baishuihe landslide). Then, transfer learning techniques and parameter fine-tuning are applied to transfer knowledge from the Baishuihe landslide to the target domain (the Bazimen landslide). The results show that the CNN-BiLSTM-Bayesian model is better than other models, such as BiLSTM and gated recurrent unit (GRU). Compared with BiLSTM, the F1-score and area under the curve (AUC) of the proposed model improved by 4.94% and 4.88% for the Baishuihe landslide, respectively. The CNN layer can extract features of data, and the BiLSTM layer can capture temporal information within displacement data. The proposed model not only acquires knowledge from similar landslide cases but also has excellent accuracy despite limited new data. Therefore, the optimized transfer learning model can accurately predict the evolutionary stage and provide a reference for landslide assessment.