Landslide Susceptibility Assessment Using a CNN–BiLSTM-AM Model

Landslides are common geological hazards worldwide, posing significant threats to both the environment and human lives. The preparation of a landslides susceptibility map is a major method to address the challenge related to sustainability. The study area, Nyingchi, is located in the southeastern region of the Qinghai-Tibet plateau, characterized by diverse terrain and complex geological formations. In this study, CNN was used to extract high-order features from the influencing factors, while BiLSTM was utilized to mine the historical data. Additionally, the attention mechanism was added to adjust the model weights dynamically. We constructed a hybrid CNN–BiLSTM-AM model to assess landslide susceptibility. A spatial database of 949 landslides was established using remote sensing images and field surveys. The effects of various feature selection methods were analyzed, and model performance was compared to that of six advanced models. The results show that the proposed model achieved a high prediction accuracy of 90.12% and exhibits strong generalization capabilities over large areas. It should be noted, however, that the influence of feature selection methods on model performance remains uncertain under complex conditions and is affected by multiple mechanisms.