Enhancing Landslide Displacement Prediction Using a Spatio‐Temporal Deep Learning Model With Interpretable Features

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Wang, Jia, Zhu, Hong‐Hu, Zhang, Wei, Tan, Dao‐Yuan, Pasuto, Alessandro
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• DOI: 10.1029/2025JH000592
• ISSN: 2993-5210

Landslides cause significant economic losses and pose severe risks to human safety, making accurate predictions of landslide displacements essential for effective early warning systems. Many prediction models focus primarily on time series forecasting at individual monitoring points. Consequently, challenges are faced in capturing the spatial correlations of landslide displacements. In addition, the black‐box characteristics of the model limit the interpretability of the decision‐making process, which may make the prediction results difficult to use effectively for decision‐making. This paper proposes a deep learning model with interpretable features, which combines graph neural networks and gated recurrent units (GRUs) to capture the spatio‐temporal characteristics of landslide displacements. Using the Outang landslide in the Three Gorges Reservoir area as a case study, the effectiveness of the proposed model is validated through comparisons with long short‐term memory and GRU models. The results demonstrate that the spatio‐temporal graph neural network model provides accurate predictions of landslide displacement, exhibiting strong robustness and stability. The attention module enhances the interpretability of the model, revealing the influence of factors such as the groundwater table, reservoir water level, and temperature at different monitoring points. This work provides insights into advancing the understanding and forecasting of complex landslide behaviors. Landslides pose a serious threat to human safety and property. Accurate prediction is crucial for early warning and risk management. However, current methods often focus on individual monitoring points and lack interpretability, making it difficult to understand the rationale behind the predictions. This study proposes a new approach that integrates temporal and spatial information from various monitoring stations to improve prediction accuracy. The model uses an attention mechanism to highlight the relative importance of different environmental drivers, such as water levels and temperature, making the predictions more transparent. By identifying key factors and explaining their impacts, the model improves the understanding of landslide behavior and fosters enhanced community protection. The spatio‐temporal graph neural network captures spatial and temporal dependencies, improving landslide displacement prediction accuracy Interpretable attention mechanisms quantify the influence of key factors, balancing prediction precision and model transparency Hydrological drivers exhibit time‐lag effects on displacement, underscoring the necessity of lag‐aware models for early warning systems