Landslide susceptibility analysis is a process that evaluates the spatial likelihood of future landslides based on the correlation between historical landslide records and conditioning factors. In data-driven susceptibility analysis, the selection of ...
Landslide susceptibility analysis is a process that evaluates the spatial likelihood of future landslides based on the correlation between historical landslide records and conditioning factors. In data-driven susceptibility analysis, the selection of appropriate mapping unit is critical. Slope units have recently been widely adopted for reflecting actual terrain boundaries. Furthermore, a graph-based machine learning approach has been employed to analyze landslide susceptibility on the basis of slope unit mapping units. However, previous graph studies have either focused on capturing the adjacency relationships between slope units by reducing internal spatial variability to a single representative value, or on the internal variability within slope units, without reflecting inter-unit correlations. It has been demonstrated that both types of spatial information affect landslide occurrence; however, they have not yet been incorporated into model training simultaneously.
In order to address the aforementioned limitation, the present study proposes a Hierarchical Graph Attention Network (H-GAT). The H-GAT is designed to hierarchically integrate both types of spatial information by using node embeddings from a contour-based graph (Cont-graph) within each slope unit as the node input of a slope unit graph (SU-graph). The proposed approach was applied to landslide susceptibility analysis in Jecheon-si, Chungcheongbuk-do, where landslides occurred in 2020. To facilitate comparison, three models were constructed: an MLP devoid of any graph structure, an SU-GAT reflecting only inter-unit relationships, and a Cont-GAT reflecting only intra-unit information. Each model was trained five times with different random seeds and evaluated using a variety of metrics, including confusion matrix-based metrics, Monte Carlo Dropout-based predictive uncertainty, Integrated Gradients-based factor contribution, and analysis of attention weights and node embeddings.
H-GAT demonstrated the optimal predictive performance, as indicated by the area under the curve (AUC) of 0.935 and the area under the precision-recall curve (AUPRC) of 0.927. Monte Carlo Dropout analysis revealed that the uncertainty distributions of the training and test datasets were analogous, suggesting that the model had undergone stable generalization without the occurrence of overfitting. Factor contribution, attention weight, and node embedding analyses consistently demonstrated that H-GAT effectively incorporated hydro-geomorphological factors, such as planform curvature and TWI. The model also concentrated attention on slope units adjacent to actual landslides and formed node representations that most clearly distinguished landslides from non-landslide units. These findings demonstrate that integrating fine-scale intra-unit structures and inter-unit spatial correlations into a hierarchical framework enables more accurate and reliable landslide susceptibility prediction than approaches relying on a single type of spatial information.