ConToGCN: A landslide susceptibility assessment model considering contour topographic features in slope units using graph convolution network

Abstract

Landslide susceptibility assessment (LSA) plays a critical role in disaster prevention and mitigation. While machine learning techniques have been widely applied to LSA with notable progress, they face limitations in LSA precision, and struggle to capture micro-topographic features with multi-points instead of single one in slope units. To address these limitations, this study proposes an LSA model based on contour topographic features and Graph Convolution Networks (ConToGCN). First, the graph structure for each slope unit is built by generating nodes and edges from contour lines. Next, 9 factors, including elevation, are extracted for graph nodes using feature engineering, and the steepness between nodes is calculated to identify critical features such as steep scarps and flow zones. The ConToGCN model is then developed to aggregate adjacent node features and generate landslide probability. Changxing County in Zhejiang, China, was selected as the study area. A dataset comprising 124 landslide and 124 non-landslide samples was collected, with 70% used for training and 30% for testing. A comparative performance evaluation was conducted against Graphormer, a GCN model based on slope unit centroids (CentGCN), Random Forest (RF), and Support Vector Machine (SVM). The results demonstrate ConToGCN model outperforms these other models, with an area under the ROC (AUROC) of 0.93 and an area under the PR curve (AUPR) of 0.94. This represents improvements of 14.81% and 14.63% over the SVM model, respectively. The ConToGCN model effectively captured the complex topographic structures and improved the LSA precision significantly. We believe the model provides a novel approach to landslide disaster prevention and mitigation.