Spatial Block-Evaluated Graph Neural Networks for Landslide Susceptibility Mapping in North Kashkadarya, Uzbekistan

Landslide susceptibility mapping in spatially clustered mountain terrain requires models that represent spatial dependence and geographically structured evaluation.This study compared five graph neural network (GNN) architectures for North Kashkadarya, Uzbekistan: graph convolutional network (GCN), Graph Sample and Aggregate (GraphSAGE), relational graph convolutional network (R-GCN), Chebyshev graph convolutional network (ChebNet), and a TransformerConv graph neural network (TransformerConv GNN) that does not implement the full Graphormer architecture.An inventory of 272 mapped landslide features was converted to 900 presence samples and paired with 900 distance-constrained absence samples.Twenty-one geological, topographic, hydrological, and environmental factors were harmonized at 20 m.Hyperparameters were selected within nested spatial block cross-validation, and performance was summarized using receiver operating characteristic-area under the curve (ROC-AUC), precision-recall-area under the curve (PR-AUC), F1, and Brier score.R-GCN produced the highest observed scores (ROC-AUC = 0.9383, PR-AUC = 0.9401, F1 = 0.8778, Brier score = 0.1033), followed by ChebNet.The five model-specific susceptibility maps consistently emphasized the dissected eastern and northeastern mountain belts, although class boundaries varied among architectures.Integrated-gradient attribution ranked northness, normalized difference vegetation index (NDVI), downslope curvature, geomorphons, and terrain-surface texture as the leading model controls.The results support GNN-based regional susceptibility screening, while the transductive graph, unbuffered block design, static inventory, and 1 km prediction-node spacing limit claims of external generalization or site-scale prediction.