A scalable graph transformer for large-scale landslide susceptibility mapping: a case study in the Upper Yellow River Basin

Landslide susceptibility mapping (LSM) is crucial for disaster prevention and regional planning; however, existing approaches face challenges in large-scale and heterogeneous environments owing to high computational costs and limited capacity for modeling long-range dependencies. To overcome these limitations, we proposed a scalable Graph Transformer with a dual-branch design. It integrates a linear attention-based Transformer, which reduces the computational complexity for global dependency modeling, and a Scalable Inception Graph Neural Network (SIGN), which emphasizes local feature extraction and multiscale feature integration. Additionally, consistency regularization loss was incorporated to enhance the robustness and generalizability. Experiments in the Upper Yellow River Basin, which is characterized by complex and diverse geomorphology, show that the proposed model achieves 81.7% accuracy and 88.8% recall, outperforming traditional machine learning methods and mainstream Graph Neural Networks (GNNs) with improvements of 5% and 3.7%, respectively. Furthermore, we introduced the Mean Average Distance (MAD) and the gap of MAD values (MADGap) to quantify the over-smoothing problem in a GNN-based LSM. The proposed model exhibited consistently higher MADGap values across multiple subgraphs than mainstream GNNs, indicating its superior capacity in mitigating over-smoothing and confirming its overall accuracy and effectiveness for large-scale LSM.