Investigating the landslide susceptibility assessment methods for multi-scale slope units based on SDGSAT-1 and Graph Neural Networks

Abstract

Accurate assessment of landslide susceptibility is vital for risk prevention, yet existing methods often overlook remote but interconnected geographical features, leading to unreliable maps. To effectively address this issue, the complex mountainous terrain and geomorphological features involved in landslide formation were fully considered in this work. This was attained by introducing geographical environmental correlations from the perspectives of mapping units and susceptibility assessment models to achieve comprehensive linkage between the landslides and the affected environments, thereby enhancing accuracy. Significantly, in this work, the SDGSAT-1 data were innovatively applied to the field of landslide research and the landslide susceptibility in Jiulong County, Ganzi, was evaluated based on optimal-scale slope units and Graph Neural Networks (GNN). The results showed that: (1) SDGSAT-1 offers significant benefits to landslide research. Our analysis compared LANDSAT with similar resolutions from multiple perspectives and found that SDGSAT-1 has substantial advantages. (2) The landslide susceptibility assessment method proposed in this work, based on optimal-scale slope units and GNN, demonstrated superior performance, with various evaluation metrics, such as AUC, Accuracy, and Precision far exceeding those of other machine learning models.