A Graph-Based Inference Framework for Active Landslide Identification by Integrating InSAR Data and Landslide Prior Knowledge
Year: 2026
Type: article
Abstract: Interferometric synthetic aperture radar (InSAR) enables large-area observation of ground surface deformation and has become an important technique for identifying active landslides. However, deformation detected by InSAR is not exclusively caused by landslide activity, engineering activities, agricultural disturbance, land subsidence, and other nonlandslide processes may also produce similar deformation signals. As a result, existing deep learning methods that rely only on InSAR-derived features are prone to false identification in complex environments. To address these issues, this study proposes a graph-based inference framework for active landslide identification by integrating InSAR with landslide prior knowledge to distinguish landslide-related deformation candidates from other deformation signals. First, an improved Yolov11n-based detection model is employed to extract candidate deformation areas from InSAR data. Subsequently, a topological graph is constructed using joint constraints of spatial proximity and environmental similarity, thereby explicitly encoding the spatial-environmental relationships among different deformation candidates. In this process, each detected deformation area is treated as a node, and ten categories of landslide influencing factors related to terrain, geology, hydrology, and human activities are assigned as node attributes. Finally, a graph neural network-transformer model is employed to perform node-level inference by integrating local graph connectivity with multihop contextual information, thereby distinguishing landslide-related deformation from nonlandslide deformation. Experiments conducted in the Bailong River Basin suggest that the proposed framework improves the discrimination of landslide-related deformation candidates. The results indicate that incorporating landslide prior knowledge into graph inference can reduce the risk of false identification caused by nonlandslide deformation and provide a feasible technical pathway for constructing regional-scale active landslide candidate inventories.
Institutions: Lanzhou Jiaotong University, Lanzhou University of Technology
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7204549009
