Weathered regolith is a key predisposing factor that controls landslide initiation and potential failure volume, particularly in high-relief gorge regions such as southeastern Tibet. There, it may govern the occurrence of large landslides and the reactivation of ancient landslides. This study develops a geophysics-centered, multi-source data fusion framework with graph neural network perception to map regional highly weathered regolith thickness. The mapping covers a 401 km² area with widespread ancient landslides in Basu County, Changdu, southeastern Tibet. First, transient electromagnetic data along 158 survey lines and at 602 measurement stations were acquired. This derived 286 highly weathered regolith thickness control points through standardized geophysical processing and interpretation. To address the strong uncertainty, sparsity, and spatial heterogeneity of these observations, we formulated thickness mapping as a semi-supervised regression problem. A 43-variable predictor set spanning topography, geology, hydrology, seismicity, climate, and surface observations was constructed, and a slope-unit-constrained semi-supervised graph neural network was trained for regional prediction. The final model achieved an R² of 0.825 in 10-fold cross-validation. These results provide a quantitative basis for assessing the occurrence, reactivation, and physics-based hazard potential of regional giant landslides. More broadly, the proposed multi-source fusion and deep learning framework offers a transferable paradigm for large-scale mapping of landslide-prone environmental conditions.

Highly Weathered Regolith, spatial modelling, ancient-landslide, graph neural network