GT-IBTR-3D: Graph Transformer for 3D Ice-Bed Topography Reconstruction

Accurately mapping subglacial bed topography is critical for understanding ice dynamics and climate impacts. Ice-penetrating radar provides direct returns from the bed, yet reliable bed picking remains difficult due to low signal-to-interference-and-noise ratio (SINR), strong spatial variability in subglacial relief, and acquisition artifacts. We propose GT-IBTR-3D, graph transformer for 3D ice-bed topography reconstruction. GT-IBTR-3D is a geometric deep learning approach that represents surface observations as graphs and reconstructs 3D ice-bed topography with a graph transformer. It combines the GraphSAGE inductive framework for localized structure with transformer layers for long-range dependencies, and it reformulates the prediction target from absolute bed elevation to surface-to-bed thickness. The bed is then recovered by subtracting the predicted thickness from the observed surface. This design isolates informative surface variation while avoiding outliers in full radargrams, stabilizes training by removing scene-level offsets, and unifies local and global reasoning. On Canadian Arctic Archipelago data, GT-IBTR-3D achieves substantially lower mean absolute error than prior probabilistic graphical methods and traditional deep neural networks.