GRIT-LP: Graph Transformer with Long-Range Skip Connection and Partitioned Spatial Graphs for Accurate Ice Layer Thickness Prediction

TL;DR

This work tackles the challenge of accurately predicting ice-layer thickness from polar radar imagery by enhancing graph transformers to model both local spatial coherence and long-range temporal dependencies. The proposed GRIT-LP framework combines GraphSAGE-based spatial learning, multi-head temporal attention, and an adaptive long-range skip connection, together with partitioned spatial graphs to maintain locality and reduce noise. The approach yields a 24.92% RMSE improvement over the prior state-of-the-art ST-GRIT on a representative l=5, m=15 setting, while remaining robust to variations across radargrams. These advances enable deeper temporal reasoning in cryospheric data, with potential benefits for understanding snow accumulation, ice-sheet dynamics, and sea-level rise projections.

Abstract

Graph transformers have demonstrated remarkable capability on complex spatio-temporal tasks, yet their depth is often limited by oversmoothing and weak long-range dependency modeling. To address these challenges, we introduce GRIT-LP, a graph transformer explicitly designed for polar ice-layer thickness estimation from polar radar imagery. Accurately estimating ice layer thickness is critical for understanding snow accumulation, reconstructing past climate patterns and reducing uncertainties in projections of future ice sheet evolution and sea level rise. GRIT-LP combines an inductive geometric graph learning framework with self-attention mechanism, and introduces two major innovations that jointly address challenges in modeling the spatio-temporal patterns of ice layers: a partitioned spatial graph construction strategy that forms overlapping, fully connected local neighborhoods to preserve spatial coherence and suppress noise from irrelevant long-range links, and a long-range skip connection mechanism within the transformer that improves information flow and mitigates oversmoothing in deeper attention layers. We conducted extensive experiments, demonstrating that GRIT-LP outperforms current state-of-the-art methods with a 24.92\% improvement in root mean squared error. These results highlight the effectiveness of graph transformers in modeling spatiotemporal patterns by capturing both localized structural features and long-range dependencies across internal ice layers, and demonstrate their potential to advance data-driven understanding of cryospheric processes.

Figures (4)

• Figure 1: Network architecture of the proposed graph transformer network, GRIT-LP. (a) Overview of the complete network architecture. (b) Architecture of GRIT-LP (c) Details of the temporal attention block.
• Figure 2: (a) Airborne radar sensor captures the status of internal ice layers by measuring the reflected signal.(Image adapted from diagram-airborne-radar) (b) Radargram image (c) Labeled image, where boundaries of each ice layer is manually labeled out.
• Figure 3: Qualitative visualization of some prediction results by GRIT-LP with $N=8$ and $\alpha=0.25$. The blue line is used to generate the graphs. The green line is the groundtruth (manually-labeled ice layers) and the red line is the model prediction.
• Figure 4: Qualitative results that shows a comparison of different model predictions on the same radargram. The blue line is used to generate the graphs. The green line is the groundtruth (manually-labeled ice layers) and the red line is the model prediction. "Best" means GRIT-LP with 8 attention blocks and $\alpha=0.25$.