Accurate 3D reconstruction of ice-bed topography via graph transformer

Zesheng Liu, Maryam Rahnemoonfar

Abstract: Accurately mapping the bed topography of ice sheets is essential for understanding ice dynamics and monitoring the impact of climate change. Ice penetrating radar has become a powerful tool, as it emits signals that can penetrate thick ice and directly reach to the bed, providing a direct measurement of the bed topography. However, accurately identifying the ice bed boundary remains challenging due to low signal to interference and noise ratio, high variability of subglacial topography, and presence of artifacts in the collected data. To address these limitations, we proposed a novel geometric deep learning approach. Unlike previous deep learning methods that focus on convolutional operations, our proposed approach leverages the irregular structure of the bed by representing it as graphs. We developed a graph transformer that learns from the ice-surface structural information, aiming to better reconstruct the 3D bed topography. Experiment results show that compared with previous method that focus on reconstruct 3D bed topography from radar echograms, our graph transformer utilized the ice-surface elevation that contains less outliers and thereby achieved a lower mean absolute error.