Locate and extend: a geometric deep learning strategy for predicting polar ice layer structures using graph neural networks

Zesheng Liu, Maryam Rahnemoonfar

Abstract: Spatiotemporal patterns in polar ice layers play a crucial role in understanding ice accumulation and layer formation processes. Accurately capturing these patterns is essential for monitoring snow mass balance changes, estimating future ice sheet melting rates and serving as high-fidelity inputs for other climate models. While traditional approaches leverage convolutional neural networks on raw echogram images collected by airborne snow radar sensors, the inherent noise in these images, together with the need of high-quality labels on radargram, often limits the model overall performance. To address these limitations, we focus on geometric deep learning, employing graph neural networks to extract meaningful patterns from thickness information of shallow ice layers and make prediction for deeper layer thickness. In this work, we propose a novel network that deploy a “locate and extend” learning strategy. In the ”locate” process, we make predictions for the mean thickness of the ice layer, and thereby locate the rough position of each layer. In the “extend” process. we extend the in-layer variations by predicting the shift of each node corresponding to the layer mean thickness, forming the structural details. Experiments shows that compared with other methods that directly predict the detailed layer structure in a single step, our proposed approach can have competitive performance in both accuracy and efficiency. This work also highlights the potential of decomposing deep ice layer thickness prediction into two independent yet complementary tasks, which can be further improved individually in future research.