Physics-Informed Spatio-Temporal Graph Neural Network for Efficient Deep Ice Layer Thickness Estimation in Radar Imagery

A better understanding of the thickness and variability of internal ice layers can provide crucial insights into monitoring snow mass balance change, assessing ice dynamics, and reducing uncertainties in climate models. Currently, the most effective way to capture the status of the internal ice layer is through radar sensors, as they can penetrate through the thick snow accumulation. While convolutional networks are widely used to extract layer boundaries from radargram images, noise in the radargrams prevents researchers from achieving high-quality results. Moreover, deeper ice layers are typically less contrastive and tend to be even harder to track due to the low signal-to-noise ratio in radar imagery. Instead of convolutional neural network, we focus on geometric learning using spatio-temporal graph neural network in this work, aiming to learn a mapping between the thickness information of shallow and deeper ice layers. We developed a novel physics-informed multi-branch spatio-temporal graph neural network with an inductive graph framework and a temporal convolution in different branches and added physical node features synchronized from the Model Atmospheric Region weather model. We found that with the inclusion of snow mass balance, meltwater refreezing, height change due to melting, and snowpack height as physical features, our proposed network can consistently outperform other non-physical models in both accuracy and efficiency.