A snow properties-aware deep learning framework for penetration bias estimation of TanDEM-X DEMs over ice sheets

Abstract

The accurate assessment of glacier volume and mass changes as well as snow depth is crucial for understanding glaciological processes and the impact of climate change. TanDEM-X, an X-band spaceborne interferometric synthetic aperture radar (InSAR) mission, offers global, high-resolution digital elevation models (DEMs) that are invaluable for these studies. However, the inherent variability in radar wave penetration into snow and ice creates challenges in accurately estimating surface elevation changes and snow depth. Variations in the acquisition geometry and snow properties can affect the estimation of the radar mean phase center, leading to penetration bias and an underestimation of the surface topographic height. In this work, we propose a novel deep learning framework for estimating the penetration bias of TanDEM-X DEMs over ice sheets, by combining the knowledge of the physical properties of snow and the InSAR system for the development of a robust regression framework. Due to the lack of extended reference data, which jeopardizes the use of fully-supervised data-driven approaches, we propose a deep learning approach based on two intrinsically connected tasks: a first unsupervised snow facies segmentation model designed to capture the overall properties of the snowpack independently of the single-pass InSAR acquisition geometries; and a subsequent downstream penetration bias regression model. To ensure that the robustness against the InSAR geometries of the first model is preserved, we propose two approaches: first, we employ a fine-tuning approach to transfer the weights of the segmentation model for a downstream regression task, leveraging the knowledge acquired by the pretext segmentation task for the regression of the penetration bias of TanDEM-X DEMs; and second, a multitask learning approach for the downstream task by jointly training both the segmentation and regression models, ensuring that the snow-related feature representations identified during the segmentation task are consistently leveraged to improve the final regression performance. We demonstrate that utilizing the first model as a pretext task improves convergence and overall performance, whereas the multitask approach enables better generalization. Experimental results over the Greenland Ice Sheet during boreal winter, using IceBridge laser altimeter measurements as reference data, demonstrate that our method estimates the penetration bias with a coefficient of determination $R^2$ = 90% and RMSE of 0.65 m, independently of the InSAR acquisition geometry and snow properties. The work performed here is crucial for enhancing the accuracy of TanDEM-X DEMs over snow and ice-covered regions, thereby improving our understanding of glaciological processes and their climatic responses.