AI-Driven TanDEM-X Penetration Bias Estimation in Antarctica Using ICESat-2 and ECMWF Data: Implications for the NASA Surface Topography and Vegetation Decadal Survey Incubation study

Abstract. TanDEM-X radar penetration into polar ice introduces significant uncertainties in digital elevation models (DEMs), consequently affecting the accuracy of glacial mass balance estimations. This limitation arises from the variable penetration depth of radar signals in snow and ice, which can cause the DEM surface to deviate from the true physical surface. X-band radars, operating in the 8–12 GHz range, interact with ice through absorption, reflection, and transmission. These interactions heavily depend on variables such as ice temperature, water content, snow density, and salinity, which directly influence microwave penetration depth and signal behaviour. Advancements in remote sensing have leveraged X-band synthetic aperture radar (SAR) to characterize snow and ice surfaces, facilitating applications such as sea ice classification and snow layer analysis. However, X-band SAR signals often exhibit biases in elevation measurements due to their partial penetration into snow and ice. This bias complicates efforts to integrate SAR-derived digital elevation models (DEMs) with laser altimetry data like ICESat-2, where penetration is negligible. Therefore, correcting X-band biases is critical for achieving high-accuracy surface elevation models and reliable glaciological assessments. Here we address these challenges by integrating neural network techniques with TanDEM-X (TDX) DEMs, Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) altimetry data, and environmental parameters from European Centre for Medium-Range Weather Forecasts (ECMWF). We leverage about 300,000 ICESat-2 pointwise measurements acquired within 30 days of TDX measurements spanning 2021–2024 in Antarctica. Additionally, we consider a diverse dataset of snow and atmospheric variables including temperature, snowfall, snow depth, and wind speed to model and predict X-band penetration biases across Antarctica. This approach advances existing methods by automating bias correction and enhancing the integration of SAR and altimetry datasets achieving mean bias correction of the order 1 cm with a Root Mean Square Error (RMSE) of about 1 m and maximum errors of the order of 10 m. Findings from this work provide actionable insights for improving elevation model accuracy in different ways. First, we offer the retrieved pointwise TanDEM-X, ICESat-2 and ECMWF dataset open access for future studies through a dedicated Zenodo page. Second, we provide both our trained network weights and our algorithms in the form of Jupyter Notebooks on Github for improved reproducibility. Most importantly, we discuss the broader efforts in the NASA Surface Topography Elevation (STV) decadal survey incubation study and polar ice monitoring by addressing penetration biases which is one of the most critical uncertainties in radar remote sensing of snow and ice.