Speaker
Description
Ice sheets play a critical role in regulating global climate and sea-level rise. To enhance the accuracy of climate modeling and predictions, it is essential to comprehend the intricate processes of snow accumulation, transformation, and melting that transpire on ice sheets. A fundamental component of monitoring ice sheets involves identifying distinct layers or units within the snowpack, known as snow facies, each possessing unique physical properties. These snow facies can be classified as: the inner dry snow zone, where melt is absent; the percolation zone, where limited annual melt leads to larger snow grains and small ice structures; the wet snow zone, characterized by substantial summer melt and the presence of multiple ice layers; and the outer ablation zone, where the previous year's accumulation fully melts during summer, revealing bare ice and surface moraine. Variations in melt levels across these snow facies can influence the radar wave penetration when using spaceborne synthetic aperture radar (SAR) to monitor snow- and ice-covered regions. This impacts, e.g., the estimation of the radar mean phase center, which is necessary for the generation of digital elevation models (DEM), leading to penetration bias and an underestimation of the surface topographic height. Accurately estimating this penetration bias is therefore pivotal for reducing uncertainties in determining snow depth, ice thickness, and glacier mass balance through DEM differencing. Previous works have demonstrated the effectiveness of interferometric SAR (InSAR) features for segmenting snow facies and for penetration depth estimation. Bistatic SAR missions like TanDEM-X are particularly well-suited for these tasks thanks to the absence of temporal decorrelation effects, which allows for the generation of high-quality InSAR products. In this work, we explore the capability of bistatic SAR features and deep learning methods for snow facies segmentation in Greenland and Antarctica. We propose a weakly supervised learning convolutional neural network (CNN)-based approach that leverages raster intersections during the mosaicking process of multiple TanDEM-X acquisitions, allowing us to optimize a robust model for various bistatic acquisition geometries. We compare the segmentation results with cumulative melt maps and in situ measurements, demonstrating the potential of bistatic missions for monitoring the evolution of snow facies. Lastly, we discuss how the features extracted for this purpose can be used to estimate the penetration bias of bistatic X-band SAR systems, independently of acquisition parameters. Given the lack of large reference data sets for the estimation of the penetration bias, this work aims at understanding the impact of an effective bistatic feature extraction and utilization, which allows for reducing the amount of labelled data required to fulfill a fully supervised regression task. This is a crucial aspect for, on the one hand, improving the accuracy of the TanDEM-X DEMs over snow and ice-covered areas, and, on the other hand, preparing for future bistatic InSAR missions, such as the forthcoming ESA Harmony mission, which will require similar paradigms for coping with radar penetration into volumes.
