Application of Deep Learning Techniques to Ice Sheet Surface and Bed Interface Detection

Abstract:

Ice penetrating radar (IPR) at VHF frequencies is used to measure the properties of ice sheets. IPR systems are typically mounted on aircraft that fly several hundred meters above the surface. The detection of the air-ice and ice sheet bottom (“bed”) interfaces in the IPR data is necessary for the calculation of ice thickness measurements, which is crucial in evaluating climate change-induced effects on glacier mass balance over time, among other applications. Automating the ice surface and bed identification process has the potential to vastly reduce the amount of human effort expended in these tasks, thus leading to faster ice thickness measurements with fewer resources. In practice, the identification of these interfaces from airborne radar sounder data is a time-consuming task and generally requires supervision by a human expert. Previous work to employ more automated methods include the application of convolutional neural networks to identify the surface and bed interfaces in ice-sounding radar data by reading in radargram records as images.

Here we present an alternative approach using a recurrent neural network (RNN), a class of deep neural networks, operating on individual radar data traces in a sequential manner. This input method has, to our knowledge, not previously been attempted on radar sounding data. The model is trained on a human-identified ice surface dataset and receives the sequential traces of radar sounding data as input. The model is able to achieve good results on the holdout validation set. We also present the results of applying our surface identification model to the task of bed identification. Apart from the direct applications to ice thickness measurements, this work is an example of the possible capacity that deep learning techniques may have in the analysis of IPR data in the coming years, including the identification of layers within the ice and identifying and quantifying the distribution of subglacial lakes at the bed.