Abstract: The ice-rock interface at the bottom of the polar ice sheet records the historical evolution of the ice sheet, reflecting the geometric characteristics of the ice layer and the environmental properties of the ice bottom. Moreover, it is an important indicator both for inferring the dynamics of the ice sheet and for explaining the subglacial topography. Airborne ice radar is an effective method for detection of polar ice cover; however, the radar data contain many types of noise due to the detection environment and to the limitations of the instrument itself. For efficient and accurate extraction of the bedrock interface and the ice surface in an ice radar image, and for reduction in noise interference, we established an automatic method for extraction of the ice surface and bedrock interface in a radar image based on the deep learning pix2pix algorithm and observational airborne ice radar data of Princess Elizabeth Land acquired during the 32nd Chinese National Antarctic Research Expedition in 2015–2016. Experimental results showed that the accuracy of the model proposed for extracting the ice surface/bedrock interface was 0.863/0.948, and that the peak signal-to-noise ratio was 24.814 dB; i.e., higher than that realized previously using the K-SVD and CycleGAN algorithms. Thus, the proposed method was demonstrated to be more effective in removing noise, improving image quality, and restoring the common manual extraction effect with high accuracy.

Key words: pix2pix, ice-penetrating radar, bedrock extraction, subglacial topography, Antarctic ice sheet