Deep Clustering in Subglacial Reflections Reveals New Insight into Subglacial Lakes

Radar images imply subglacial features, including distinct reflections from ice bottom. Different from bedrock interfaces, subglacial lakes generally display smooth and continuous highlights as a special type of ice bottom reflectors in radar images. In this study, we construct a dataset of ice bottom reflectors based on CReSIS radar sounder dataset. A deep learning method is applied to downsample and convert peak ice bottom reflectors to latent space. Unsupervised clustering later separates different types of subglacial reflectors. One reflector type with a sharp shape and high reflect power reveals smooth and continuous distributions in the radar images. The spatial distribution of this reflector type also matches the known subglacial lake distribution. We further applied this workflow to indicate candidate groups of subglacial reflectors similar to the conventional lakes. Results show more lakes are marked in the same radar sounder dataset. This method can automatically indicate subglacial lakes in radar images with high efficiency. The other types of subglacial reflectors can also provide potential references for subglacial studies.