Deep Clustering in Radar Subglacial Reflector Reveals New Subglacial Lakes

Abstract. Ice-penetrating radar (IPR) imaging is a valuable tool for observing the internal structure and bottom of ice sheets. Subglacial water bodies, also known as subglacial lakes, generally appear as distinct, bright, flat, and continuous reflections in IPR images. In this study, we collect and generate a dataset of one-dimensional reflector waveform features from IPR images of the Gamburtsev Subglacial Mountains region in the CReSIS database, to investigate these features. We apply a deep learning method to reconstruct the reflector features, and subsequently downsample the features to a low-dimensional vector representation. An unsupervised clustering method is then used to separate different types of reflector features, including a reflector type corresponding to subglacial water bodies. The derived clustering labels are used to detect features of subglacial water bodies in IPR images. Using this method, we compare the new detections with the known lakes inventory. The results indicate that this new method identified additional subglacial lakes that were not previously detected, and some previously known lakes are found to correspond to other reflector clusters. This method can offer automatic detections of subglacial lakes and provide new insight for subglacial studies.