Signal similarity-informed generative adversarial network for prediction of basal wetness conditions in Antarctica: a case study in the AGAP region

Identifying basal wetness conditions beneath the Antarctic ice sheet is essential for understanding subglacial hydrology, basal traction and long-term ice dynamics. However, conventional reflectivity-wetness analysis methods face significant challenges, such as the reliance on inconsistent reflectivity thresholds and manually labelled training datasets. In this contribution, we propose a signal similarity-informed generative adversarial network (SSIGAN), an unsupervised anomaly detection framework, to predict and analyse the basal wetness conditions using focused radio-echo sounding data. Taking the AGAP region as a case study, the method reformulates the wetness classification as an anomaly detection problem, removing the need for labelled data. It extracts radar waveform similarity features to generate an anomaly score (R1), which is combined with an inverted geometrically corrected and normalized bed return power score (R2) to form a wetness score (R3) that enhances the classification separability. Validations are conducted through spatial comparison with existing inventoried subglacial lakes and consistency analysis at radar intersection zones. The results demonstrate that this method can reliably distinguish the transition between wet and dry basal interfaces, highlighting its potential for continental-wide mapping of basal wetness in Antarctica. This article is part of the Theo Murphy meeting issue 'Next generation ice-sheet bed measurements'.