Automatic detection of cold-temperate transition surface in polythermal glaciers using GPR and machine learning

Ground-penetrating radar (GPR) has been shown to be an effective tool to infer the hydrothermal structure of polythermal glaciers. Knowledge of this structure is fundamental to the study of their dynamics. The cold-temperate transition surface (CTS) is the englacial boundary between cold and temperate ice. It can be identified by GPR because of the contrast in permittivity between dry cold ice and water-rich temperate ice. However, the interpretation of the CTS using GPR has traditionally been a very time-consuming and manual process. Here we show a procedure based on machine learning for detecting CTS automatically. The data used for training a convolutional neural network were collected in both Svalbard, in the Arctic (radar with central frequency of 25 MHz), and the South Shetland Islands in the Antarctic Peninsula region (200 MHz central frequency). Various metrics revealed success rates in the classification in the order of 90%. The size of the training dataset is limited, so current work is focused on enlarging its size by using random variations of synthetic radargrams generated by forward modelling with gprMax.