IceAnatomy: a benchmark dataset and methodology for automatic ice boundary extraction from radio-echo sounding data

Abstract. The measurement of ice thickness is of great importance for the accurate estimation of glacier volume and the delineation of bedrock topography. In particular, this is a crucial factor in forecasting the future evolution of glaciers in the context of a changing climate. In order to derive the ice thickness, the travel time of electromagnetic waves in radargrams acquired by radio-echo sounding (RES) systems is analyzed. This can only be achieved by identifying the ice surface and underlying ice bottom in corresponding radargrams. Manually identifying these two reflection horizons in RES data is a laborious and time-consuming process. Consequently, scientists are attempting to automate this task through the use of techniques such as deep learning. Such automation can significantly reduce the time between a field campaign and the calculation of the glacier's ice thickness distribution. In this paper, we present the first benchmark dataset for delineating the ice surface and bottom boundaries in RES data to facilitate standardized comparisons of deep learning models in the future. The “IceAnatomy” dataset comprises radargrams and the corresponding manual picks, amounting to a total of over 45 000 km of observations. The RES data originate from three sources: FAU (Friedrich-Alexander-Universität Erlangen-Nürnberg, Institute of Geography), CReSIS (Center for Remote Sensing and Integrated Systems), and AWI (Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research). The dataset comprises different RES systems as well as different pre-processing methods. In addition, the data were acquired over a large range of geographical and glaciological settings, featuring different thermal regimes present in Antarctica and the Southern Patagonian Ice Field. This diversity ensures that the models' behaviors can be analyzed in different scenarios. We define a standardized train–test split for each source in the dataset. This allows us to introduce not only a baseline model trained on the entire training set (the “omni”-model), but also three source-specific baseline models. The source-specific models are trained exclusively on the subset of the training data acquired by the specified source. The baseline models provide an initial benchmark against which subsequent models can be compared. The source-specific models demonstrate more accurate results than the omni-model. For the FAU, CReSIS, and AWI test sets, the source-specific models achieve mean meter errors of 2.1, 23.1, and 4.9 m for the ice surface and 9.1, 78.2, and 29.3 m for the ice bottom. In relation to the mean measured ice thickness of the test set, these errors equate to 1.2 %, 3.1 %, and 0.3 % for the ice surface and 4.9 %, 10.4 %, and 1.5 % for the ice bottom. The dataset and implementation are available at https://doi.org/10.5281/zenodo.14036897 (Dreier et al., 2024) and https://doi.org/10.5281/zenodo.14038570 (Dreier, 2024).