| 21 Jul 2026
Automatic classification of the hydrothermal structure of polythermal glaciers from ground-penetrating radar using deep learning techniques
Abstract. Knowing the internal spatial distribution of cold and temperate ice within polythermal glaciers is essential for understanding and modelling their dynamics. The boundary between both types of ice, termed as cold-temperate transition surface (CTS), is usually determined from ground-penetrating radar (GPR) data, given the permittivity contrast between cold ice and water-rich temperate ice. This task is traditionally manual and time-consuming. We here present an approach based on deep learning for the automatic classification of cold ice, temperate ice and bedrock. We used deep learning algorithms based on convolutional neural networks (CNN), following a feature pyramid network architecture. The training data were collected in Svalbard, using a GPR with central frequency of 20–25 MHz, along various campaigns spanning the period 2008–2022, carried out on glaciers in Sabine Land, Nordenskiöld Land, Wedel Jarlsberg Land and Nordaustlandet. To address data scarcity, we evaluated two dataset expansion strategies: generating synthetic radargrams by forward modelling of electromagnetic wave propagation, and a data augmentation scheme in which each GPR profile is split into 200 m segments treated as independent samples. Starting from a baseline trained on the original GPR dataset, these strategies were applied individually and in combination, defining four experiments. Our results show in general satisfactory values of the performance metrics. For instance, the intersection over union (IoU) metric reaches values of 64.2 ± 2.7 %, 77.0 ± 2.4 % and 98.1 ± 0.2 % for the classes cold ice, temperate ice and bedrock, respectively.
Unai's pre-print
================
I would like to thank the authors for this good paper. The automatic classification of cold and temperate ice from GPR data is a relevant and new application of deep learning in glaciology, and the open data and code are a valuable contribution to the community. Below are my comments, which I hope help to improve the manuscript.
>> The introduction could use more recent citations, particularly in the background on polythermal glaciers (lines 15 to 37).
>> A citation is needed for the claim between lines 26-29.
>> Also for the claim in 31-35 there needs to be reference(s)."
>> Lines 86 to 88, use of DL: I suggest also citing the review article "Feature tracing in radio-echo sounding products of terrestrial ice sheets and planetary bodies" (https://doi.org/10.5194/tc-19-2159-2025), because it gives a comprehensive overview of the methods developed for the detection of horizons and other features in radar data, so readers could get more information there.
>> Line 89: CNNs can be trained in a supervised, semi-supervised or unsupervised way.
>> Lines 89 to 98: please discuss in more detail how the synthetic radargrams compare to the real data, as the pixel intensity distribution in synthetic and real data could be different. Also, line 103 states that the CNN is trained on synthetic data alone, but none of the four experiments does this.
>> Line 123: GPUs do not reduce the computational cost, they reduce the computation time.
>> Line 161: picking of the CTS in the real data was done in ReflexW, right? Since the CTS for the synthetic data is given by the model, it is not picked. Please make this clear, and explain how the labels of the synthetic data were converted from depth to time.
>> Lines 170 to 182 need a more rigorous description (e.g. channel dimensions, the segmentation head, and how single-channel radargrams are fed to the ImageNet pre-trained backbone).
>> Lines 198 to 207 and Table 5: more information on the training data and the training is needed for each of the experiments. Please give, for each experiment:
- the number of samples used for training, validation and testing, and whether the numbers in Table 5 include the validation and test sets
- the number of synthetic radargrams. This is not stated in Section 2.2. From Table 5, experiment 2 adds 92 synthetic samples (230 - 138), while experiment 4 adds 85 (1668 - 1583). This difference needs to be explained.
- the share (%) of synthetic data in experiments 2 and 4
- the number of training samples made from each radargram
- the original size of the radargrams (number of traces and samples) before downsampling
- how the ten cross-validation runs and the five weight initialisations are combined into the reported results, and whether the hyperparameters in Table 4 are the same for all experiments
>> It should be explained more whether every radargram is downsampled to 256x256 or broken into windows of this size for training. If a whole profile of several km is downsampled to 256 pixels in width, the horizontal resolution is very different between experiments 1-2 and 3-4. Please clarify this.
>> In the results, define what each of TP, TN, FP and FN are in this case (e.g. pixel counts, and whether they are summed over the whole test set or averaged per radargram).
>> Experiment 4 has about 5% more data than experiment 3, which is synthetic data. However, this small increase in data has increased the accuracy considerably.
>> Figures 5 and 6 do not have axes; they should be clearer in terms of depth and distance.
>> Line 233: experiments 1 and 2 do not have augmentation. The 200 m splitting needs to be better clarified (e.g. overlap between segments, and how the remaining part shorter than 200 m is handled).
>> Lines 233 to 236: either tell why you think the addition of synthetic data does not help in experiment 2 but does in experiment 4, or at least give your hypothesis. You are just reporting.
>> Line 260: it is a remarkable and comprehensive view of the authors to also include the shortcomings of such models. Very good work.
>> In general, synthetic data do not seem to be worth the effort, as they require high computational power, as mentioned. Also, they do not always capture the same pixel intensity distribution as the target radar system. Therefore, they might practically be the results of "a different radar system", which, due to the limited generalisability of DL models, takes the focus of the model away from learning the specifics of one radar system and makes it learn features from two different radar systems.