The authors have demonstrated commendable expertise and innovation in their work by introducing CNN in the field of radar altimetry retracking. Their research represents a significant contribution to the field, offering valuable insights and promising results. By employing CNN, the authors have opened new possibilities for improving the accuracy and efficiency of altimetry data processing, and possibly (not addressed) speeding up otherwise time-consuming retracking. The paper introduces this innovative approach and provides convincing evidence of its advantages. While comprehensive research is crucial for presenting a well-rounded perspective, the excessive length of the papers can be challenging for readers to digest. One suggestion would be to separate the retracking from a second paper on the applications or move some of the backgrounds to an appendix.  This being said, I only have minor comments on the paper.

One aspect of AWI-ICENet1 which could be addressed is the possible increased efficiency in the processing time. As the authors have great experience in retracking radar altimetry data, it would also be beneficial to highlight the efficiency of the AWI-ICENet1 compared to other methods. We are often faced with very long reprocessing times from the agencies. 

In general, the caption for the many figures is very shallow, please read through them and elaborate on them, so a reader who is not reading all 61 pages can follow the main conclusions of the figures.

L2: “long-term observations of surface elevation change are required to”, agree with the meaning but I would suggest that you also acknowledge other methods and hence remove the “Surface”

L5: The snow penetration is an issue in most places, just leave this sentence open and remove “especially over the…”

L15: This shows a broader application and suggests writing “This technique provides new opportunities to utilize convolutional neural networks in the processing of satellite altimetry data, which can be applied to historical, recent, and future missions.”

L21: missing reference at “2010 onwards”

L30 may start with “Ku-band satellite altimeters…” as the mentioned satellites are all Ku.

L40 Why add the PLRM, this is a pure post-processing product.

L43: ICESat-2 operating at green-wavelength is penetrating the snow. And add references for ICESat-2.

L46: please elaborate on the sentence “Because…”,

L73: the common abbreviation is CNN, please consider using this

L124: why exclude Greenland?

L128: to help the reproducibility of the study please give more insights into the chosen backscatter cross-section values.

L144: “rate in the following”, the paper has a couple of these please proofread the paper once more.

L153 What is the resolution of the applied DEM in the modelling?  (it might not fit here but should be discussed in this relation)

L177: move the specific tensorflow package to the acknowledgement, and add a reference to this library.

L189: With possible differences/drift in Bs in the processing baseline only one should be used. For consistency use E.  

L194: add a reference for ATL06.005

L196: Add the resolution used.

L210: so it is the 2 km product which is used, why is this chosen?

L221: “the correction adapted from Nilsson et al. (2022)” elaborates on what is used.

F5: add a plot of the model vs. test point cloud. Maybe retracked range vs model range. 

L238: As Greenland is different and possibly more complicated it would be nice the see an ROI in Greenland.

F7: thank you for this very convincing figure. How does this look with respect to slope?

L265: Roemer is a better solution however the LEPTA relocation seems to improve even further, how does this affect the results?

L268: monthly crossovers are a very long time span for cross-overs on ice sheets please add a lower time constraining on the timing between orbital crossing evaluated.

L270: how many fall within this outlier filtering (maybe in %)

F8-F10: The ESA ICE2 seems as an outlier compared to the others, suggest removing this from the plots to see the specific difference in the others.

L423: “…focus on the time from January 2019 to December 2021…”

F20: Is the trackiness due to errors in the CS2 or ICESat-2 data?

L472: Guess this is ATL15 this should be mentioned.

General comments:

This paper presents a convolutional neural network (CNN) approach to measure and quantify surface elevation change in Greenland and the Antarctic ice sheets via satellite radar altimetry data. Through extensive analysis, the authors show that their proposed method displays improved performance and reduced uncertainty over traditional retrackers.

The primary strengths of this paper are in the thoroughness of analysis of the performance of AWI-ICENet1 and in comparisons to conventional retracking algorithms. Cross point error analysis is a good way of comparing the performance of each method for identifying the ice surface, as it does not rely on a ground truth (as is typical in supervised machine learning).

Another strength of the paper is the construction of a synthetic dataset that, after training a CNN on it, performs at least as well as (if not better than) conventional methods. It is an impressive contribution in itself to be able to construct a synthetic dataset that is sufficiently close in distribution to the training and testing data such that a deep learning model can be adequately trained on the synthetic data alone.

Specific comments:

Despite the strengths and contributions, my main concern for this paper is that it does not situate itself within the context and literature of deep learning approaches applied on data from satellite or airborne sounding of ice sheets. To my knowledge, the majority of this work has involved using deep learning to track ice and bedrock layers beneath the ice surface, but these approaches still seem quite relevant, at least to briefly discuss. These are some such prior works:

1. S. Dong, X. Tang, J. Guo, L. Fu, X. Chen, and B. Sun, “EisNet: Extracting bedrock and internal layers from radiostratigraphy of ice sheets with machine learning,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–12, 2021.
2. M. Liu-Schiaffini, G. Ng, C. Grima, and D. Young. “Ice thickness from deep learning and conditional random fields: application to ice-penetrating radar data with radiometric validation,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1-14, 2022.
3. M. H. Garcia, E. Donini, and F. Bovolo, “Automatic segmentation of ice shelves with deep learning,” in Proc. IEEE Int. Geosci. Remote Sens. Symp., Jul. 2021, pp. 4833–4836.
4. H. Kamangir, M. Rahnemoonfar, D. Dobbs, J. Paden, and G. Fox, “Deep hybrid wavelet network for ice boundary detection in radra imagery,” in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Jul. 2018, pp. 3449–3452.
5. R. Ghosh and F. Bovolo, “TransSounder: A hybrid TransUNet-TransFuse architectural framework for semantic segmentation of radar sounder data,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–13, 2022.
6. E. Donini, F. Bovolo, and L. Bruzzone, “A deep learning architecture for semantic segmentation of radar sounder data,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–14, 2021.
7. Y. Cai, S. Hu, S. Lang, Y. Guo, and J. Liu, “End-to-end classification network for ice sheet subsurface targets in radar imagery,” Appl. Sci., vol. 10, no. 7, p. 2501, Apr. 2020.

The authors discuss some prior machine learning methods applied to the cryospheric sciences, but this discussion only includes one deep learning approach (Fayad et al. (2021)). I would recommend that the authors include a brief discussion of what distinguishes Fayad et al. (2021)’s setting/model from the current paper. I would also recommend the authors incorporate an additional discussion of the above (and related) references on page 3, or where relevant.

Most of these prior approaches applying CNNs to identify ice and bedrock layers beneath the ice surface use 2D CNNs to capture spatial correlations in the along-track direction. However, to my understanding AWI-ICENet1 only performs 1D convolutions in the radar return at a specific waveform in time. Why was this design choice made? It seems likely that capturing spatial correlations could aid the prediction of a deep learning model, especially in regions where data is noisy and measurements are highly variable. Please add a discussion/comparison of AWI-ICENet1 to prior 2D CNNs methods in the paper.

The authors motivate the use of a synthetic dataset by discussing how ground truth data cannot be obtained by using airborne or ground-borne sounders due to the different footprint sizes. While the answer may be clear to someone in the cryospheric community, some members of the machine learning community may ask why the ground truth cannot simply be set to be the output from a retracking algorithm that the CNN can simply learn to approximate (albeit potentially improving runtime). I would recommend that the authors briefly address this question in the introduction as well.

Can the authors also provide a brief description/comparisons of runtimes between the algorithms?

Technical corrections:

There are several typos in the paper, and some of the language is unclear; please proofread the paper closely again. For instance, there are two typos in line 144, and on line 30 “esa” should be “ESA.” On line 270, there seems to be an extra $x$. In lines 504-505, it is unclear what is meant by “the nature of things.” The wording in line 93 should also be tweaked for grammar and combined with the previous sentence: “Reason is the very different footprint size of the two systems.”