General Comments:

This paper uses deep learning to extract glacier termini over Greenland, Alaska, and Antarctica. The author uses two glaciers that are not included in the training set to test the generalization ability of the deep learning network. One has good results while the other has quite a large uncertainty. Overall, I think the paper is well-written and clear, and I appreciate that the author has a comprehensive review of the recent studies on a similar topic. However, I have several concerns about the manuscript, which are listed below.

Major concerns:

• Compared with previous studies, this manuscript has room to improve regarding innovation. The previous study has used deep learning to extract the glacier termini from multiple remote sensing datasets. If combining multi-sensor and long-term images brings innovations, the author needs to justify the reason (where are the challenges and how do you solve them).
• The second concern is about the data quality and quantity. The uncertainty is high for one of the two test glaciers, and the terminus results are not usable for further analysis. Also, deep learning aims to produce a large number of termini by automating terminus delineation. But the number of terminus traces derived from this study is limited. Considering the ESSD is a journal focusing on data, I think the manuscript needs to improve in this regard.

Specific Comments:

Page 2 Line 58:  It would be nice if the author could explain more about why the dataset from this study is beneficial for bridging the gap regarding the evaluation among different studies.

Page 10 Line 212: Please explain more about re-mapping.

Page 10 Line 222: What is the rationale behind identifying four zones since this study will only pick the boundary between glacier and ocean.

Page 11 Line 246: Morphological dilation can resolve the imbalance to some degree but will also cause uncertainty in terminus delineation. The more balanced between positive and negative pixels, the larger uncertainties the terminus will have.

Page 13 Line 277: I like the idea of applying Gaussian importance weighting before taking the average. It would be nice if the author could provide more details.

Page 15 Line 298: Is this a universal threshold? Please explain more about setting this threshold and what will happen if the threshold is too large. Usually, the threshold would be 0.5. The threshold can indeed be different as long as it is justified.

Page 21 Line 436: the uncertainty value might be wrong (150 m?). I couldn’t find this value in Table 7.

Page 22, Line 461: Maybe consider including such post-processing in this study. Using the bedrock mask can eliminate the fjord boundaries.

Table 8, Table 9, and Figure 9: The predicted terminus position for the Columbia glacier deviates from its true position, and the uncertainty for the Columbia glacier is too large.

General Comments:

This manuscript primarily describes the release of a deep learning focused glacial feature tracking benchmark dataset on multi-mission SAR data. Additionally, it describes a reference deep learning implementation that performs automatic glacial termini extraction/land classification on the benchmark dataset, and provides tools for further use/standardization for easier comparison across different deep learning methodologies. The literature review is robust, and covers existing work done in the field regarding glacial feature datasets and automated extraction algorithms. This study uses data sources with high spatio-temporal and satellite variety for its benchmark dataset, exemplifying robustness and rigor in the final data product. Specifically, the dataset consists of 7 diverse tidewater glaciers from Greenland, Alaska, and the Antarctic Peninsula, from 1985-2020, across 6 separate SAR satellite families. The dataset itself consists of 681 total image pairs and accompanying metadata, which is a significant contribution to the field. This will encourage the development of more accurate and robust deep learning models for glacial feature extraction and land classification of cryospheric regions. Furthermore, the reference deep learning methodology provides a baseline for easy cross-comparison across future studies in glacial feature extraction. Robust pre- and post- processing techniques are employed to produce usable data products from the deep learning classification module. Accuracy is determined using the Mean Distance Error and Intersection over Union metrics that are commonly used by other studies in the field. While the deep learning model on the dataset testing set shows room for improvement, it fulfills the goals of this study to provide benchmarks for future models.

Overall, this study provides contributions to the field of deep learning as applied to automated glacial feature extraction and classification. Ultimately, this data will help encourage further development in the field, allow for increases in accuracy in derived data products, and ultimate enable better projections/modeling of sea level change and in AOGCMs. The overall manuscript is well done, free of major grammatical errors, with only minor remarks (see below and specific comments) to be made and addressed before acceptance at the editor’s discretion.

Major Comments:

• It would be helpful to include the type of data covered by each study in Table 1, such what types of glacial features are provided by each dataset. These may include (but are not limited to) one or more of the following: full-line delineation (ESA), a centerline position (King), or glacial outline (GLIMS).
• Since the primary data of interest is the training/testing benchmark dataset, it would be beneficial to further emphasize this in some way throughout the manuscript, such as including the total number of image pairs/metadata in the abstract. This will enable readers to see the primary contribution more easily, and enhance visibility/usage of this work within the field.

Specific Comments:

P5 L108:  “Lansat” -> “Landsat”

P19 L364: “So there is a trade-off between patch size and batch size” Rephrase.

P24 L463: “are” -> “is”

P32 L539: “25rd” -> “25th”