General Comments:

The work described in this publication details a glacial termini dataset for Svalbard, which covers 149 marine-terminating glaciers from 1985-2023, and is comprised of 124919 fully-delineated calving front positions. This data is generated using an automated processing pipeline to download optical (Landsat, Sentinel-2) and SAR (Sentinel-1, Terra-ASTER) remote sensing imagery via Google Earth Engine, and process the images into vectorized polylines using deep machine learning. This methodology is sound, building on and improving existing machine learning methods, which has been published alongside the data as the CORBA framework. The manuscript contains uncertainty and validation of the method/dataset, which provides appropriate bounds and checks on the accuracy and rigor of the automated methodology. The overall error is within human levels of accuracy of error (46 ± 21 m < 78m, Goliber et al. 2022). Additionally, scientific analysis is performed on the dataset, showing calving front change in agreement with existing literature (R^2 = 0.98 during the time period of 2008-2022 with Moholdt et al., 2022) concerning the changes to Svalberd’s marine-terminating glaciers.

The dataset itself consists of a single GeoPackage containing 4 outputs, which include the glacial centerlines, rectangular polygonal domains, front change time series/rates (in m/yr), and the calving front polyline traces. Metadata provides relevant information for reference and potential reproduction/reprocessing. It is well prepared, and in a common format that is easy for community members to use in future work. It is well documented both within the manuscript and in accompanying materials provided along with the dataset.

The publication is well done, and is largely free of grammatical errors and typographical issues. There are only minor remarks to be addressed by the authors, after which I can recommend acceptance at the editor’s discretion.

Specific Comments:

• For modelers and other community members, it would be useful to have the glacial termini data in the form of areal change or land/ocean polygonal masks, in addition to just polyline (as in Kochtitzky and Copland, 2022). This would involve connecting the existing calving front polyline endpoints to some static coastline that coincides with the boundaries of the polygonal domains already provided in the dataset. While this is not necessary to do and may be outside the scope of the publication, this would help reduce the processing required for community members to use the calving front data to mask ice extent changes in ice sheet models, or for measurement of areal change.
• Complementary with the previous comment, the glacial fjord mask is inexact, and enforces sharp cutoffs on the calving front endpoints where they join the fjord walls. While it may be out of scope to reprocess the dataset with improved fjord masks, this may prove useful to the community to avoid errors when connecting the calving fronts positions to land.
• For the glacial fjord boundaries, if it is possible or seen fit to include the glacial fjord masks in the final dataset, it might prove useful for further processing of the calving front traces. Again, this Is not strictly necessary if it is outside the scope of this work, but should increase the ease of using the data provided by this publication.
• Figure 10 shows front rate change differences w.r.t. existing data, and Figure 12 shows a spatial distribution of calving front rate changes. In addition to these, it would be a good visual summary to show a histogram or distribution of calving front rate changes. Any relevant statistics (mean/median) would be good to include as well.
• While the standard error measurement seems to be the Area/Front length or Median Mean distance between predicted and ground truth fronts, this should be clarified in the text, as it may not be obvious how this is calculated without prior knowledge in the field.

Technical Comments:

• Figure 8: Consider adding more histogram bins for better granularity of the uncertainty/error distribution (i.e., 10m or 5m bins instead of 25m bins).
• Figures 3, 9, 10: Some text is small/hard to read – consider increasing the font size in these plots.

General Comments

This paper presents a comprehensive dataset, generating 124,919 glacier termini for 149 marine-terminating glaciers in Svalbard. Employing an innovative automated deep learning pipeline, the dataset integrates multiple optical and SAR satellite images to enhance temporal coverage. The pipeline encompasses a GEE-based automated data collection method, the CORBA deep learning framework, and a suite of post-processing techniques designed to filter out inaccuracies, ensuring the integrity of the dataset.

Overall, the paper is well written, the method is solid, and the dataset is of high quality. The dense calving fronts will benefit future scientific research about the estimation of glacier mass loss and calving mechanism. Nonetheless, there are specific comments that require attention before acceptance, at the editor's discretion, as outlined below.

Specific Comments

Line 108: Why does the number of marine-terminating glaciers listed here (220) differ from the 149 mentioned in the abstract?

Line 115: Did the author merge images captured on the same date before applying a non-data pixel threshold? The satellite stripe footage might cover only a portion of a glacier, but merging it with its adjacent stripe could provide complete images. This is generally applicable to optical images but not as much for SAR images. Therefore, combining images from the same satellite on the same date before applying the non-data threshold might yield a greater number of images.

Line 117: A total of 1,135,074 satellite images were downloaded, yielding 124,919 glacier termini. This suggests an abandonment rate of approximately 90%. I am concerned about this aspect because the utilization of the inter-quartile range in results filtering relies on the assumption that the majority of the results are of high quality.

Line 152: The test error for CALFIN test set is 99 ± 10 m. What about the Baumhoer dataset?

As a suggestion for the post-processing step, I recommend applying 2.3.2 first, followed by 2.3.1. The rationale behind this recommendation is that a result might be accurate on the glacier terminus but incorrect on the fjord boundary, and we can still make use of this result. Applying 2.3.2 first would allow for the retention of such results, while applying 2.3.1 first might discard them. It's worth noting that this is a suggestion for the authors to consider, rather than a strict requirement.

Line 270: What about before 2014?

Line 274: The full-thickness calving that produces tabular icebergs could cause glacier terminus to retreat by several kilometers within several days.

Line 277: I assume the visual checking is performed manually, potentially impacting the pipeline's level of automation. How many incorrect calving fronts are identified through this visual checking?

Figure 7: Is retreating symbolized by an upward trend or a downward trend? I suppose both (b) and (e) experienced retreating until recent years, but (c) has an upward trend while (d) has a downward trend. Please consider adding labels to show the retreat in all the figures of terminus variation.

Line 327: The test error for the Baumhoer dataset is missing.

Line 331: If I understand correctly, the uncertainty is assessed by quantifying the variance in model-predicted terminus traces from different satellite images captured on the same date. Is that accurate? If so, I recommend utilizing the test error to represent the dataset error and suggest corresponding adjustments to the abstract. Additionally, could you provide insight into how to deal with the duplicate results originating from different satellites on the same date?

Section 3.2.2 Some components of this section, such as the dataset used for result validation and the methodology employed for validation, may be more appropriately placed within the Methods section. I recommend restructuring this section to ensure a clearer delineation of content.

Section 3.2.2: A more critical question arises regarding the necessity of using the rate of terminus changes for result validation. It seems more straightforward to calculate the difference between the results of this study and Moholdt et al. (2022) data on the same date, similar to the author's approach for quantifying uncertainty and test error. Such a direct comparison could serve as an additional test error, complementing the other two (CALFIN and Baumhoer test errors) to more comprehensively represent the error in this study's results.

Figure 12: The arrows in (a) can be misleading. It could make readers think the white period within the two arrows is the surging period.

General Comments:

Li et al. present a calving front dataset for 149 marine-terminating glaciers in Svalbard over the time span 1985-2023. The fronts were extracted from different satellite missions (Sentinel-1/2, Terra-ASTER, Landsat) available at Google Earth Engine (GEE). The fronts are extracted by the novel COBRA model based on a CNN and active contour model in contrast to previous U-Net-based approaches. The mapping accuracy along centerlines is very high (46 ± 21 m) and the calculated calving front change rates are consistent with existing studies (R² of 0.98) based on manual front delineations. The created dataset is freely available and uses well-known RGI glacier domains making it a valuable resource for glaciological studies on calving front dynamics and ice-ocean interactions.

The manuscript is of high-quality, very well written and accompanied by good graphics. Only a few minor remarks should be addressed by the authors to improve the manuscript:

Specific Comments:

L70: Are the short acquisition intervals of 1-3 days possible year-round or only in summer when polar night is not present?

L93: Why did you chose to use Sentinel-1 EW data and not the higher resolution IW data? For the glaciers in Svalbard I would assume the higher resolution would be very beneficial.

Section 2.3.4 Excluding erroneous calving front detections has been a challenging task for a while and different approaches were developed such as the automated screening module based on front geometries (Zhang et al. 2023) or a time series approach (Baumhoer et al. 2023). Could you provide some numbers on how many fronts were identified by the automated approach and how many were removed by visual inspection? Additionally, it would be very beneficial to state the ration between successful front delineations and removed ones based on the number of available input images. Are these ratios more or less the same or are there fronts where COBRA preforms better/worse and less/more fronts are excluded?

Section 3.2.1 The uncertainty measurement is based on fronts delineated on the same day from different image sources. This is for sure a good approach as it does not require tedious manual delineations. Nevertheless, it should be kept in mind, that several front positions at the same day are more likely available since the launch of Landsat-8/9 and Sentinel-1/2. Could you provide a time line to see the temporal distribution of your accuracy assessment? Does the accuracy assessment cover the entire time series between 1985 to 2023 or is it temporally clustered? Furthermore, the accuracy is calculated along one centerline not accounting for errors at the margins of the glaciers. That makes it difficult to compare the accuracy to other studies as they measured the accuracy by the mean difference between two fronts. Please consider providing accuracy measures of the mean distance error considering the entire front to make your accuracy assessment comparable to other studies (Cheng et al 2021, Loebel et al. 2023).

Section 3.2.2 Why did you decide to compare the rates and not the front positions itself? It would be very interesting to provide a mean distance error also for the Moholdt et al. 2022 dataset. For sure, also manual front delineations are not 100% accurate but the comparison would give a good estimate for general deviations that have to be considered. Also see the comment above on accuracy measures with the mean distance error.

Figure 12 (b): Nice exemplary figure but unfortunately it is hard to see the Moholdt et al. fronts below the purple fronts of your dataset. Additionally, you could provide a magnified part of the figure in the area where fronts are overlapping.

Figure 12 (c/d): It is a bit misleading that the retreat distance is positive and the advance negative. I would recommend to flip the y-axis and display frontal change where a surge is positive as the front advances and retreat is negative.

References mentioned in this review:

Baumhoer, C. et al. (2023): IceLines – A new data set of Antarctic ice shelf front positions, Scientific Data, 10(1), p. 138. Available at: https://doi.org/10.1038/s41597-023-02045-x

Cheng, D. et al. (2021):Calving Front Machine (CALFIN): glacial termini dataset and automated deep learning extraction method for Greenland, 1972–2019, The Cryosphere, 15(3), pp. 1663–1675. Available at: https://doi.org/10.5194/tc-15-1663-2021.

Loebel, E. et al. (2023): Calving front monitoring at sub-seasonal resolution: a deep learning application to Greenland glaciers, The Cryosphere Discussions, pp. 1–21. Available at: https://doi.org/10.5194/tc-2023-52.

Moholdt, G., Maton, J., Majerska, M., and Kohler, J. (2022): Annual frontlines of marine-terminating glaciers on Svalbard, https://data.npolar.no/dataset/d60a919a-9cc8-4048-9686-df81bfdc2338.

Zhang, E., Catania, G. and Trugman, D.T. (2023): AutoTerm: an automated pipeline for glacier terminus extraction using machine learning and a “big data” repository of Greenland glacier termini, The Cryosphere, 17(8), pp. 3485–3503. Available at: https://doi.org/10.5194/tc-17-3485-2023.