Corr et al. present the first inventory of supraglacial lakes and channels across the West Antarctic Ice Sheet. The data quality and data presentation are excellent in the linked Zenodo repository (Corr et al., 2021). This paper is an excellent pair with Stokes et al. (2019), and, together, the papers provide a baseline dataset of supraglacial hydrology across the Antarctic (as highlighed in the author's Conclusion). The figures are clear overall, and I just have a few suggestions on some rough locations in the manuscript.

L12: The sentence that begins with percentage is a comma splice. Suggest a re-write to not begin the sentence with a number and to split this sentence in two.

L34: Cite this inventory for the EAIS.

Figure 1: Delineate WAIS and the EAIS on Figure 1.

Introduction: This section could use proofreading to identify a few typos here and there. Several sentences could be shortened to increase writing clarity (e.g., run-on sentence from L61–65). 

Section 2.1.4. Lake vs channel classification: It would be great to have an additional figure that shows how the shape index metrics are used within the K-Means clustering approach. The figure cited in this section (Figure 4) only shows supraglacial channel and lake outlines, but not how the various metrics are combined. Additional details on how the K-Means algorithm was applied communicated via a descriptive figure would help better illustrate the "20 distinct clusters" (L195) identified. What do the top 3 or 4 clusters in the dataset look like? Suggest revise Figure 4 to illustrate the K-Means clustering approach, or add a new figure in addition to Figure 4.

L310: Move this equation up closer to where it is cited in the text. 

Review of the manuscript “An inventory of supraglacial lakes and channels across the West Antarctic Ice Sheet” submitted to the ESSD special issue on “Benchmark datasets and machine learning algorithms for Earth system science data“

The manuscript presents an approach to map supraglacial lakes and channels for the year 2017 across the West Antarctic Ice Sheet on the basis of Sentinel-2 and Landsat-8 data and spectral thresholding. It complements the study of Stokes et al 2019 that used a similar approach to map supraglacial lakes around the margin of the East Antarctic Ice Sheet. The resulting inventory of mapped lakes and channels is suggested as benchmark datasets for Earth system science data.



Although the manuscript is generally well written and well presented, I have some major concerns, especially regrading the fit to the SI.

I don’t think that the resulting inventory can be regarded as a benchmark dataset for machine learning. First, the inventory is the result of a modelling procedure (by spectral thresholding) and hence does not present reference data with “ground truth quality”, which however would be required for agood  benchmark dataset to be useful for algorithm and model comparison. Second, the dataset is limited to the labeled locations of lakes and channels but no predictor data are given so that the data cannot be used to test different machine learning approaches without extensive further data acquisition.

Further, I have concerns about the thresholding approach used to delinete water pixels. The authors applied a series of thresholds (see Fig. 2 and 3) to the spectral channels but it is unclear to me how these thresholds were derived. Is that based on try and error, and if so, how was the error assessed (visual? statistically by comparing to the manually digitized reference data?). I also wonder why the authors did not use a supervised classification algorithm but decided for a series of manually selected thresholds instead.

The validation is not complete. If I get it correctly, the accuracy/sensitivity/specificity values that are given in the manuscript refer to the water delineation. What is missing is the validation for the classification into channels and lakes.

The authors present a methodology to create a benchmark dataset of supraglacial hydrology across the Antarctic. Their approach takes advantage of the large amount of multi-spectral satellite imagery provided by Landsat-8 and Sentinel-2. They propose a pipeline consisting of a cascade of spectral thresholding for the creation of a binary mask and subsequent K-means clustering to further distinguish lakes and channels.

The manuscript is clear and the exposition is easy to follow and linear.

I have only one specific comment related to the validation approach that was used, particularly thinking about future use by researchers in the area of machine learning.

As stated by the authors already in the abstract, the dataset provides a partition of the superglacial features into non-water, lakes and streams. However, from my understanding in Table 1 the authors are showing aggregated results (binary water/non-water) for the accuracy and sensitivity metrics. I think that for the future users to benefit the most from the proposed inventory, it would be beneficial to show a validation considering all the classes.