Automatic Supraglacial Lake Extraction in Greenland Using Sentinel-1 SAR Images and Attention-Based U-Net

Round 1

Reviewer 1 Report

This manuscript is presenting a study on a very interesting topic, the detection of supraglacial lakes on the Greenland Ice Sheet by means of radar imagery employing an AI application. This extends previous attempts to automatically detect supraglacial lakes using Bayesian classifiers and is advancing the research field substantially, but it lacks currently several aspects, that are discussed below.  

One major topic that needs to be extended is how well the classification is dealing with discriminating between former lake ground and lake ice. While lake ice area is accounted (well justified) as lake area, the former lake ground is not lake area. This is quite tricky, but can change the area quite a bit. The approach may able to deal with that well, but it is not presented in the manuscript. To make clear what I refer to, I try to explain it a bit better: the glacier is moving with a speed of a few hundred meters per year, while the topographic sink remains always a sink. This way the glacier surface that was formerly a lake ground is moving further downstream and eventually being again glacier surface. Due to the melt going on at the base of the lake, this ice is different from ‘normal’ glacier surface, but also different from lake ice. Only after 8-10m of snow are accumulated or the surface has undergone more melt, this is no longer well visible in SAR imagery. Fig 8b shows lakes where this feature is prominent, but on the scale presented one cannot assess it well.

I have concerns that rivers between supraglacial lakes are falsely detected as lake bodies. As floating tongues and ice shelves also exhibit lakes that are rather long and narrow, rivers and such features may be mixed up easily. This will have strong effect on the time series and trends of lake area. This is well visible in e.g. Fig 8j, 10.

The trickiest time period to track with Sentinel-1 data supraglacial lakes is in the peak melt season. The false classification in this time period is not presented here. For example Fig. 8 (no dates given) even presents snow covered frozen surface at KNS in Sentinel-2, which of course cannot be used for checking how well the DNN is doing in peak melt season. Here, a comprehensive chapter is needed.  

Ice marginal lakes have been falsely detected as supraglacial lakes. This needs to be changed.

Penetration depth of C-band in pure ice is not 0.7m (what is this value based on??) but several metres! Rignot, E., Echelmeyer, K. and Krabill, W. (2001), Penetration depth of interferometric synthetic-aperture radar signals in snow and ice. Geophys. Res. Lett., 28: 3501-3504. https://doi.org/10.1029/2000GL012484

Monthly average temperature values are not very useful for comparison, rather present mean daily or maximum daily temperature.  

Why do the authors expect the lakes in Antarctica to be different from lakes on the Greenland Ice Sheet? This remains unclear.

What do the authors mean by ‘instantaneous hydrological features such as supraglacial rivers and streams’? This seems to be a misconception.

The west coast of Greenland contains the so called dark ice zone. There the surface is considerable different from other ‘clean’ locations such as the ones chosen for this study. I highly recommend to include a chapter on the capability of the network to detect there supraglacial lakes. Actually, the reader would learn by far more from that then from any temperature trend.

Avoid giving trends for five years of time series – you do not even cover the entire Sentinel-1 time series in you manuscript and five years are not a robust data basis for trends.

Minor:

Check in the entire text that names, such as ** Ice Shelf and ** Glacier are written with capital letters for Ice Shelf or Glacier respectively. Check in the entire manuscript, that Sentinel-1 is correctly written.

Line 113 typo snowSentineltine‐

Table 1: It is confusing that there should be only one date for the test dataset per area. I would have expected an entire season, year etc..

Fig. 8 needs to be split in different figures. Acquisition dates of the imagery needs to be given.

Fig. 10: Can you please spread you colour scale such that the glacier area presents some more color range.

Fig 11: align the panels for 79NG and JI next to each other. Plot a zero degree line into the temperature plot.

Fig 12: this figure is hard to understand. Can the authors please put an optical or SAR image as background and plot lakes that are present in July and August in for example purple. The elevation contours need to be smoothed. Sure you need coordinates in seconds when you only present ticks in full degrees?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes an automatic method using the attention‐based U‐Net and Sentinel‐1 SAR data to map the supraglacial lakes. Compared with optical remote sensing, SAR has more advantages in all‐day, all‐weather detection of supraglacial lakes, identification of covered lakes, and long‐term sequence analysis. The proposed method gets a higher validation coefficient with an F1 score of 0.971 and was applied to lake extraction in Antarctica's GeorgeVl ice shelf region. However, the paper does not explain clearly the reason for the U-Net improvements. Are these improvements linked to the difficulties of supraglacial lake extraction? It also lacks comparison with other glacial lake extracting modified U-Net. Apart from this, there are not enough arguments to prove that the method is indeed well applied in the GeorgeVl area of Antarctica. This paper is recommended for minor revision.

 Major comments

1) Models running on the GPU cannot be considered an innovation. Abstract (line 13), Section 4.1.1 (line 576) and Section3.1.(line 480) emphasize the advantages of running the model on the GPU compared to running it on the CPU, but running deep learning models on GPU is already a widespread application. If you want to emphasize the GPU, you can compare the performance of different GPUs and whether the one used in this paper has improved efficiency.

2) The overview of the method of extracting the supraglacial lake is written too briefly. The introduction section provides a detailed overview of the difficulties of supraglacial lake extraction, what can be solved by traditional methods, what cannot be solved by traditional methods but can be solved by deep learning, and what are the advantages of deep learning over traditional methods. Also, some studies such as Wu et al. and Qayyum et al. have used modified U-Net to extract glacial lakes in the Himalayan region, can their models be transferred to Antarctica for use, and do your modified U-Net models have advantages over those modified U-Net?

3) The connection between proposed method and supraglacial lakes’ characteristics is not clearly written. Section 2.3.2. describes in detail the characteristics of supraglacial lakes on SAR images. But the reader has no way of knowing how these characteristics are related to the proposed method design.

4) No mention of the reason for using the BCE loss function. Wu's study says that the BCE loss function is inapplicable to data with extremely unbalanced positive and negative samples. So they changed to the Tversky function in the glacial lake extraction. Have you not encountered this problem when using the BCE loss function?

5) There were six test sites, but the results were only analyzed in two. Section 3.2. only analysis Jakobshavn Glacier and 79N Glacier, the analysis of the other four test sites should also be added.

Specific comments

Line138: km2 revise to km².

Line483: Figure 8p does not adequately support the conclusion that the proposed method has been better applied to migration in the Antarctica's GeorgeVl ice shelf region.

Line558: 4.1.1 revise to 4.1., corresponding to the format of section 3.

Line615: 4.1.2 revise to 4.2., corresponding to the format of section 3.

Figure 1: delineate all the supraglacial lake extent with red lines.

Figure 2: ①* not show on the figure 2, suggest use a separate color to represent sites that both train and test; ②It lacks of legend, scale and latitude/longitude; ③Application area(Antarctica's GeorgeVl ice shelf) should be marked out.

Figure 3: Part ”ground truth labelsets making” ‘s arrow is reversed.

Figure 8: It is recommended that the extraction result and the original image be displayed on one image and that only the boundaries of the extraction result be displayed. This will allow the reader to see more intuitively how the extraction result corresponds to the original image.

Figure 9: The same advice as given for Figure 8. And the results extracting by different methods could be displayed using different boundary color.

Figure 10: The same advice as given for Figure 8.

Table 1.: ①ID S1-1, S1-2, S1-3… revise to S1-01, S1-02, S1-03…, for alignment; ②Uniform font size.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have addressed all my comments, and I have no more comments and agree with the publication of this manuscript.