Calving front monitoring at sub-seasonal resolution: a deep learning application to Greenland glaciers
Abstract. The mass balance of the Greenland ice sheet is strongly influenced by the dynamics of its outlet glaciers. Therefore, it is of paramount importance to accurately and continuously monitor these glaciers, especially the variation of their frontal positions. A temporally comprehensive parameterization of glacier calving is essential to understand dynamic changes and to constrain ice sheet modelling. However, current calving front records are often limited in temporal resolution as they rely on manual delineation, which is laborious and not feasible with the increasing amount of satellite imagery available. In this contribution, we address this problem by applying an automated method to extract calving fronts from optical satellite imagery. The core of this workflow builds on recent advances in the field of deep learning while taking full advantage of multispectral input information. The performance of the method is evaluated using three independent validation datasets. Eventually, we apply the technique to Landsat-8 imagery. We generate 9243 calving front positions across 23 Greenland outlet glaciers from 2013 to 2021. Resulting time series resolve not only long-term and seasonal signals but also sub-seasonal patterns. We discuss the implications for glaciological studies and present a first application analysing the interaction between calving front variation and bedrock topography. Our method and derived results represent an important step towards the development of intelligent processing strategies for glacier monitoring, opening up new possibilities for studying and modelling the dynamics of Greenland outlet glaciers. Thus, these also contribute to advance the construction of a digital twin of the Greenland ice sheet, which will improve our understanding of its evolution and role within the Earth's climate system.
Erik Loebel et al.
Status: open (until 29 Jun 2023)
- RC1: 'Comment on tc-2023-52', Anonymous Referee #1, 01 Jun 2023 reply
Erik Loebel et al.
Erik Loebel et al.
Viewed (geographical distribution)
Review of Loebel et al., The Cryosphere
The authors present an exciting deep learning method for tracing glacier calving fronts in Landsat 8 and 9 images. The manuscript presents the method based on a specialized Artificial Nueral Network, the resulting dataset of 9243 calving front traces for 23 of Greenland’s outlet glaciers, and an example of how the data may be useful for examining glacier dynamics. The method produces calving fronts that are on average within 80 meters of manually traced calving fronts, which is less than the uncertainty of manual calving front delineations according to a study by Goliber et al. (2022).
The authors thoughtfully developed the deep learning method. They considered different illumination conditions and terminus morphologies when training the model. There is good documentation of the time and storage requirements for training the model. I applaud the contribution to open-source code and datasets, which are valuable to the glaciological community. The 698 manual delineations used for training the model would also be valuable to the community and I recommend submitting them to a new or existing data repository.
Overall, the manuscript is well-presented and concisely written. The figures are particularly well-constructed and compelling. However, I think the main text currently lacks detail on the deep learning method. I think the information included in Appendix A and B should be included in the main manuscript since it is relevant to understanding how the ANN algorithm was developed.
Spatial transferability of the method is mentioned throughout the manuscript and described as an advantage to using this method compared to other existing automated calving front tracing methods. The deep learning model is tested on glaciers from regions outside of Greenland (e.g., Antarctica, Svalbard, and Patagonia). I would be really interested in formal discussion of how the method, trained on Greenland’s outlet glaciers, performed with the glaciers in other regions specifically. How does the accuracy calculated for those test glaciers compare to the accuracy of the Greenland test glaciers? Discussing this would provide appropriate support for the spatial transferability of the method.
In general, this manuscript presents a valuable contribution to the field and I would like to see this work published after these more major comments and the minor comments listed below are addressed.
In general, proof read for compound adjectives that need to be hyphenated, e.g., Greenland-wide (L176).
L10: You should include a statement about the accuracy of your method that you calculated here.
L14-15: The phrase “digital twin” of Greenland ice sheet is not clearly defined. Unnecessary in abstract unless explained in more detail. It’s not discussed throughout the paper so I don’t think it’s appropriate to include here or in the conclusion without further elaboration.
L52-55: This is not the first automated method that captured sub-seasonal resolution time series of calving front change (see Liu et al., 2021). Reframe the language here.
L69: What is the fixed window size and how was it chosen?
L72: “Built” instead of “build”
L86-100: This section discussing the method performance should be moved to the Results or Discussion section.
L106: Elaborate on how the completely clouded Landsat scenes are filtered.
L136: Include citations for how glacier geometry impacts terminus retreat. At the very least, Felikson et al., 2020 (https://doi.org/10.1029/2020GL090112) should be cited here since it directly discusses the impact of bed topography on glacier retreat.
L140: Looks more like 2016 and 2017, not 2018 showed the rapid retreat for Ingia Isbræ.
L164-166: Is it that the algorithm performs better at overcoming challenging cloud, illumination, and mélange issues than manual delineations? The way this sentence is currently structured implies that. I think this sentence could be removed altogether since the sentence that follows already emphasizes the high temporal resolution of the time series.
Figures and Tables
Fig. 2. In the caption, write out TU Dresden or just refer to it as the testing dataset for this study. I think it’s fine to exclude the testing glaciers from other regions. Adding a location in parentheses after each of the excluded glaciers would make it more clear why they aren’t included in this map. E.g., Drygalski Glacier (Antarctica), Storbreen Glacier (Svalbard), etc.
Fig. 5. I recommend adding a colorbar for the green shading.
Fig. B1. This figure could remain in the Appendix or Supplementary Material even if the description of methodology in Appendix B is moved to main text.
Table C1. Is the right side really a confusion matrix if only done for TUD? Listing the fraction/percentage of total pixels would be more meaningful here than the raw pixel numbers. As of now, I draw much more from the mean and median errors listed on the left side than the Confusion Matrix. Consider separating the Confusion Matrix portion of this table into its own table. Clearly define TP, TN, FP, FN in the caption.