the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Outlet Glacier Seasonal Terminus Prediction Using Interpretable Machine Learning
Abstract. Glacier terminus retreat involves complex processes superimposed at the interface between the ice sheet, the ocean, and the subglacial substrate, posing challenges for accurate physical modeling of terminus change. To enhance our understanding of outlet glacier ablation, numerous studies have focused on investigating terminus position changes on a seasonal scale with no clear control on seasonal terminus change that has been identified across all glaciers. Here, we explore the potential of machine learning to analyze glaciological time series data to gain insight into the seasonal changes of outlet glacier termini. Using machine learning models, we forecast seasonal changes in terminus positions for 46 outlet glaciers in Greenland. Through the SHapley Additive exPlanations (SHAP) feature importance analysis, we identify the dominant predictors of seasonal terminus position change for each. We find that glacier geometry is important for accurate predictions of the magnitude of terminus seasonality and that environmental variables (mélange, ocean thermal forcing, runoff, and air temperature) are important for determining the onset of seasonal terminus change. Our work highlights the utility of machine learning in understanding and forecasting glacier behavior.
- Preprint
(7635 KB) - Metadata XML
-
Supplement
(8521 KB) - BibTeX
- EndNote
Status: open (until 26 Sep 2025)
-
RC1: 'Comment on egusphere-2025-3483', Anonymous Referee #1, 29 Aug 2025
reply
Summary
This study applies machine learning to investigate the drivers of seasonal terminus changes in Greenland outlet glaciers. The analysis uses freely available datasets on glacier geometry, climate, and glacier dynamics for the period 2000–2020. The authors find that geometric variables are the strongest controls on the magnitude of seasonal change, and that their approach performs best for glaciers with limited long-term retreat.
The study addresses a highly relevant topic for the glaciological community and represents a valuable step toward understanding and predicting glacier calving front changes. The manuscript is generally well written and engaging. However, I believe several issues regarding data and data preparation must be addressed to strengthen the paper. Below I provide general comments, followed by specific points.
General points
Data preparation:
Before your timeseries of data enter the machine learning setup, you remove the multi-annual component from the data leaving the seasonal component as input. You do this for every input variable (L153 - 154). Does this detrending apply to geometric variables such as bed slope and thickness, and to dynamic variables such as strain rate? If yes, what is the physical interpretation of removing the long-term component of bed slope or strain rate near the calving front (which changes position)? For instance, the velocity pattern and thus the strain rates at a given location are highly dependent on the geometry of the glacier (e.g. is it in a narrow or wider spot). Is it actually necessary to detrend the geometric data? Perhaps you can show the reader examples of the full and detrended time series of input variables like you have in Figure 2 for terminus advance. Also adding some more details of the method and on the resulting seasonal component would be useful to the reader, like does the timeseries have a zero mean?
Because glaciers are not in steady state, more clarification is also needed on how the “magnitude of seasonal terminus change” is defined. Since termini do not necessarily return to the same position at the end of summer, how is this accounted for? In other words, do the long term changes still enter your prepared data through this? Are these situations where the algorithm struggles to make accurate predictions (L394)? Similarly, in Figure 2b some years show double peaks (e.g., 2013) —how are such cases treated?
Mélange proxy:
I am not convinced that SST is a reliable proxy for mélange conditions. SST does not necessarily capture mélange strength or thickness, as you also note (L406). I understand the motivation, given the lack of a Greenland-wide mélange product, but perhaps it would be more accurate to label this variable “fjord surface temperature,” acknowledging that it reflects mixed influences from open water, fjord ice, and mélange. Alternatively, additional examples establishing its validity as a mélange proxy would strengthen the argument.
Key findings:
In the abstract you write ‘We find that glacier geometry is important for accurate predictions of the magnitude of terminus seasonality and that environmental variables (mélange, ocean thermal forcing, runoff, and air temperature) are important for determining the onset of seasonal terminus change.’ I think these are the most important findings in your study, although perhaps not surprising? However, this is not really emphasized in your conclusions and outlook, where the focus is mostly on the potential of machine learning methods to improve our understanding e.g. seasonal terminus changes from a data driven perspective and how they can be implemented with numerical ice flow models. While these are important points, they are more general/outlook, more weight should be given to the glaciological insights gained—specifically, what your study reveals about the processes governing seasonal terminus changes.
Specific comments/suggestions
L125: Should be Danish Meteorological Institute (DMI) not Danish Meteorological Society
L125: Figures 1 and 2 in Jensen et al. (2023) show that the station network is densest in South and Southwest Greenland. It is often very far from the glaciers in your study to the nearest weather station and the observations might not be representative of a particular glacier even when choosing the nearest one. For instance, the weather station is located much closer to the open ocean than the glacier terminus in question. While direct observations are great, would it not be better to apply re-analysis data from e.g. CARRA or similar?
L131: Just to clarify, is the timeseries of ocean thermal forcing near the glacier fronts for the entire column of water or only at a specific depth or interval?
L162: ‘…heterogeneity in this population (Table 4).’ Should be Table 3, I think.
L200: When evaluating the NRMSE, is it done once pr year or at every minimum that you evaluate the amplitude of the seasonal cycle or is it evaluated at every timestep?
L328 Perhaps change the subsection from ‘Feature Importance for Seasonal Terminus Prediction’ to ‘Feature Importance for Magnitude of Seasonal Terminus Prediction’?
L319: You have a mean temporal offset of 6.7 weeks between the predicted and observed peak. This is ~1.5 months so quite a large difference in timing. Is it a systematic bias or is it random?
L357 forward: The studies in L357-359 you cite for findings on mélange impact on timing of retreat of the while you’re the focus of your study is the magnitude and not the timing (L338-340) of the seasonality. I am just wondering if you are comparing the right things.
L394: You write that the best performing models are for glaciers that have experienced little overall retreat during the period. Does this mean, that geometric parameters are the main controls for outlet glaciers close to a steady state?
L455: suggests -> shows
Figure 2: A Figure 2c is mentioned in the caption but is not present. Are both trends now in Figure 2b?
Citation: https://doi.org/10.5194/egusphere-2025-3483-RC1
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
571 | 35 | 5 | 611 | 12 | 22 | 21 |
- HTML: 571
- PDF: 35
- XML: 5
- Total: 611
- Supplement: 12
- BibTeX: 22
- EndNote: 21
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1