the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Future scenarios of albedo and radiative forcing resulting from changes in snow depth in Austria
Abstract. The presence of a seasonal snowpack decisively modulates the albedo of terrestrial land surfaces. Global warming-driven decreases in the duration of the seasonal snow cover are thus expected to lower annual albedo, result in a positive radiative forcing and thus represent a positive feedback to climate change. Here we quantify future (up to the year 2100) scenarios of albedo change and the associated radiative forcing for Austria using a machine learning approach which leverages satellite-derived albedo data and future scenarios of snow depth. Albedo was calculated from the MODIS MCD43A1 v006 BRDF/albedo product for the period 2002–2019. Snow depth was taken from a novel dataset for Austria (FuSE-AT) covering the period 1951–2100. A machine-learning model (using LightGBM) was then trained to predict albedo separately for each land cover type (MODIS MCD12Q1 v006) using snow depth, days since last snowfall, as well as several predictors related to plant canopy structure (leaf area index, canopy height) and topography (latitude, longitude, elevation above sea level, inclination, exposition). LAI turned out to be an important predictor in simulating albedo in both snow free and snow covered points in time. Time since last snowfall, as a surrogate for snow aging, was more important for short land cover types than for forests. The correlation coefficients of the trained models varied widely across the different land cover types, ranging from 0.70 to 0.94. In 5 out of the 6 scenarios used, a significant decline of albedo could be observed. The cumulative time-dependent emission equivalent resulting from the albedo changes between 2020–2100 corresponds to 0.25–1 (RCP 2.6), 0.8–2.25 (RCP 4.5) or 1–5 (RCP 8.5) times the annual CO2-equivalent emissions of Austria for the year 2021.
- Preprint
(1792 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on egusphere-2024-881', Anonymous Referee #1, 14 May 2024
Overview:
Kiem and co-authors present an analysis of remote sensing and climate model-based changes in albedo and radiative forcing, as it interacts with changes in snow depth and land cover type. The approach appears to use high spatial resolution (1000 m) snow depth data from a regional climate model, MODIS albedo data, GEDI lidar data (for canopy height), BESS radiative forcing data together with a machine learning model (LightGBM) to project future changes in albedo and associated radiative forcing and convert to an emissions equivalent. The authors conclude that albedo significantly declines in “5 out of the six scenarios used”, though it looks like only three scenarios were used (RCP2.6, RCP4.5 and RCP8.5).
I provide some major comments and suggestions below and encourage the authors to resubmit when they have been addressed.
Major Comments:
The introduction is rather short and thin on references that would help lay the groundwork for justifying the study and identifying what we already know about future declines in snow, albedo, and changes in land cover. I recommend the authors dive deeper into the literature to include a review of papers that look at changes in snow loss and albedo change with land cover. Several of these show up in the discussion. I have also provided some at the end of the review as a starting point.
The methods section would benefit greatly from a workflow diagram and clear description of what the LightGBM machine learning model is doing with the model snow data and remote sensing data. Currently, it is a bit of a black box. The description of the FuSE-AT dataset is also lacking. Is this a regional climate model? Statistically downscaled or dynamically downscaled? Were the historical climate model runs completed using reanalysis? Or did the authors use the CMIP6 historical runs? Each of the model runs selected come from a different variant ids (r1i1p1 and r12i1p1 - unless that’s a typo in the text and figure caption?), differing in the initial conditions. The upsampling technique for the remote sensing datasets is also not described, nor are any of the remote sensing datasets described beyond that they were downloaded as netCDF and their spatial resolution. It’s difficult to review the significance of the machine learning results and discussion of said results with so little information on the methods, data and justification for the approach.
Lastly, the significance of the study is unclear. The land surface models of most GCMs include process-based modeling of the variables deemed important in the machine learning model (LAI/vegetation height, snow aging). How would machine learning provide an advantage over processed-based land surface models that account for canopy interception and shading, radiative transfer, and full energy balance?
Minor Comments:
Please define all abbreviations. Many are lacking throughout the manuscript.
Lines 72-73: Please include more recent references, such as Sabatini et al. (2021) and Ceccherini et al. (2020).
Sabatini et al. (2021): https://onlinelibrary.wiley.com/doi/full/10.1111/ddi.12778
Ceccherini et al. (2020)
https://www.nature.com/articles/s41586-020-2438-y
Lines 93-99: Please confirm and specify that two-way coupling is used between the land and atmosphere models.
Line 109: Remove unnecessary comma after CCLM-VEG3D
Figures:
Figure 2. It’s not clear why this was included in the methods section. Please move to results, or better, the appendix.
Figure 4a. What is the reference data set used for the Taylor diagram? The figure caption says “reference standard deviation”, which doesn’t make much sense.
Figure 4b. Predicted from the machine learning model vs. MODIS? Be specific about what is considered “real” and “predicted”
Additional references:
Jones et al. 2015. Accounting for radiative forcing from albedo change in future global land-use scenarios. https://link.springer.com/article/10.1007/s10584-015-1411-5
Kvalevåg et al. 2009. Anthropogenic land cover changes in a GCM with surface albedo changes based on MODIS data. https://doi.org/10.1002/joc.2012
The Land Use and Climate Across Scales (LUCAS) project publications (Mooney et al. 2022; Daloz et al. 2022; Davin et a. 2020 come to mind):
https://ms.hereon.de/cordex_fps_lucas/079634/index.php.en
Citation: https://doi.org/10.5194/egusphere-2024-881-RC1 -
RC2: 'Comment on egusphere-2024-881', Anonymous Referee #2, 20 May 2024
General comments:
This is an interesting study modelling the future evolution of albedo in Austria in response to changes in snow depth. While most studies have focused on the influence of snow cover area and duration on albedo, it is true that fewer have focused on snow depth only as a driver of albedo change. The authors use a novel dataset of future snow depth scenarios for Austria and train a machine learning model to predict albedo change. The authors use state of the art remote sensing data for snow cover to train their model and evaluate their predictions for the historical period 2002-2019. The main findings predict a decline albedo over Austria in response to reduced future snow depths. Interestingly, the authors put into perspective the decline in albedo by comparing it with the equivalent radiative forcing from CO2 emissions in Austria, showing that a small decline in albedo may be equivalent to 0.5-6 times the annual CO2 emissions from Austria. I believe the study findings and methodology are interesting, valid, and fall within the scope of the journal The Cryosphere. However, the authors need to improve/clarify certain aspects of the manuscript, particularly some of the figures and discussions associated with the figures, as well as more details on the model they use, before the manuscript is published.
Specific comments:
- Even though I found the evaluation of the model sound for what the authors are using it, which is mean albedo conditions, I am a bit sceptical about some of the model results. For instance, in Figure 3, top row, albedos are quite low (max 0.58) even for 2 metre of snow depth and 0 days since last snowfall. Considering that fresh snow albedos can be as high as 0.9, and that grasslands should be fully snow covered even with low snow depth, how is this explained? Is this because of grid/spatial averaging? Or is it a flaw of the model? In Barren and Ice (Fig A3) albedos may reach monthly means of up to 0.7, which is higher than what Figure 3 shows. Please explain/clarify this in the manuscript where you discuss Figure 3.
- Still in Figure 3: For forests, I understand that the snow may be under the canopy and therefore have a lower influence on albedo. However, for 0 days since last snowfall, I expect a higher albedo as there may still be considerable snow on the canopy. However, the increase in albedo is not significant at all. I guess the data/model does not distinguish between above or below canopy snow? Please explain/discuss this in the manuscript when you discuss Figure 3.
- Another main point to address is that the authors say the novelty of the study is characterising the influence of snow depth on albedo, as opposed to the influence of snow presence/absence. However, I wonder whether snow presence/absence is already an implicit predictor variable within snow depth. If snow depth is 0, that implies snow absence, so how important is this compared to snow depth going above 0? This should be discussed both when discussing snow depth as a predictor variable, and in the results/discussion.
- I think the primary objectives of the study in lines 53-56 have a strange order. I suggest the following order: 3-4-1-2. At least this makes more sense with the manuscript structure.
- Please provide more information about the model workflow. I know it is a machine learning model and therefore a bit of a black box, but right now the only information about the modelling and the workflow is referring to the original model paper. Please provide more details about how the model is set up.
- Please explain what the L2-loss metric is. (Line 111).
- Please explain what “combined feature importance” and “split feature importance” means, as this is not defined and may be a machine learning specific term? Does this refer to a combined land cover type model and the separate land cover models? Please discuss this in the paragraph starting line 155. Similarly, explain what it means that the model is making decisions (splitting) in line 158, as it is unclear. This would make Figure 4 clearer.
- The taylor diagram is not clearly explain in my opinion. Please describe in more detail what this figure shows and how it shows it.
- I am not sure the Figure 4b is the most useful in terms of validation, as it shows the yearly average. I actually found Figure A3 and A4 much more valuable for the model evaluation. Looking at Figure A3 and A4 made me trust the model findings more than Figure 4. I suggest taking Figure A3 and bringing it to the main part of the manuscript, as this is a great evaluation Figure per land cover type and per month.
- Figure 6: please explain why the cumulative sum is fluctuating so much. Is the TDEE negative in some hears and reduces the cumulative sum? Why so?
- Line 188: Why during winter months the modelled albedo is worse for forests and croplands? Please discuss further. Also discuss the influence of canopy (snow on canopy) in your results.
- Line 207: their emission-equivalent MAY cumulate to years worth of…
- Line 214-216: I would remove this sentence. I am not sure the main discussion of a 2024 study like this one should include a recommendation to use a biogeoengineering technique that was published in 2009 to affect global-scale albedo values. Instead, factors such as greening of mountains should be discussed (e.g. https://iopscience.iop.org/article/10.1088/1748-9326/aa84bd/meta)
Technical corrections:
Line 9-10: I think it should be slope and aspect instead of inclination and exposition.
Line 19: affect à affecting
Line 25: I think “dust” should be included as influencing albedo.
Line 29: “Many” instead of “a lot”
Figure 1: add Latitude (⁰N) and Longitude (⁰E) on the map axes.
Figure 2: Yearly mean SD of Austria for the available FuSE-AT datasets (2000-2100) and for 3 RCP scenarios.
Figure 3: Here you can type out the land cover types in the subpanel titles. Also add the subpanels numbering (a-d), and modify the text accordingly.
Line 114-115: among others the Nash-Sutcliffe efficiency (NSE) (Nash and 115 Sutcliffe, 1970) and the Kling-Gupta efficiency (KGE) (Gupta et al., 2009) of what? Daily albedo?
Figure A3-A4: If there is space, I would type out the land cover types.
Citation: https://doi.org/10.5194/egusphere-2024-881-RC2
Status: closed
-
RC1: 'Comment on egusphere-2024-881', Anonymous Referee #1, 14 May 2024
Overview:
Kiem and co-authors present an analysis of remote sensing and climate model-based changes in albedo and radiative forcing, as it interacts with changes in snow depth and land cover type. The approach appears to use high spatial resolution (1000 m) snow depth data from a regional climate model, MODIS albedo data, GEDI lidar data (for canopy height), BESS radiative forcing data together with a machine learning model (LightGBM) to project future changes in albedo and associated radiative forcing and convert to an emissions equivalent. The authors conclude that albedo significantly declines in “5 out of the six scenarios used”, though it looks like only three scenarios were used (RCP2.6, RCP4.5 and RCP8.5).
I provide some major comments and suggestions below and encourage the authors to resubmit when they have been addressed.
Major Comments:
The introduction is rather short and thin on references that would help lay the groundwork for justifying the study and identifying what we already know about future declines in snow, albedo, and changes in land cover. I recommend the authors dive deeper into the literature to include a review of papers that look at changes in snow loss and albedo change with land cover. Several of these show up in the discussion. I have also provided some at the end of the review as a starting point.
The methods section would benefit greatly from a workflow diagram and clear description of what the LightGBM machine learning model is doing with the model snow data and remote sensing data. Currently, it is a bit of a black box. The description of the FuSE-AT dataset is also lacking. Is this a regional climate model? Statistically downscaled or dynamically downscaled? Were the historical climate model runs completed using reanalysis? Or did the authors use the CMIP6 historical runs? Each of the model runs selected come from a different variant ids (r1i1p1 and r12i1p1 - unless that’s a typo in the text and figure caption?), differing in the initial conditions. The upsampling technique for the remote sensing datasets is also not described, nor are any of the remote sensing datasets described beyond that they were downloaded as netCDF and their spatial resolution. It’s difficult to review the significance of the machine learning results and discussion of said results with so little information on the methods, data and justification for the approach.
Lastly, the significance of the study is unclear. The land surface models of most GCMs include process-based modeling of the variables deemed important in the machine learning model (LAI/vegetation height, snow aging). How would machine learning provide an advantage over processed-based land surface models that account for canopy interception and shading, radiative transfer, and full energy balance?
Minor Comments:
Please define all abbreviations. Many are lacking throughout the manuscript.
Lines 72-73: Please include more recent references, such as Sabatini et al. (2021) and Ceccherini et al. (2020).
Sabatini et al. (2021): https://onlinelibrary.wiley.com/doi/full/10.1111/ddi.12778
Ceccherini et al. (2020)
https://www.nature.com/articles/s41586-020-2438-y
Lines 93-99: Please confirm and specify that two-way coupling is used between the land and atmosphere models.
Line 109: Remove unnecessary comma after CCLM-VEG3D
Figures:
Figure 2. It’s not clear why this was included in the methods section. Please move to results, or better, the appendix.
Figure 4a. What is the reference data set used for the Taylor diagram? The figure caption says “reference standard deviation”, which doesn’t make much sense.
Figure 4b. Predicted from the machine learning model vs. MODIS? Be specific about what is considered “real” and “predicted”
Additional references:
Jones et al. 2015. Accounting for radiative forcing from albedo change in future global land-use scenarios. https://link.springer.com/article/10.1007/s10584-015-1411-5
Kvalevåg et al. 2009. Anthropogenic land cover changes in a GCM with surface albedo changes based on MODIS data. https://doi.org/10.1002/joc.2012
The Land Use and Climate Across Scales (LUCAS) project publications (Mooney et al. 2022; Daloz et al. 2022; Davin et a. 2020 come to mind):
https://ms.hereon.de/cordex_fps_lucas/079634/index.php.en
Citation: https://doi.org/10.5194/egusphere-2024-881-RC1 -
RC2: 'Comment on egusphere-2024-881', Anonymous Referee #2, 20 May 2024
General comments:
This is an interesting study modelling the future evolution of albedo in Austria in response to changes in snow depth. While most studies have focused on the influence of snow cover area and duration on albedo, it is true that fewer have focused on snow depth only as a driver of albedo change. The authors use a novel dataset of future snow depth scenarios for Austria and train a machine learning model to predict albedo change. The authors use state of the art remote sensing data for snow cover to train their model and evaluate their predictions for the historical period 2002-2019. The main findings predict a decline albedo over Austria in response to reduced future snow depths. Interestingly, the authors put into perspective the decline in albedo by comparing it with the equivalent radiative forcing from CO2 emissions in Austria, showing that a small decline in albedo may be equivalent to 0.5-6 times the annual CO2 emissions from Austria. I believe the study findings and methodology are interesting, valid, and fall within the scope of the journal The Cryosphere. However, the authors need to improve/clarify certain aspects of the manuscript, particularly some of the figures and discussions associated with the figures, as well as more details on the model they use, before the manuscript is published.
Specific comments:
- Even though I found the evaluation of the model sound for what the authors are using it, which is mean albedo conditions, I am a bit sceptical about some of the model results. For instance, in Figure 3, top row, albedos are quite low (max 0.58) even for 2 metre of snow depth and 0 days since last snowfall. Considering that fresh snow albedos can be as high as 0.9, and that grasslands should be fully snow covered even with low snow depth, how is this explained? Is this because of grid/spatial averaging? Or is it a flaw of the model? In Barren and Ice (Fig A3) albedos may reach monthly means of up to 0.7, which is higher than what Figure 3 shows. Please explain/clarify this in the manuscript where you discuss Figure 3.
- Still in Figure 3: For forests, I understand that the snow may be under the canopy and therefore have a lower influence on albedo. However, for 0 days since last snowfall, I expect a higher albedo as there may still be considerable snow on the canopy. However, the increase in albedo is not significant at all. I guess the data/model does not distinguish between above or below canopy snow? Please explain/discuss this in the manuscript when you discuss Figure 3.
- Another main point to address is that the authors say the novelty of the study is characterising the influence of snow depth on albedo, as opposed to the influence of snow presence/absence. However, I wonder whether snow presence/absence is already an implicit predictor variable within snow depth. If snow depth is 0, that implies snow absence, so how important is this compared to snow depth going above 0? This should be discussed both when discussing snow depth as a predictor variable, and in the results/discussion.
- I think the primary objectives of the study in lines 53-56 have a strange order. I suggest the following order: 3-4-1-2. At least this makes more sense with the manuscript structure.
- Please provide more information about the model workflow. I know it is a machine learning model and therefore a bit of a black box, but right now the only information about the modelling and the workflow is referring to the original model paper. Please provide more details about how the model is set up.
- Please explain what the L2-loss metric is. (Line 111).
- Please explain what “combined feature importance” and “split feature importance” means, as this is not defined and may be a machine learning specific term? Does this refer to a combined land cover type model and the separate land cover models? Please discuss this in the paragraph starting line 155. Similarly, explain what it means that the model is making decisions (splitting) in line 158, as it is unclear. This would make Figure 4 clearer.
- The taylor diagram is not clearly explain in my opinion. Please describe in more detail what this figure shows and how it shows it.
- I am not sure the Figure 4b is the most useful in terms of validation, as it shows the yearly average. I actually found Figure A3 and A4 much more valuable for the model evaluation. Looking at Figure A3 and A4 made me trust the model findings more than Figure 4. I suggest taking Figure A3 and bringing it to the main part of the manuscript, as this is a great evaluation Figure per land cover type and per month.
- Figure 6: please explain why the cumulative sum is fluctuating so much. Is the TDEE negative in some hears and reduces the cumulative sum? Why so?
- Line 188: Why during winter months the modelled albedo is worse for forests and croplands? Please discuss further. Also discuss the influence of canopy (snow on canopy) in your results.
- Line 207: their emission-equivalent MAY cumulate to years worth of…
- Line 214-216: I would remove this sentence. I am not sure the main discussion of a 2024 study like this one should include a recommendation to use a biogeoengineering technique that was published in 2009 to affect global-scale albedo values. Instead, factors such as greening of mountains should be discussed (e.g. https://iopscience.iop.org/article/10.1088/1748-9326/aa84bd/meta)
Technical corrections:
Line 9-10: I think it should be slope and aspect instead of inclination and exposition.
Line 19: affect à affecting
Line 25: I think “dust” should be included as influencing albedo.
Line 29: “Many” instead of “a lot”
Figure 1: add Latitude (⁰N) and Longitude (⁰E) on the map axes.
Figure 2: Yearly mean SD of Austria for the available FuSE-AT datasets (2000-2100) and for 3 RCP scenarios.
Figure 3: Here you can type out the land cover types in the subpanel titles. Also add the subpanels numbering (a-d), and modify the text accordingly.
Line 114-115: among others the Nash-Sutcliffe efficiency (NSE) (Nash and 115 Sutcliffe, 1970) and the Kling-Gupta efficiency (KGE) (Gupta et al., 2009) of what? Daily albedo?
Figure A3-A4: If there is space, I would type out the land cover types.
Citation: https://doi.org/10.5194/egusphere-2024-881-RC2
Model code and software
Future scenarios of albedo and radiative forcing resulting from changes in snow depth in Austria Joseph Kiem https://doi.org/10.5281/zenodo.10907185
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
279 | 77 | 30 | 386 | 20 | 23 |
- HTML: 279
- PDF: 77
- XML: 30
- Total: 386
- BibTeX: 20
- EndNote: 23
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1