the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Separating the albedo reducing effect of different light absorbing particles on snow using deep learning
Abstract. Several different types of light absorbing particles (LAPS) darken snow surfaces, enhancing snow melt on glaciers and snowfields. LAPs are often present as a mixture of biotic and abiotic components at the snow surface, yet methods to separate their respective abundance and albedo-reducing effects are lacking. Here, we present a new optimisation method enabling the retrievals of dust, black carbon and red algal abundances as well as their respective darkening effects from spectral albedo. This method includes a deep learning emulator of a radiative transfer model (RTM), and an inversion algorithm. The emulator alone can be used as a fast and lightweight alternative to the full RTM with the possibility to add new features, such as new light absorbing particles. The inversion method was applied to 180 ground field spectra collected on snowfields in Southern Norway, with a mean absolute error on spectral albedo of 0.0056, and surface parameters that closely matched expectations from qualitative assessments of the surface. The emulator predictions of surface parameters were used to quantify the albedo reducing effect of algal blooms, mineral dusts and dark particles represented by black carbon. Among these 180 surfaces, the albedo reduction due to light absorbing particles was highly variable and reached up to 0.13, 0.21 and 0.25 for red algal blooms, mineral dusts and dark particles respectively. In addition, the effect of a single LAP was attenuated by the presence of other LAPs by up to 2–3 times. These results demonstrate the importance of considering the individual types of light absorbing particles and their concomitant interactions for forecasting snow albedo.
- Preprint
(6839 KB) - Metadata XML
-
Supplement
(56754 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-2583', Urs Niklas Bohn, 28 Sep 2024
Chevrollier et al.: “Separating the albedo reducing effect of different light absorbing particles on snow using deep learning”, Review by Niklas Bohn.
Chevrollier et al. present a new forward model emulator that predicts spectral snow albedo based on a given set of input parameters, including snow grain size, liquid water content, and concentration of different light-absorbing particles (LAPs) such as dust, black carbon, and red algae. The authors train and test the emulator with simulations from the BioSNICAR radiative transfer model (RTM) and advertise it as a fast and lightweight alternative to running the full RTM in an inversion scheme. The study demonstrates that the new emulator enables accurate retrievals of LAPs from selected field spectra collected on snowfields in Southern Norway. In addition, the authors quantify the albedo reducing effect as well as radiative forcing for each type of LAP separately, highlighting the importance of considering the individual types of LAPs and their interactions for forecasting snow albedo.
Chevrollier et al. prepared an interesting, nicely compact, and useful study, which is important for future algorithm development and improvement, particularly in the context of upcoming spaceborne imaging spectrometer missions, such as NASA’s Surface Biology and Geology (SBG). The proposed emulator is a promising alternative to existing inversion schemes of complex RTMs. Given this high value, I recommend publication in The Cryosphere. I only have a few comments that need to be addressed in a minor revision before publication. These comments are outlined below.
General comments
- A short introduction to BioSNICAR/SNICAR would be good. What are its basic concepts? Which input options exist? What's the output quantity? Which radiative transfer approach is used (two-stream vs. multi-stream)? In particular, Section 2.1.1 would better go as an introduction to the utilized radiative transfer model, i.e., BioSNICAR, with changing the section title accordingly. You could still keep the description of input and output data, but expand a bit more on the underlying physics of SNICAR.
- You represent snow as a granular medium with spherical grains. However, many controversies exist in the literature about how to model the shape of snow grains. I usually apply the 'collection of spheres' approach from Grenfell and Warren (1999) myself, but a brief discussion about potential impacts of assuming the spherical representation, and possible alternatives would be good.
- There are different sets of dust optical properties available in SNICAR, depending on their source and/or sampling region. Please provide more detail about which exact dataset you used and why.
- It would be good to have a paragraph in the introduction highlighting the potential of hyperspectral/multispectral sensors for remote sensing of LAP in snow.
Specific comments
Lines 25 - 26: Can you provide any references to studies that did these forward modeling experiments?
Line 37: Bohn et al. (2021) do not actually use Gaussian processes, but simply optimal estimation that acts on the assumption that state parameters and their errors show a Gaussian distribution.
Line 45: Why snow algae in particular? From my experience, modeling the effects of black carbon particles is clearly more challenging because of their minor absorptivity and weak occurrences.
Lines 59 – 61: I don't think you need to mention all the used Python libraries here. Maybe only tensorflow and keras as those two are important for understanding the applied deep learning approach.
Line 73: To better justify this, you could create non-linear but homogeneous grids by using Sobol sequences (https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.qmc.Sobol.html).
Table 1: Why only going up to 15%? Previous studies have shown that up to 25% could be realistic (Green et al. 2002; Bohn et al. 2021).
Line 92: Can you further substantiate this assumption?
Lines 94 – 95: Hyperspectral/multispectral comes out of the blue here. See general comments.
Line 123: How fast is this inversion scheme? Could you give a few numbers?
Line 134: What does 'homogeneous enough' mean? Did you have minor influences from other surface types or roughness?
Line 145: Calculated radiative forcing is not instantaneous when you use 24h daily averaged shortwave incoming radiation. To get the instantaneous radiative forcing, you would need to multiply the BBA reduction by the incoming radiation at the exact time of the measurement.
Line 195: You need to clarify a bit better if you applied the ARFs to the HCRF spectra before doing the inversions. I guess you did not, but it's not fully clear from the text.
Lines 208 – 210: So why not using a multi-stream RTM such as DISORT to account for reflectance anisotropy?
Figure 2: I don't see the values of the retrieved grain radius in these figures.
Line 239: You need to clarify in the methods section how you calculated the IRF.
Line 253: Again, please clarify at the beginning, which type of dust OPs you're applying in this study. See general comments.
Line 256: The expression light absorbing particles darkening snow surfaces sounds odd, please try to revise.
Figure A1: Again, the grain size values are missing in this plot.
Technical corrections
Line 1: Several different types of light absorbing particles (LAPs) darken snow surfaces, …
Line 49: BioSNICAR
Line 56: BioSNICAR
Table 1: The caption should be located above the table.
Line 93: spectral albedo
Line 134: … an hour prior to the measurements …
Line 140: Since the measurements are not equivalent to …
Line 148: … the spectral albedo output by the …
Line 171: The emulator is therefore a practical …
Fig. 1, caption: … to the (b) highest and (c) lower mean …
Table 2: Again, I think the caption goes above the table.
Line 216: … if the spectral diffuse and direct partitioning …
Fig. 2, caption: … is approximately 45x45 cm, centered on …
Line 252: … could be integrated in the …
Line 259: … emulating a radiative transfer model, and an …
Fig. A1, caption: … is approximately 45x45 cm, centered on …
Fig A2, caption: The retrievals of LAP concentrations are compared to the retrievals when applying ARFs from …
References
Bohn, N., Painter, T. H., Thompson, D. R., Carmon, N., Susiluoto, J., Turmon, M. J., Helmlinger, M. C., Green, R. O., Cook, J. M., and Guanter, L.: Optimal estimation of snow and ice surface parameters from imaging spectroscopy measurements, Remote Sensing of Environment, 264, https://doi.org/10.1016/j.rse.2021.112613, 2021.
Green, R. O., Dozier, J., Roberts, D. A., and Painter, T. H.: Spectral snow-reflectance models for grain-size and liquid-water fraction in melting snow for the solar-reflected spectrum, Annals of Glaciology, 34, 71–73, https://doi.org/10.3189/172756402781817987, 2002.
Grenfell, T. C. and Warren, S. G.: Representation of a nonspherical ice particle by a collection of independent spheres for scattering and ab- sorption of radiation, Journal of Geophysical Research: Atmospheres, 104, 31 679–31 709, https://doi.org/10.1029/1999JD900496, 1999.
Citation: https://doi.org/10.5194/egusphere-2024-2583-RC1 -
RC2: 'Comment on egusphere-2024-2583', Anonymous Referee #2, 01 Nov 2024
The authors developed a deep learning emulator of a snowpack radiative transfer model (RTM), and an inversion algorithm to retrieve key contributions from different LAPs. They found very high accuracy of the emulator and the inversion algorithm. This work provides very useful modeling tools for snow albedo calculations affected by LAPs and retrieving LAPs’ respective contributions, which would have broad application for satellite data. I have a few comments and suggestions for the authors to consider and address.
Major comments:
- I am a little confused by the necessity of developing/using the machine learning (ML) emulator of RTM. Does the optimising retrieval algorithm only work with the ML emulator? Can the algorithm also work/couple directly with the physics-based RTM e.g., BioSNICAR? Or is it just for the computational efficiency purpose?
- The authors made a few key assumptions when creating the training dataset, such as two snow layers, spherical snow grains, only upper layer for LAPs, and constant snow density, which limit the applicability of the emulator. Among these assumptions, the top 2 cm snow layer containing LAPs and constant density are probably two most important limitations, which could be relaxed to allow them to vary during the emulator training to increase the applicability of the emulator for future studies. Particularly, only the top 2 cm containing LAPs is not realistic.
- Does the inversion algorithm search for local optima or global optima? Would there be equifinality issue? Also, how sensitive is the algorithm retrieval result to the initial guess and how to effectively select the initial guess?
Minor comments:
- Line 95: What is the reason for the biosnicar discontinuity around 2.5um?
- Lines 223-250: How did the authors compute the radiative forcing from albedo reduction? What downward solar radiation data did the authors use?
- I would suggest adding a section to discuss uncertainties involved in the ML-based emulator and the inversion algorithm.
Citation: https://doi.org/10.5194/egusphere-2024-2583-RC2 -
RC3: 'Comment on egusphere-2024-2583', Anonymous Referee #3, 15 Nov 2024
The authors use a deep learning emulator of a radiative transfer model (RTM) with an inversion model to determine contributions of dust, black carbon, and snow algae from snow spectrums obtained in the field. The deep learning emulator performed extremely well, simulating RTM albedo within <0.05% error at 30 times the speed of the RTM. Similarly, the inversion model was able to reproduced ground spectra from Southern Norway, including the optical properties of LAP. The authors provide a high quality and useful study for retrievals of LAP properties and radiative impacts to snow. This work is especially exciting considering current and future hyperspectral spaceborne instruments such as EMIT and SBG
Major Comments:
1. A short description in the methods of biosnicar/SNICAR would be useful. This should include the optical schemes used and any error associate with the model.
2. This work is very exciting for potential use in larger scale hyperspectral measurements. The conclusions could use some discussion of any potential issues or challenges of scaling the model up to regional or gloabal scale, especially with the mention of satellite use. For example, dust optical properties can vary by region. Would this impact use of the model?
Minor Comments
- Line 27 "..allow to study the impacts of LAPs..." is a bit difficult to read, consider rewording
- In the RTM model setup, two snow layers are used, with the lower layer being a semi-infinite layer. Dust and black carbon tend to be deposited in layers throughout the season, often together. The assumption of the near semi-infinite lower layer can introduce some error during the melt season as these buried LAP layers get close to the surface
- The title of Section 2.3 should be updated to say daily radiative forcing instead of instantaneous based on the methods described. Instantaneous radiative forcing would multiply the BBA reduction by the incoming solar radiation at the time of the measurement.
- The methods in Section 2.3 could use some more information. When calculating reduction in BBA from a LAP, is a spectrum with the same grain size and other LAP concentrations being used
- The results mention both daily average and instantaneous radiative forcing. If both are being used, the calculation of instantaneous radiative forcing should be covered in Section 2.3
Citation: https://doi.org/10.5194/egusphere-2024-2583-RC3
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
291 | 71 | 99 | 461 | 19 | 3 | 5 |
- HTML: 291
- PDF: 71
- XML: 99
- Total: 461
- Supplement: 19
- BibTeX: 3
- EndNote: 5
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1