A sensor-agnostic albedo retrieval method for realistic sea ice surfaces &ndash; Model and validation

Zhou, Yingzhen; Li, Wei; Chen, Nan; Fan, Yongzhen; Stamnes, Knut

doi:https://doi.org/10.5194/tc-2021-397

Preprints

https://doi.org/10.5194/tc-2021-397

Preprints

18 Jan 2022

Status: a revised version of this preprint is currently under review for the journal TC.

A sensor-agnostic albedo retrieval method for realistic sea ice surfaces – Model and validation

Yingzhen Zhou¹, Wei Li¹, Nan Chen¹, Yongzhen Fan², and Knut Stamnes¹ Yingzhen Zhou et al.

Show author details

¹Light and Life Laboratory, Department of Physics, Stevens Institute of Technology, Hoboken, NJ 07307, USA
²Cooperative Institute for Satellite Earth System Studies (CISESS), Earth System Science Interdisciplinary Center (ESSIC), University of Maryland, College Park, MD 20740, USA

Received: 28 Dec 2021 – Discussion started: 18 Jan 2022

Abstract. The cryosphere's surface (snow, sea ice, and water) regulates global climate through several feedback mechanisms. Broadband albedo is a critical parameter determining the radiative energy balance of the complex atmosphere-cryosphere system, but there is currently no reliable, operational albedo retrieval product capable of assessing the global sea-ice albedo with sufficient spatial-temporal resolution for studies of sea-ice dynamics and for use in global climate models.

A framework was established for remote sensing of sea ice albedo that integrates sea-ice physics with high computational efficiency, and can be applied to any optical sensor that measures appropriate radiance data. A scientific machine learning (SciML) approach was developed and trained on a large synthetic dataset (SD) constructed using a coupled atmosphere-surface radiative transfer model (RTM). The resulting RTM/SciML framework combines the RTM with a multi-layer artificial neural network SciML model. In comparison to the NASA MODIS MCD43 albedo product, this framework does not depend on observations from multiple days, and can be applied to single angular observations obtained under clear-sky conditions. Compared to the existing melt pond fraction-based approach for albedo retrieval, the RTM/SciML framework has the advantage of being applicable to a wide variety of cryosphere surfaces, both heterogeneous and homogeneous. Validation of the RTM/SciML albedo product using MODIS and SGLI data against measurements obtained from aircraft campaigns revealed excellent agreement, with mean absolute error of 0.047 for above 2000 clear-sky pixels.

Yingzhen Zhou et al.

Status: final response (author comments only)

RC1:
'Comment on tc-2021-397', Anonymous Referee #1, 09 Feb 2022
General comments: I have read the manuscript entitled “A sensor-agnostic albedo retrieval method for realistic sea ice surfaces - Model and validation”. In this manuscript, the authors declared that they developed a sensor-agnostic sea ice albedo estimation method using a coupled atmosphere-surface radiative transfer model (RTM) and a multi-layer artificial neural network (MLANN) model. The results were validated with the measurements of ACLOUD and AFLUX campaigns, and compared with the MCD43D, MERIS, and OLCI albedo product. The validation and comparison results indicated that the albedo estimated by the MLANN method are in good agreement with the measurements, and can provide better estimation of sea ice albedo than the other albedo products. The framework for estimating sea ice albedo using RTM/MLANN is interesting for the remote sensing society and polar studies. However, there are still several issues need to be addressed before publication. The main issues of this manuscript are listed as follows.

Specific comments:

Although detailed information of the coupled radiative transfer model AccRT can be found in the literatures, I suggest to add a concise description about it in the manuscript.

The method of how to construct the synthetic dataset (SD) with the coupled RTM is not clear. The detailed information about the inherit optical properties (IOPs) listed in Table 2 are needed, such as the data ranges, probability distribution, and constraints.

The framework of the RTM/MLANN is not clear. I suggest to add a flowchart for it.

In the manuscript, the MLANN method to used estimate the sea ice albedo. What are the performances of training, validating, and predicting accuracies of this artificial neural network model?

The authors declared that the sensor-agnostic albedo retrieval method has the ability to apply to any optical sensor, however few explanations about this are shown in the manuscript. I suggest the authors to further explain the major theories of this method. In fact, other methods such as the MPD and direct-estimation algorithm, can also be adopted to other sensors easily. Please add a discussion about it.

The comparisons with MCD43D, MERIS, and OLCI datasets were not easily for reader to interpret. I suggested to add scatter plots to compare the differences of these datasets.

Figure 13, the authors declared that the MERIS albedo product are higher than the albedo estimated by the MLANN method in the areas with large melt pond fraction (greater than 50%). However, this difference is not obvious, and the major differences appeared in the upper right corner. Please provide an explanation for it.

Figure 13, the measurements of campaigns were not shown in this figure. Why? Please add the validation data for comparison.

In the abstract, the mean absolute error (MAE) of 0.047 was used for indicating the accuracy of this method. I suggest to use root mean standard error (RMSE) to represent the estimation accuracies for the visible, near infrared, and shortwave albedo.

Technical corrections:

Figure 7. The color ramp of this figure is not easily to interpret. Please change it.

Line 493, the sentences of “Istomina et al. (2015); Istomina (2020)” can be rewritten as “Istomina et al. (2015; 2020)”.

Caption of Figure 13. “(Qu et al. (2015), this study, and Istomina et al. (2015))”. The reference Qu et al. 2015 is not related with this figure.
Citation: https://doi.org/10.5194/tc-2021-397-RC1
- AC1: 'Reply on RC1', Yingzhen Zhou, 23 Apr 2022
  
  The comment was uploaded in the form of a supplement: https://tc.copernicus.org/preprints/tc-2021-397/tc-2021-397-AC1-supplement.pdf
  
  Citation: https://doi.org/10.5194/tc-2021-397-AC1
RC2:
'Comment on tc-2021-397', Anonymous Referee #2, 13 Feb 2022
Summary of Comments on

tc-2021-397_review-1round.pdf

Thanks for this review invitation. The manuscript describes a method combining a RTM model and a machine learning method to estimate broadband sea-ice albedos from TOA reflectance. The validation of method with ground measurements over sea ice is interesting. However, the method description is not clear enough while some statements are not objective and eager to emphasize the advantage of the proposed method. I have the following suggestions/questions before recommendation for publication.

The coupled RTM is used to simulate TOA reflectance from various sea ice surface and atmospheric properties. The surface parameters are listed but the values were not mentioned as well as the sampling strategy. I cannot figure out how the authors determine the distribution and relevance among the parameters. Similar concern for the atmospheric parameters and the solar/view angles.

The machine learning method needs more detailed description about how it was used. How to deal with the invalid retrievals from the relationship? Is there a post-processing? It was mentioned there are two models trained. What are their difference and advantages?

The author emphasized many times about the advantage of the proposed method than the previous MPD or direct-estimation methods. However, many descriptions needs to be clarified or discussed more. What are the advantage of the coupled RTM rather than the separate radiative transfer models? Is there any quantitative comparison about this? Is a classification within sea-ice surface needed in previous methods? The method is claimed as independent on sensor or spatial resolution, how that is realized without considering the spectral response function difference? Did the previous method restrict to a specific spatial resolution?

Minor comments:

Please refer to the highlights in the PDF document.

Thanks.
Citation: https://doi.org/10.5194/tc-2021-397-RC2
- AC2: 'Reply on RC2', Yingzhen Zhou, 23 Apr 2022
  
  The comment was uploaded in the form of a supplement: https://tc.copernicus.org/preprints/tc-2021-397/tc-2021-397-AC2-supplement.pdf
  
  Citation: https://doi.org/10.5194/tc-2021-397-AC2
RC3:
'Comment on tc-2021-397', Anonymous Referee #3, 06 Mar 2022

General comments:

This manuscript describes the retrieval of broadband albedo using coupled machine learning and radiative transfer model. Although retrieved albedo was validated and compared, the methods and results part are not clear enough and inconsistent. I don’t see novelties and advantages in methods and results section. The motivations and justifications should be more addressed before publication. For example, flowcharts, tables, explanation, Justification are needed.

Specific comments

Table 1: Please add advantages and disadvantages of each product in the table.

P4, L 104: Please Just give some short explanation, as we don’t see the paper ready to submit

Please make data section and explain satellite used for the retrieval, validation dataset, comparison dataset before methodology.

Section 2.2: I would be merited to have a flowchart to understand better.

Section 2.4: The details of structure of MLANN must be addressed. For example, the number of layers, activation functions, weight initialization, input variables (should be synthetic dataset, SD), target variables, how to train and validate, accuracies.

P10, L251: the cloud screening method used MODIS bands? If it’s right, how can it be used for SGLI?

Figure 2: This figure should go data section.

Figure 3: Can you explain what is the difference between c and d?

Section 3.4: I don’t understand the link between surface metamorphism and two days (Morning-noon-early afternoon, late afternoon) albedo changes. Figure 8 is not mentioned in section 3.4. If they have some links please elaborate more.

Section 3.5: The retrieved albedo using SGLI is also comprehensively validated like a MODIS and analyzed with solar zenith angle, surface metamorphism. The retrieved albedo using SGLI should be validated and compared in parallel.

Section 4.1 and 4.2: In 4.1, albedo retrieval map against MCD is daily but in 4.2, 5-day mean albedo map against MERIS. Can you elaborate why they are different?

The retrieved albedo maps are only shown near Svalbard islands but Pan-Arctic retrieved albedo map should be shown and have to be compared with other comparison dataset.

Minor comments:

All captions in the table should be above table.

L 99, 100: Please mention SGLI MCD 43 full name

P4, L101-105: should go to the discussion section

3 validaiton: The authors mentioned MOSAiC. Have you used the data from MOSAiC for the validation?

L 497-499: should go comparison dataset

Citation: https://doi.org/10.5194/tc-2021-397-RC3
- AC3: 'Reply on RC3', Yingzhen Zhou, 23 Apr 2022
  
  The comment was uploaded in the form of a supplement: https://tc.copernicus.org/preprints/tc-2021-397/tc-2021-397-AC3-supplement.pdf
  
  Citation: https://doi.org/10.5194/tc-2021-397-AC3
RC4:
'Comment on tc-2021-397', Anonymous Referee #4, 07 Mar 2022

AccuRT/RTM looks novel but needs sharing as open source as well as documentation to be subject to rigorous peer review. For example compare this situation with the RAMI experiments (Widlowski et al., 2007) or the MYSTIC cloud simulator (Mayer et al., 2010). Similarly, MLANN looks like a significant advance but again needs sharing as an open source resource to have any impact on the community. Very limited examples are not really a proper “validation” when the uncertainties are unknown of the “truth” data-sets. The authors do not present convincing evidence that MLNN will work on a time series of MODIS (let alone other instruments) to show the evolution of sea ice albedo during the Arctic spring/summer. They ignore the work of the NOAA group on VIIRS and the NASA group at UMD on the VIIRS-SNPP and MODIS time series and the UCL group on MISR instantaneous albedo retrievals all of which have long time series datasets publicly available which this paper does not. This technique and the paper is of high interest to the community but needs less hyperbole (on line 3 the authors claim there are no reliable albedo products, this reviewer would strongly dispute this) and more quantitative intercomparison with the aforementioned datasets before it can be considered for publication. Otherwise, this paper will represent cherry-picking results without any serious self-critical analysis.

1. There is an incorrect assertion in the abstract: “there is currently no reliable, operational albedo retrieval product capable of assessing the global sea-ice albedo with sufficient spatial-temporal resolution for studies of sea-ice dynamics and for use in global climate models “

2. NOAA have had an operational spectral and shortwave albedo product multiple times per day derived from NOAA-20 VIIRS since September 2018.

3. There are a bewildering number of acronyms that are not defined in the order that they are introduced. The paper needs to include a list of acronyms that the reader can consult.

4. One example is “comprehensive SD” on line 179 which is not defined previously. What is “SD”?

5. The authors should provide evidence for the negligible differences of NIR and SW albedos for the differences given the upper wavelengths of 2.1µm, 2.5µm, and 3µm (lines 278-279)

6. Absolute albedo is not very helpful when the range in albedos is so large. It is better to show the coefficient of variation (stdv/mean) to see how the albedo varies in uncertainties. (Line 437)

7. The so-called validation shown here is usually referred to as stage 1 (CEOS-WGCV-LPV, see https://lpvs.gsfc.nasa.gov/) as there are very limited dates and there is no uncertainty specified for the aircraft measurements.

8. Why was MLANN not adapted for uses with SGLI, VIIRS, and OLCI?

9. Also, what about comparisons with the OLCI product derived using the Kokhanovsky et al. 2020 (Line 687) SNAP processor?

10. Where are the open-source repositories of AccuRT and the RTM/SciML as well as MLANN?

Cited references

1. Widlowski, J.-L.; Taberner, M.; Pinty, B.; Bruniquel-Pinel, V.; Disney, M.; Fernandes, R.; Gastellu-Etchegorry, J.-P.; Gobron, N.; Kuusk, A.; Lavergne, T.; Leblanc, S.; Lewis, P. E.; Martin, E.; Mõttus, M.; North, P. R. J.; Qin, W.; Robustelli, M.; Rochdi, N.; Ruiloba, R.; Soler, C.; Thompson, R.; Verhoef, W.; Verstraete, M. M.; Xie, D. Third Radiation Transfer Model Intercomparison (RAMI) exercise: Documenting progress in canopy reflectance models. Journal of Geophysical research Solid earth 2007, 112, D09111.

2. Mayer, B.; Hoch, S. W.; Whiteman, C. D. Validating the MYSTIC three-dimensional radiative transfer model with observations from the complex topography of Arizona's Meteor Crater. Atmos Chem Phys 2010, 10, 8685–8696.

Annotation Summary of tc-2021-397 - annotated.pdf.

Note [page 1]: Line 3: NOAA have had an operational DAILY spectral and shortwave validated albedo product derived from VIIRS since September 2018. There is also a paper which describes the sea ice product specifically which you reference below from Peng et al. (2018) but which you ignore in your paper. Where is the evidence that this is not reliable and operational? Is your proposed product operational? This sentence should be modified.

Note [page 1]: Line 9: But neither does the MISR (Kharbouche & Muller, 2018) nor does the GLASS product both of which are produced from instantaneous measurements.

Note [page 1]: Define acronym

Note [page 1]: Line 14: This is not a very helpful measure of error if you don’t provide the range and mean?

Note [page 1]: Line 23: Extent? Thickness? Concentration? Which attribute is in decline?

Note [page 2]: Spatial resolution?

Note [page 2]: What is the resolution?

Note [page 2]: Spatial resolution?

Note [page 2]: Omits MISR products from Kharbouche & Muller (2018)

Also, needs spatial and temporal resolution and time range adding as well as URLs of where the product is described and available.

Note [page 2]: Table 1 is very poor. Needs consistency in spatial resolution, needs a column for time range for which they are available. Needs an additional column for validation level (see CEOS comment later)

Strikeout [page 2]: (2018)) have been validated using ground truths from the Greenland Ice Sheet and snow-covered land, but the validation is not repre- sentative of the highly variable sea-ice surface.

Note [page 2]: L38: This is because there are no reliable long-term measurements of sea ice albedo publicly available.

Note [page 2]: Line 41: But that is true of all the so-called validation exercises including your aircraft data. This I sonly for a few dates, can be up to 5 hours different in time with the satellite overpass and dos not have any uncertainties associated with the aircraft measurements.

Strikeout [page 3]: The CLARA-SAL product (Riihelä et al., 2013) calculates the ‘black-sky albedo’, the App-x and MCD43 products generate both black- and white-sky albedo (Schaaf et al., 2002), and all other albedo products directly yield ‘blue-sky albedo’, which is also the albedo that can be directly compared to with…

Highlight [page 4]: Section 3 is devoted to validation of the albedo-retrieval product, discussion of potential causes of uncertainty, and a sensor-to-sensor comparison of the retrieval outcomes using MODIS and SGLI radiance data.

Highlight [page 4]: (2021 (ready to submit), as well as its applicability to cloud radiative forcing (CRF) analysis and to enhancement of current regional/global climate models.

Note [page 5]: Line 118: Define acronym

Highlight [page 6]: A coupled RTM that is used in conjunction with a realistic modeling of the system based on its IOPs can address both of these challenges.

Highlight [page 8]: (i) the complicated surface and atmosphere conditions by varying the optical properties in Table 2, and…

Note [page 8]: Line 185: all the parameters need to be elaborated in a table as this is an open journal. Also, is AccuRT open source? And what about the retrieval method?

Note [page 8]: Line 195: Are these available? Where are they described?

Note [page 8]: Line 203: Reference needed

Note [page 8]: Line 208: need reference and/or URL for this unknown sensor.

Note [page 8]: Line 209: It is disappointing that this sensor was not examined as it could then be compared against the operational VIIRS product from Peng.

Note [page 9]: Line 215: What does the L stand for?

Note [page 11]: Line 281: What is this footprint? How is the difference in resolution dealt with? Aggregation?

Note [page 11]: Line 295: Where does this significant decrease come from? H2O absorption?

Highlight [page 12]: matching dates in the latitude-longitude range of ï¬ight operations (identiï¬ed using the MLCM, Chen et al.

Note [page 12]: Line 311: The visible results do show the lowest value of r and slope. The authors should comment on why these produce the worst results.

Note [page 13]: Line 326: how fast did the sea ice move over the time period between the MODIS observation and the aircraft observation? Itis likely that the poorer disagreement is due to the fact that the same piece of sea ice is not observed by the aircraft.

Note [page 14]: Figure 3: caption: What is the time range shown here between these 2 sets of measurements?

Note [page 16]: Line 378: This is difficult to believe as most sea ice moves at >10 km/day at this time of year.

Note [page 21]: Line 437: Absolute albedo error is not very helpful when the range in albedos is so large. It is better to show coefficient of variation (stdv/mean) to see how the albedo varies in uncertainties.

Note [page 21]: Line 441: Remind the reader what MPD is and define in a list of acronyms.

Note [page 22]: Figure 10: What does EE mean? Define in the caption.

Note [page 23]: Lines 460-461: Is this upper range of wavelength for n2b significant?

Note [page 27]: Figure 14 caption: Why is the OLCI retrieval so much coarser in spatial resolution?

Note [page 28]: Line 529: Why on earth was this done?

Note [page 29]: Line 55: this is hyperbole. Where is this demonstrated? I only see MODIS & SGLI results.

Note [page 29]: L567: Why is this important? What impact does this have?

Highlight [page 29]: Information of both the surface BRDF and the IOPs of the atmosphere have already been taken into account.

Highlight [page 29]: (2021 (ready to submit)).

Note [page 29]: Line 574: What is a whole image? A 5-minute MODIS Level-1B data granule?

Note [page 29]: Line 585: But so are MISR (which uses MODIS cloud masks) and VIIRS & MODIS (e.g. GLASS) direct estimation algorithms?

Note [page 30]: Line 588: EGU journals should only permit open access datasets with a publication DOI. In addition, all software should be open access. This is what differentiates EGU from other comparable journals. This should not be an exception.

Note [page 30]: Table A1 caption: Where does these percentages come from?

Note [page 33]: Line 597: Exact URLs should be provided.

Note [page 33]: Line 600: Grant numbers should be listed.

Citation: https://doi.org/10.5194/tc-2021-397-RC4
- AC4: 'Reply on RC4', Yingzhen Zhou, 23 Apr 2022
  
  The comment was uploaded in the form of a supplement: https://tc.copernicus.org/preprints/tc-2021-397/tc-2021-397-AC4-supplement.zip
  
  Citation: https://doi.org/10.5194/tc-2021-397-AC4

Yingzhen Zhou et al.

Viewed

Total article views: 787 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
575	178	34	787	14	14

HTML: 575
PDF: 178
XML: 34
Total: 787
BibTeX: 14
EndNote: 14

Views and downloads (calculated since 18 Jan 2022)

Month	HTML	PDF	XML	Total
Jan 2022	167	46	5	218
Feb 2022	157	32	7	196
Mar 2022	82	24	7	113
Apr 2022	58	23	9	90
May 2022	40	16	1	57
Jun 2022	11	7	1	19
Jul 2022	6	4	0	10
Aug 2022	16	11	0	27
Sep 2022	8	3	0	11
Oct 2022	15	6	2	23
Nov 2022	12	5	1	18
Dec 2022	3	1	1	5

Cumulative views and downloads (calculated since 18 Jan 2022)

Month	HTML	PDF	XML	Total
Jan 2022	167	46	5	218
Feb 2022	157	32	7	196
Mar 2022	82	24	7	113
Apr 2022	58	23	9	90
May 2022	40	16	1	57
Jun 2022	11	7	1	19
Jul 2022	6	4	0	10
Aug 2022	16	11	0	27
Sep 2022	8	3	0	11
Oct 2022	15	6	2	23
Nov 2022	12	5	1	18
Dec 2022	3	1	1	5

Viewed (geographical distribution)

Total article views: 717 (including HTML, PDF, and XML) Thereof 717 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 08 Dec 2022

Short summary

We present a method to compute albedo (percentage of the light reflected) of the cryosphere surface using observations from any optical satellite sensors. This method surpasses all existing models, may be applied globally (Arctic, Antarctic, Sea of Okhotsk) and to any realistic cryosphere surface: sea ice, snow-covered ice, melt-pond, open-ocean, and their mixing. Evaluation of the albedo values calculated using the approach demonstrated excellent agreement with observations.


Total:	0
HTML:	0
PDF:	0
XML:	0