BRDF Estimations and Normalizations of Sentinel 2 Level 2 Data Using a Kalman-Filtering Approach and Comparisons with RadCalNet Measurements

Round 1

Reviewer 1 Report

Thanks for inviting me to review the manuscript, which aims to normalize the Sentinel 2 data using BRDF estimations. The method is useful in applications such as vegetation products. The method looks reasonable while the conclusions still need to be fully validated. To address this, I would suggest the authors do some revisions on validation of the methods on more sites. I would  recommend major revision before publication.

1. Three sites looks too limited for validation of the nadir reflectance output and demonstrate the capability/limitation of the method. To reach reliable conclusions or objective assessment of the method, more global sites are recommended to be used and the sites should represent different surface types.
2. Questions about Section 2.4 Aerosol and Rayleigh corrected Sen2Cor reflectances, How to do this correction for the pixels without in situ observations at the Radcalnet stations?
3. Question about Section 2.5. What is the SZA and VZA range of the observations from Sentinel? I just wonder if there are sufficient multi-angular observations to allow the application of this NASA and MODIS method on the claimed dataset. Also, it is recommend to demonstrate the MCD43A1 BRDF output in the result to be compared with the derived BRDF as the MODIS method has been introduced as a reference.
4. Section 2.4 appeared twice.
5. Line 127 in Page 3, suggest to clarify the angles more clear: the solar zenith, view zenith and relative azimuth angles.

Author Response

Dear,

First of all I would like to gratefully thank for the review.

My first point is that I noticed some formatting differences between the word document I uploaded and the word you got. Therefore the text has been less readable that it should have been. Sorry for this, it is a word issue, and as I used the template I do not understand why it happened. I will try to upload in addition a revised version in pdf format to try avoiding the same issue.

Below I answer to your questions:

1. Three sites looks too limited for validation of the nadir reflectance output and demonstrate the capability/limitation of the method. To reach reliable conclusions or objective assessment of the method, more global sites are recommended to be used and the sites should represent different surface types.

This is true that statistically speaking three selected sites do not represent all the surface types. In practice, the issue is the very limited number of in-situ directional spectral reflectance datasets freely available. To my knowledge, Radcalnet is the only network providing such data with “true” in-situ measurements, i.e. from a precise spectrometer or a photometer. I was part of the ESA ACIXII atmospheric correction experiment, and I know that for land they developed some “pseudo” in-situ dataset starting from the aeronet in-situ data (sky measurements). In this case they subtract to the measurements the atmospheric signal estimated using a radiative transfer model (RTM) to obtain the “pseudo in-situ surface reflectance measurements”. It must be noticed that, in this case, the “pseudo in-situ reflectance measurements” depend on the used RTM…  It must be noticed too that, to my knowledge, the ESA ACIXII dataset is not publicly available.

Regarding the obtained number of matchups used in this paper, we obtain 122 matchups for Gobabeb, 198 for Lacrau and 151 for Railroad Valley, which is very suitable for a statistical analysis. For all these reasons, I made the choice to focus on these three stations. The conclusions in the paper are explicitly related to the studied data.

2. Questions about Section 2.4 Aerosol and Rayleigh corrected Sen2Cor reflectances, How to do this correction for the pixels without in situ observations at the Radcalnet stations?

From the paper: “In order to unmix the errors performed by the atmospheric corrections from the BRDF normalization, we apply before the normalization an aerosol and Rayleigh correction given the aerosol loads and the surface pressure measured at the Radcalnet stations. “

It must be noted that we consider in this paper only the matchups, i.e. collocated observations in time between the in-situ and the satellite. Therefore, by definition, we do not perform this correction when there is no in-situ observation available.

3. Question about Section 2.5. What is the SZA and VZA range of the observations from Sentinel? I just wonder if there are sufficient multi-angular observations to allow the application of this NASA and MODIS method on the claimed dataset. Also, it is recommend to demonstrate the MCD43A1 BRDF output in the result to be compared with the derived BRDF as the MODIS method has been introduced as a reference.

All the geometry of acquisition is available in Figure 1 for the three sites. It must be noted that the three Radcalnet sites are very well known and used routinely by the MODIS teams for BRDF estimation and validation. This should answer the question.

In the paper we show in Figure 3, as you request, the comparisons with the MODIS BRDF MCD43A1 band 2 product. MODIS Band 2 was chosen as it very close spectrally to the S2 B08. We may have introduced one or two other MODIS bands but the size of the dataset to download is huge (several terabytes for MODIS and one band for many years as here). More important is that it must be noted that the other MODIS bands are not so close spectrally to the used S2. For these bands, Introducing a required “spectral shifting methods” to “adapt” the MODIS BRDF product to cope with the S2 spectral response would introduce some uncertainties in the results that will make the analysis difficult.

Considering the differences in resolution between the sensors and the fact that we use a bounded OLS for the MODIS-like method, results in Figure 3 show a good agreement (especially in mean, top of the figure 3) between the MODIS-like approach we derived and the results obtained using the MCD43A1 products for Band 02. This validates for me this part as you request. The differences in variance (bottom of the figure 3) are for me led by the use of a bounded OLS in our case and perhaps the atmospheric corrections.

4. Section 2.4 appeared twice.

Thanks for that.

5. Line 127 in Page 3, suggest to clarify the angles more clear: the solar zenith, view zenith and relative azimuth angles.

Thanks for that, I added the following sentence and reference: “In Figure 1 thetas is the sun zenith angle, thetav the viewing angle and dphi the relative azimuth angle (see [20] for the definition of the angles).”

Author Response File: Author Response.docx

Reviewer 2 Report

The paper proposes a method to account for BRDF effects in the processing of Sentinel-2 satellite signal. In practice, this is related to the Sen2cor tool now used operationally in the processing of Sentinel-2 data. The tool does not account for BRDF effects.

Overall, I find the manuscript a meaningful contribution. Especially the method is well described. However, there are aspects of this manuscript which I would suggest should be improved. 

As a first small remark, I would suggest that this improvement (ie accounting for BRDF effects) would benefit more people then just the climate modelling community mentioned at the end of the Abstract.

Moreover, the structure of the text could be improved. There are many very short sections with headings that are either numbered or not. This negatively impacts the clarity of argument.

l. 106-106 - the normalisation with locally-derived BRDF is of key importance, as it plays a role in validation of the proposed method. However, has that normalisation itself been validated?

My major concern is the Discussion, which is obviously short. Also, its last paragraph is very vague ("the KF allows correcting not only for the observation noise, but also potentially for some issues in the atmospheric corrections that may introduce some shifts and uncertainties.")

In the Discussion the author briefly mentions the issue which I find to be the biggest shortcoming of the analysis: none of the analysed sites has any significant vegetation cover, while it is the vegetation which is the primary reason why there is a BRDF bias in the satellite signal. Without a site covered by trees or at least shrubs, the proposed method cannot be fully validated. The author concludes that the method is better than the one created by the NASA MODIS team; however, without looking at the cases where significant vegetation occurs, this conclusion is not globally valid.

 

Author Response

Dear,

First of all I would like to gratefully thank for the review.

My first point is that I noticed some formatting differences between the word document I uploaded and the word you got. Therefore the text has been less readable that it should have been. Sorry for this, it is a word issue, and as I used the template I do not understand why it happened. I will try to upload in addition a revised version in pdf format to try avoiding the same issue.

Below I answer to your questions:

The paper proposes a method to account for BRDF effects in the processing of Sentinel-2 satellite signal. In practice, this is related to the Sen2cor tool now used operationally in the processing of Sentinel-2 data. The tool does not account for BRDF effects.

Overall, I find the manuscript a meaningful contribution. Especially the method is well described. However, there are aspects of this manuscript which I would suggest should be improved. 

As a first small remark, I would suggest that this improvement (ie accounting for BRDF effects) would benefit more people then just the climate modelling community mentioned at the end of the Abstract.

This is true, thanks for that. I updated the abstract and the discussion accordingly.

Moreover, the structure of the text could be improved. There are many very short sections with headings that are either numbered or not. This negatively impacts the clarity of argument.

I checked again and corrected the numbers for the sections. The non-numbered sections , i.e.  headings are in fact “title 3”.  I re-arranged this part in the paper and put the numbers for the “title 3” to ease the reading.

1. 106-106 - the normalisation with locally-derived BRDF is of key importance, as it plays a role in validation of the proposed method. However, has that normalisation itself been validated?

I am sorry, I couldn’t understand the reference “l. 106-106”. I checked the line 106 but without finding anything relative to this comment. Generally speaking, the normalisation process, i.e. “Normalization of the Sentinel 2 reflectances” subsection of §2.3 is a well known process used by NASA (MODIS) and Eumetsat. I added nevertheless two references [23] & [24] in this section (see below). I hope it will address your request:

“Nadir-normalized Sentinel 2 reflectances

Normalized reflectances are computed using the estimated BRDF to replace virtually the sensor at nadir. The nadir-normalized reflectances are then directly comparable with the Radcalnet nadir reflectance measurements. The goal of such normalization is also to favour inter-comparisons within and between satellite time-series that are acquired with different geometric conditions. Finally, with the BRDF parameters we apply the following equation to derive the nadir- normalized reflectances [23], [24]:

(3)

“

My major concern is the Discussion, which is obviously short. Also, its last paragraph is very vague ("the KF allows correcting not only for the observation noise, but also potentially for some issues in the atmospheric corrections that may introduce some shifts and uncertainties.")

Following your comment, I enhanced the discussion in the paper.

In the Discussion the author briefly mentions the issue which I find to be the biggest shortcoming of the analysis: none of the analysed sites has any significant vegetation cover, while it is the vegetation which is the primary reason why there is a BRDF bias in the satellite signal. Without a site covered by trees or at least shrubs, the proposed method cannot be fully validated

In practice, the issue is the very limited number of in-situ directional spectral reflectance datasets freely available. To my knowledge, Radcalnet is the only network providing such data with “true” in-situ measurements, i.e. from a precise spectrometer or a photometer. I was part of the ESA ACIXII atmospheric correction experiment, and I know that for land they developed some “pseudo” in-situ dataset starting from the aeronet in-situ data (sky measurements). In this case they subtract to the measurements the atmospheric signal estimated using a radiative transfer model (RTM) to obtain the “pseudo in-situ surface reflectance measurements”. It must be noticed that, in this case, the “pseudo in-situ reflectance measurements” depend on the used RTM…  It must be noticed too that, to my knowledge, the ESA ACIXII dataset is not publicly available.

Regarding the obtained number of matchups used in this paper, we obtain 122 matchups for Gobabeb, 198 for Lacrau and 151 for Railroad Valley, which is very suitable for a statistical analysis. For all these reasons, I made the choice to focus on these three stations. The conclusions in the paper are explicitly related to the studied data.

But I agree that the added value of the method should be validated using reflectance measurements over the forests. As a conclusion for this topic, i provide in this paper the needed information for the users to test themselves and compare the proposed method with their products or local in-situ datasets (especially over vegetation).

 The author concludes that the method is better than the one created by the NASA MODIS team; however, without looking at the cases where significant vegetation occurs, this conclusion is not globally valid.

As said before on the used three sites the improvement is clear compared with the NASA-MODIS BRDF MCD43A1 BAND 2 product. I checked and It was already clear in the paper (the band number was clearly cited). Nevertheless, I updated the paper, including the abstract and the discussion to specify that the improvement was demonstrated on the selected sites using the MODIS MCD43A1 BAND 2 product to be, as you required, more precise. For others bands the users can compare themselves the proposed method with the other MODIS MCD43A1 products.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I have no further comments. Thanks.