Cloud–Snow Confusion with MODIS Snow Products in Boreal Forest Regions
, Hui Guo 1, Jian Wang 3,4 and Xiaohua Hao 3Round 1
Reviewer 1 Report
Review for paper
Cloud-snow confusion of MODIS snow products in boreal forest regions
The paper introduces a method for revising the cloud mask in MODIS collection 6 snow cover product, using temporally frequent NDSI-data by geostationary satellite data. The idea is to compare the variability of NDSI over a few hours’ period, based on the (very valid) assumption that if high variability indicates cloud while low variability indicates snow. The study was made for an extensive area in Northeast china, including also vast areas of boreal forest where MODIS snow product often seems to have problem with cloud mask (snow incorrectly flagged as cloud).
The study is well arranged, firmly reported; the results are interesting and probably feasible also in other areas, given that temporally frequent NDSI data are available. It fits well to be published.
However, there are some points where the authors should pay attention to before acceptance.
General:
- some sentences are directly copied from Reference [18], which does not follow a good practice. For example, this is the case in lines 136-140. I strongly encourage the authors to use their own phrasing throughout the paper.
- Please change figure captions more informative. it is difficult for the reader to catch the meaning of different colors etc. if they are only embedded in the text.
- all abbreviations and notations should be opened: e.g. SWE, Rgreen
- Particularly for figures 5 and 8 it would be good to have the applied RBG bands also in the figure caption or, even better, overlaid in each figure. Accordingly, explanation of which color indicates cloud, which indicates snow etc. should be also in the figure caption. This would help the reader a lot.
- in figures: 2,3,8,10,11, 12 please use slightly bigger font.
More detailed comment/questions:
Introduction is a bit ambiguous, and it gives the impression that authors are not spent so much time to really comprehend the spectral characteristics of snow and clouds, not to mention the cases of thin snow cover or patchy snow. Certainly NDSI is not the only variable for judging the pixel as cloudy, but this fact is not brought up at all (also the abstract is misleading in this respect as it is said there that the forest snow is misclassified as clouds because of the low NDSI; as if they were not any other tests??) . Some detailed comments in the following:
lines 51-66: This introductory paragraph is somewhat unclear to the reader. It is not clearly stated from where the MODIS snow cover collection 6 cloud mask comes from. Please describe the origin of the cloud mask with reference. MOD35 cloud mask even lacks in the reference list.
Lines 83-85: Should be: “ …because snow typically has a higher NDSI than clouds” . You could also mention a situation where the ground has snow-free patches (i.e., sub-pixel fractional snow) or where the snow layer is so thin it can be seen through, resulting in ground contribution. In those cases, the NDSI is not definitely higher than that of clouds (depending of course on cloud type).
line 88: it is not correct to say ‘NDSI value of forest snow is much less than that of pure snow”. I understand that you refer to scene reflectance observation, i.e., to the observed reflectance with both snow and forest canopy contribution. Please rephrase this.
Study area and Data is well written, only minor comments:
line 126: should be fractional snow Cover (FSC)
lines 136-140: see 1) in General comments
line 168: The bands information should be band information
Table 3: type in title of first column: should be Band, not Bands
Lines 170-179: Please provide some more information on this interesting data set: what is the areal coverage and is it publicly shared? If yes, please provide to contact point.
Snow-cloud Confusion in MOD10A1 provides a very interesting analysis. Some remarkable comments however:
line 191-193: is the CCD at each of the three stations extracted only from one pixel (the pixel where the station is located) of NDSI_Snow_Cover data array?
Figure 2: please provide a little bit bigger font
Figure 3: This is good, but why not to provide such a statistics also for CRU TS4.04? Although the datasets are in different resolutions, it would be very informative to see the bar presentations for both datasets. Alone, figure 3 is not so valuable.
lines 229-234: please provide reference for this 280K threshold
Cloud-snow Identification Methods is soundly written and the approach sounds quite feasible. A few remarks:
lines 241-247: please provide some considerations why in case of forest snow, low NDSI always leads to false cloud commission. It should be noted that in MOD10A1 C6 cloud detection is not dependent on NDSI only.
Equation 1: there is a typo. Should be (band2-band5) / (band2+band5)
Validation and Discussion:
Figure 8: Again, more information could be provided in the figure itself, not only in the paragraph above. In addition to the explanation of RGB bands, the meaning of red rectangles could be added in the figure caption.
Lines 348-351: please provide some considerations for also other reasons than low NDSI for false cloud commissions. Although the purpose of the paper is to approach the snow/cloud confusion through assessment of NDSI variation, it would be interesting to have your thoughts on the usability of other spectral bands to solve this problem in the future. Same applies to lines 353-359: after cloud removal, you find that NDSI_Snow_Cover data array indicates ‘no-snow’ for seemingly snow-covered pixels. Discussion on this problem would be interesting: to some extent, would it better to choose the false cloud (no information) instead of misclassified clear-sky pixel?
Comments for author File: 
Comments.docx
Author Response
The response to the reviewers is in the word file.
Author Response File: Author Response.docx
Reviewer 2 Report
The authors propose a new scheme for cloud masking incorporating a temporal analysis based on observations by a geostationary meteorological satellite.
A major weakness of the paper is the validation section. It is performed just on a small number of scenes, and just by roughly comparing with false-color images. I would highly recommend to use a different validation scheme that enables a more robust and less subjective evaluation. Furthermore, just one region in NW-China is evaluated, resulting in a lack of information about the spatial transferability of the proposed procedure (especially the threshold value) and the corresponding performance of the scheme.
Another weakness is the choice of data sources for the procedure. Why did you choose the Himawarri-8 geostationary meteorological satellite, what is the difference to other geostationary satellites? Is the data of this satellite freely available to other researchers to reproduce the results? Please extend the description on this satellite in the data section.
Finally, at some sections the language could be improved. In particular, the abstract could be written clearer (see specific remarks).
Specific remarks:
L 16: rephrase „would“. An alternative could be: “has a high potential to”
L 26: Rephrase the sentence. “Himawari-8 stationary meteorological satellite“ was already mentioned in the sentence before.
L 28/29: Rephrase the sentence to make your statement clearer.
L 29/30: Consider writing “are” instead of “will be”
L 31: Use present: “The results show that…”
L 33: What do you mean by “some ice clouds”. Can you specify the percentage?
L 54: What is meant by “more significant” here?
L 164/165: Define “new generation” and justify the statement that the satellite was launched earlier than all other comparable satellites. What are the advantages over other geostationary meteorological satellites? Justify why you chose this particular satellite.
L 223/224 and Fig 6: Specify the source for the spectral information, how did you retrieve the numbers, and how large is the standard deviation?
L 270: How did you resample the data? The spatial resolution of the geostationary satellite is 1-2 km, depending on wave length, right?
L 273: Is the threshold value stable over space and time? Or should we rather perform a histogram analysis for each scene?
Author Response
The response to the reviewers is in the word file.
Author Response File: Author Response.docx
Reviewer 3 Report
The paper propose a temporal NDSI cloud mask. The goal is to improve snow/cloud discrimination. They use time series from Himawari-8 to detected variance in NDSI to detect clouds.
Line 85. Snow does not always have a higher NDSI than clouds, I think the authors are trying to say that clouds cannot have as high of an NDSI as snow?
Section 2.5 Please explain the source of the observations in the gridded Climate Research Unit time series data. Are these surface station observations? Are these model outputs? How are they gridded?
Please turn figure 6 into a professionally formatted figure. What are the reflectance units? How many pixels are represented by each line? Or did you select a single pixel from each cloud?
Line 260 – describe how you manually selected cloud and snow pixels? Are you selections biased toward specific types of snow (mountain, plains, low NDSI, high NDSI?) are you selections biased towards specific types of clouds (thick clouds? Water clouds? Large clouds?)
Why is CRU TS4.04 not used as validation in section 5.1? is it validation data or just an independent estimate of cloud cover estimates to compare to?
The paper needs a more robust and quantitative validation of algorithm performance.
Author Response
The response to the reviewers is in the word file.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Authors have addressed my comments.
Two comments I recommend they address prior to publication:
1.) I still think the formatting of figure 6 could be improved to look more professional. Right now it looks like it was made in microsoft excel.
2.) The authors need to provide the readers with a robust written explanation for why they feel that qualitative validation of their method is sufficent validation, as opposed to per-pixel error statistics based on manually labeled snow and cloud pixels. All three L8 images are from the same satellite overpass on the same day for one H8 image. It is possible to find many more L8 images across a diversity of days and snow conditions for the study region and then perform per pixel error analysis comparing reprojected manually identified snow and cloud pixels in L8 to the equivalent H8 pixels. This would enable calculation of per pixel error statistics for the new method, which is common practice for quantitative evaluation of new method. Authors need to include a robust defense of why they did not choose to use this approach and just use subjective visual observations on select days as the validation for their method.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
