A Novel Approach to Validate Satellite Snowfall Retrievals by Ground-Based Point Measurements
Round 1
Reviewer 1 Report
It is my pleasure to review remotesensing-1492225 “A Novel Approach to Validate Satellite Snowfall Retrievals by Ground-Based Point Measurements” by Jeoung et al. The authors first pointed out the insufficiency resulting from using either SNOTEL or GHCND data alone for validating CloudSat snowfall, then proposed a new method by combining SNOTEL and GHCND data together with accounting for the stations’ elevation information. As we know, snowfall is sensitive to global and regional climate change. I found there are three fatal errors in this manuscript:
- The CloudSat snow product ranging from 2006 to 2017 was used for annual mean snowfall estimates, however, the station mean data average during 2014-2020. The time ranges between these two datasets are different, which may make the results unbelievable, e.g. CloudSat CPR mostly underestimates snowfall against SNOTEL while having a mixture of over- and underestimation against GHCND observations.
- As mentioned in this manuscript, Shi and Liu (2021) found that the ratio of snowfall to precipitation occurrence has been increasing at high latitudes in the past 40 years. It seems that snowfall is increasing with the global warming. The mechanism of this increasing trend is complex, however, this manuscript ignored the impact of global climate change on snowfall using the dataset with different time ranges, and exaggerated the dependence of snow on elevation.
- Snowfall is random with time, it is insufficient to derive the snowfall pattern using 7 years observation data.
So, the authors may need to find more data with same time ranges to improve this novel method.
Author Response
We thank the three anonymous reviewers taking their valuable time to review our paper. Their comments are well received, and a major revision was conducted based on these comments. Two major changes have been made in the revised manuscript. (1) The time period for the validation is switched to 2006-2017, overlapping CloudSat observations. (2) A discussion section is added to discuss the novelty of the proposed method and the difference of climatological means of snowfall between 2006-2017 and 2014-2020.
A point-by-point response to reviewers’ comments are as follows, with original comments in blue and responses in black.
Responses to Reviewer #1
It is my pleasure to review remotesensing-1492225 “A Novel Approach to Validate Satellite Snowfall Retrievals by Ground-Based Point Measurements” by Jeoung et al. The authors first pointed out the insufficiency resulting from using either SNOTEL or GHCND data alone for validating CloudSat snowfall, then proposed a new method by combining SNOTEL and GHCND data together with accounting for the stations’ elevation information. As we know, snowfall is sensitive to global and regional climate change. I found there are three fatal errors in this manuscript:
- The CloudSat snow product ranging from 2006 to 2017 was used for annual mean snowfall estimates, however, the station mean data average during 2014-2020. The time ranges between these two datasets are different, which may make the results unbelievable, e.g. CloudSat CPR mostly underestimates snowfall against SNOTEL while having a mixture of over- and underestimation against GHCND observations.
The comment is fair and well taken. We reprocessed all the station data and switched the time period to 2006-2017, overlapping with CloudSat observations. A major drawback of using the new time period is a significant reduction of the number of GHCND stations (from 1764 to 456), while number of SNOTEL stations is almost the same. Conclusions are mostly identical to the earlier analysis.
- As mentioned in this manuscript, Shi and Liu (2021) found that the ratio of snowfall to precipitation occurrence has been increasing at high latitudes in the past 40 years. It seems that snowfall is increasing with the global warming. The mechanism of this increasing trend is complex, however, this manuscript ignored the impact of global climate change on snowfall using the dataset with different time ranges, and exaggerated the dependence of snow on elevation.
With switching to 2006-2017, the issue mentioned in this comment is solved. However, we did make a comparison of the snowfall between the two periods and showed it in a new Figure 13. There is no systematic change of snowfall in the averages over all the stations.
- Snowfall is random with time, it is insufficient to derive the snowfall pattern using 7 years observation data.
We added a discussion section, in which we attempted to argue that the difference between multi-year (7 or 10 years) means of snowfall is far less than the mean value themselves.
So, the authors may need to find more data with same time ranges to improve this novel method.
Thanks, we used data of 2006-2017 in the new version.
Author Response File: Author Response.docx
Reviewer 2 Report
The reviewer is really grateful to the authors for such an outstanding research and article about novel approach to validate satellite snowfall retrievals by ground-based point measurements. The proposed method for validation of satellite radar snowfall retrievals used surface station observations over the west United States mountainous region, where the mean snowfall rate at a station depends on its elevation. First, all station data within a 1°x1° grid were used to develop a snowfall rate versus elevation relation. This relation was then used to compute snowfall rate in other locations within the 1°x1° grid, mimicking that surface observations were available everywhere in the grid. Grid mean snowfall rates were then derived, which should be more representative to the mean snowfall rate of the grid than using data at any one station or from a simple mean of all stations in the grid. Comparison of the so-derived grid mean snowfall rates with CloudSat retrievals showed that the CloudSat product underestimates snowfall by about 60% when averaged over all the 773 grids in the west U.S. mountainous regions. The bias did not seem to have clear dependency on elevation, but strongly depends on snowfall rate. As an application of the method, authors further estimated the snowfall to precipitation ratio using both ground and satellite measured data. It was found that the rates of increase with elevation of the snowfall to precipitation ratio are quite similar when calculating from ground and satellite data, being about 25% per kilometer elevation up, or approximately 4% per every degree Cuisses of temperature drop.
Author Response
We thank the three anonymous reviewers taking their valuable time to review our paper. Their comments are well received, and a major revision was conducted based on these comments. Two major changes have been made in the revised manuscript. (1) The time period for the validation is switched to 2006-2017, overlapping CloudSat observations. (2) A discussion section is added to discuss the novelty of the proposed method and the difference of climatological means of snowfall between 2006-2017 and 2014-2020.
A point-by-point response to reviewers’ comments are as follows, with original comments in blue and responses in black.
Response to Reviewer#2:
The reviewer is really grateful to the authors for such an outstanding research and article about novel approach to validate satellite snowfall retrievals by ground-based point measurements. The proposed method for validation of satellite radar snowfall retrievals used surface station observations over the west United States mountainous region, where the mean snowfall rate at a station depends on its elevation. First, all station data within a 1°x1° grid were used to develop a snowfall rate versus elevation relation. This relation was then used to compute snowfall rate in other locations within the 1°x1° grid, mimicking that surface observations were available everywhere in the grid. Grid mean snowfall rates were then derived, which should be more representative to the mean snowfall rate of the grid than using data at any one station or from a simple mean of all stations in the grid. Comparison of the so-derived grid mean snowfall rates with CloudSat retrievals showed that the CloudSat product underestimates snowfall by about 60% when averaged over all the 773 grids in the west U.S. mountainous regions. The bias did not seem to have clear dependency on elevation, but strongly depends on snowfall rate. As an application of the method, authors further estimated the snowfall to precipitation ratio using both ground and satellite measured data. It was found that the rates of increase with elevation of the snowfall to precipitation ratio are quite similar when calculating from ground and satellite data, being about 25% per kilometer elevation up, or approximately 4% per every degree Cuisses of temperature drop.
Thank you for your recognition of the value of our work. The accuracy of snowfall retrievals over mountainous regions has been largely unknown. We hope that the newly proposed method and the validation results can shed some light on this problem and help future algorithm developments.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors can find the comments from the PDF file attached.
Comments for author File: Comments.pdf
Author Response
We thank the three anonymous reviewers taking their valuable time to review our paper. Their comments are well received, and a major revision was conducted based on these comments. Two major changes have been made in the revised manuscript. (1) The time period for the validation is switched to 2006-2017, overlapping CloudSat observations. (2) A discussion section is added to discuss the novelty of the proposed method and the difference of climatological means of snowfall between 2006-2017 and 2014-2020.
A point-by-point response to reviewers’ comments are as follows, with original comments in blue and responses in black.
Response to Reviewer#3:
Comments for Remote sensing - 1492225
I appreciate the author's efforts to assess CloudSat snowfall retrievals' accuracy in the western United State using ground-based snowfall measurements. Though the manuscript is interesting, it has severe shortcomings in terms of language, methodology and techniques. Below are some comments for your reference. I hope these comments act as constructive to improve the quality of the manuscript further.
Thank you for your constructive comments. Major revisions have been made based on these comments. Detailed responses are as follows.
- Interpolating data is not a new technique or novelty to be mentioned in the manuscript.
There are hundreds of papers that implemented interpolations in various domains, including in snow-related researchers. I suggest authors adopt advanced interpolation techniques (shepherd interpolation or machine learning - KNN or RF) to eliminate the uncertainty caused by the basic interpolation technique adopted in the study. Suppose the authors are confident that exponential technique is accurate. In that case, I suggest authors represent multiple interpolation techniques efficiencies compared with the current technique to add some novelty to the present work.
I think that there is a misunderstanding here about our proposed approach. Our approach is not a spatial interpolation to derive a downscaled distribution. But rather it is using the unique characteristic of snowfall increasing with elevation to derive a more meaningful areal mean in a 1deg x 1deg box. As shown in Figures 5 and 6, the snowfall dependence on elevation within a 1deg x 1deg box is well behaved, which allows us to fill the gaps using the snowfall-elevation relation. This relation is derived for each 1deg x 1deg box centered at a SNOTEL station; totaling 768 boxes (or 768 relation curves) in play. To the authors’ knowledge, we are the first (only) ones who proposed this approach. For deriving the snowfall-elevation relation in each 1deg x 1deg box, we do not see the necessity of using more advanced fitting method other than exponential curve fitting because there are less than 30 points (stations) in well-equipped areas and most of boxes have much less number of points. It is our fault not explaining this well, more texts are added in the revised version to clarify the method. In particular, a paragraph is added in the new discussions section to explain the novelty of the proposed method.
- The discussion is very vague. I suggest authors include an in-depth discussion comparing their results and commenting on the advantages and disadvantages, limitations and future scope of the current work.
A new discussions section is added for clarify two points: the novelty of the proposed method (including the difference between this method and an interpolation method), and the difference of mean snowfall rates between 2006-2017 and 2014-2020. The latter point is intended to argue that the difference between multi-year (7 or 10 years) means of snowfall is far less than the mean value themselves.
- As mentioned, interpolating and finding values to the nearest grid is not new and is a decade old. Hence, I suggest authors highlight the novelty of the current work, if any that adds to the existing knowledge of the scientific community.
This is added in the new discussions section. Thanks.
- The generation of 1 degree by 1 degree is significantly coarser. Will it be possible to develop a finer grid at the resolution of ERA-5 (0.25 degrees) for better comparisons?
No. We found that for CloudSat to have a meaningful mean using the over 10 years data, the minimum size of boxes needs to be 1deg x 1deg. Otherwise, there will be many boxes without sufficient data. Song and Liu (2021) found that even 1x1 is too small in the Tibetan region, so they used 1x2. For SNOTEL and GHCND combined data, to find a relation between snowfall and elevation, we also need a few tens of points. Boxes smaller than 1x1 are too small for most the locations.
- The major flaw in the study might be considering the averages of different datasets that span for 7 years (2014 - 2020) and 12 years (2006 - 2017). Though sometimes the average smoothens the peaks, it is not recommended to consider averages of different spans. The years 2006 - 2017 have experienced multiple drought and flood events that were missed during the mean calculation of SNOTEL. If authors are sure that the mean calculations are reliable, I suggest authors exhibit the differences that occur when you calculate mean for the same time periods (time frame common for all datasets) and different time periods (as implemented in the current work)
Good comments. We switch the time for ground stations to 2006-2017, overlapping with CloudSat observations. Number of GHCND stations in the new time period is dramatically reduced (from 1764 to 456), while number of SNOTEL stations is almost the same. The conclusions, however, are almost the same, which further confirms that the multi-year means don’t change much. If averaging all stations, snowfall changed between the two time period only 3% for SNOTEL stations and 0.4% for GHCND stations.
- Based on what statistic, are you determining the best fit? How do you choose exponential and stepwise curves for different grids?
As mentioned earlier, there are only less than 30 points (most much less) in the 1x1 box, and there are 768 (773 in earlier version) boxes. We just relied on the standard python function for the fitting, and then manually examined whether the fitting looks right. If one function does not fit the points right, we divide the data into two or three elevation ranges, and do the fitting in each range. Because of the small number of data points in each fitting, we think that it would not do much good even we use a more sophisticated statistic.
- Line 247 - 254 - The methodology is very unclear. Please kindly rephrase, elaborate and try to explain clearly
We rewrote some of texts to make it clearer.
- Are you creating 240 grids between each 1 degree by 1-degree spacing? The explanation is not clear in this content.
Yes. We couldn’t draw 240x240 subgrids in the schematic figure (Figure 5a); it will make the figure unreadable. We added some texts here and in the figure caption to make this point clearer.
- Line 10 – Western
Corrected.
- Line 33 - declining --> depleting
Corrected.
- Line 34 - Abbreviate U.S when it first appears in the manuscript
Done.
- Line 57 - 59. Would you please rephrase the sentence? It sounds more like a running
Language
This sentence is rewritten.
- Lines 71 - 74 - Very long sentences. Try to break it into two sentences.
Done.
- Line 77 - 78 - No need for this line. Would you please remove the sentence? Add it as an objective rather than mentioning it here
Removed.
- Line 201 - Please rephrase the initial words of the sentence
Changed.
- Line 247 - Incomplete sentence
Corrected.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The authors have responsed carefully to some of my comments. There is one more comment: time range is from 2006 through 2017, can it be extended to 2020?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
I appreciate the author's effort in successfully addressing the comments. Though the authors have addressed all the comments raised, the discussion was unsatisfactory. The written discussion sounds like a summary or conclusion rather than a discussion for the results. I strongly suggest authors read more scientific journals to understand the procedure of writing discussion and adopt the format. Please don't write the summary of the results but instead discuss your results with other studies. I barely see references (only one reference in the entire discussion) that were cited and discussed in the discussion section. The entire discussion was drawn from reference 48 (from manuscript). Please focus on discussing your results with other articles for improved discussion. Please address limitations and future scope in the manuscript that might help other researchers pursue similar related studies.
Author Response
Please see the attachment.
Author Response File: Author Response.docx