Using FengYun-3C VSM Data and Multivariate Models to Estimate Land Surface Soil Moisture

Round 1

Reviewer 1 Report

Summary:

This manuscript describes an interesting study on improving the FengYun-3C Microwave Radiation Soil Moisture product by combining this product with NDVI, location and elevation information. Promising results and conclusions are obtained. Improvements can be made on the introduction, better explanations of methods and results, figures, and general writing.

 

Broad comments:

1.       The introduction is very general and does not lead to the specific aim of this study. Also, the FY-3C VSM product was not discussed at all in the introduction until the aim is introduced. It would be advised to focus the introduction more on what is being done in this study, so for example discuss previous studies on the standard FY-3C VSM product, previous literature on improving soil moisture products with additional information and multivariate models, and from this can follow the research aim.  

2.       Several statistical terms are used incorrectly. For example, ‘precision’ is not a synonym of accuracy and Table 4 does not list ‘errors’ but error metrics or performance metrics. Please check ‘precision’, ‘accuracy’, ‘bias’, ‘error’ and possibly more, and use them according to their statistical definitions consistently throughout the manuscript.

3.       The format of the active references must be checked (Name et al. [1] produced), see e.g. Line 82, 87, 91.

4.       The title states that FengYun-3C/MWRI data is used. Indirectly this might be the case, but if I understand the manuscript correctly it is actually the FY-3C VSM product that is used as input to the multivariate models. Also, is the word ‘-based’ necessary here? (applies to the whole manuscript). Removing redundant wording in general would improve readability of the manuscript.

5.       The FY-3C satellite data is well explained, but the method underlying the FY-3C VSM product is very sparsely explained. How is the VSM product generated from the FY-3C satellite data?

6.       The multivariate linear regression method seems to be not explained in the manuscript.

7.       It is not clear how the back-propagation neural network model was implemented. For example, which software is used, which settings or assumptions are applied (if any?), which input layer combinations are tested. This should all be explained in the methodology sections. Furthermore, it would be nice to see a table which lists the combinations of layers that are used as input to the multivariate linear regression model and the back-propagation neural network model. Then it is very easy to see for readers what is for example MLR-2.

8.       Line 250 – 251: This is an interesting finding and the impact of vegetation seems a valid explanation. Can this explanation be confirmed by plotting the soil moisture deviations between FY-3C VSM and the references against the MODIS NDVI?

9.       Figure 6 and text: The lay-out of Figure 6 must be improved. First, please provide a y-axis label (and unit). Second, spelling mistake ‘longitute’. Third, does the ( ) after e.g. NDVI apply to the y-axis? This is not clear. Perhaps better to put this in the y-axis label. Fourth, what confidence intervals are plotted?

Furthermore, the text in Line 275 – 288 does contain a pure description of the figure, but more interpretation is required. For example, to me it is not clear whether the figure indicates that the correlation with soil moisture of e.g. elevation is much lower than that of NDVI (as the y-axis number are lower). It would be interesting to compare the results for the different variables. Also, try to explain the results. For example, the correlation between longitude and soil moisture could be expected based on Figure 2.

See some more specific comments below.

10.    Table 3 and text: I see very low values in Figure 6 and the lines are even intersecting zero, so I wonder if all the QR coefficients in Table 3 can be statistically different from zero. Also, does quantile 0.1 refer to the data between percentiles 0% and 20%? And quantiles of what? (soil moisture measurements?). Lastly, better explain the reason why elevation is not included. Because it is (only) not significant at the 0.9 quantile.

11.    The discussion is more of a summary instead of a discussion. The first three paragraphs of the discussion are repetitions of things discussed earlier in the manuscript and can be removed. Only the last paragraph is a good learning point for further research and can be kept in the manuscript, either as part of the results (and discussion) section or the conclusion section. Furthermore, an additional point of discussion could be that you depend on in situ soil moisture measurements for such data-driven approaches. This study highlights the importance of in situ measurements in addition to and to enhance satellite retrievals.

12.    Although the manuscript is reasonably well written, a thorough check would really help readability. In the specific comments below also a number of suggestions are provided (not complete).

 

Specific comments:

1.       Line 23 – 25: Sentence does not read well due to repetitions of SM and FY-3C VSM.

2.       Line 27: Not sure what the authors want to say with ‘To indicate a more detailed relationship …’. ‘To find more detailed relations …’?

3.       Line 40: Start new sentence after ‘FY-3C VSM’.

4.       Line 41: What is the accuracy over all the months? (also see specific comment 46)

5.       Line 41 – 43: This sentence can probably be removed.

6.       Line 52 – 55: A contrast between optical and microwave remote sensing is created, which I think is not really true as microwave remote sensing has generally been seen as most attractive for mapping soil moisture already for a long time. The optical remote sensing part can be removed. (also see broad comment 1)

7.       Line 57: It is the sensitivity to dielectric constant we are interested in (not surface roughness).

8.       Line 57 – 60: This can also be removed. The comparison between microwave/optical/thermal is not relevant here.

9.       Line 76: Add ‘for’ after ‘correct’.

10.    Line 76 – 79: Add the references to the specific studies after each solution.

11.    Line 79 – 82: ‘synergistic approach’ is rather vague; replace this by what they actually did in that study.

12.    Line 82 – 84: ‘adjust’ is ‘improve’ and ‘precision’ is ‘accuracy’? (also see broad comment 2).

13.    Line 85 – 87: ‘useful’ in what sense? Improving accuracy of soil moisture retrievals?

14.    Line 102 – 103: This sentence does not read well. Consider replacing by something like ‘The international soil moisture network (ISMN) measurement database was employed as the reference SM.’

15.    Line 105 - 107: Reformulate this sentence. ‘multivariate linear regression’ and ‘multivariate BP neural network’ are not just examples (which ‘such as’ suggests), but the models that were actually applied in this study.

16.    Line 116: What is meant by ‘at different layers’? Perhaps ‘at various depths’.

17.    Line 118: Link does not work. Also, shouldn’t this also be part of the reference list?

18.    Line 121 – 123: ‘In general,’ can be removed.

19.    Line 123 – 124: ‘in’ is not the correct preposition here I think. Consider ‘over’ and ‘on’.

20.    Line 136 – 137: Is a RMSE of 0.096 cm3/cm3 really such a high accuracy? It seems quite high compared to the range of soil moisture values.

21.    Line 140: This sentence can be removed.

22.    Line 167: Consider replacing ‘used approach for reflecting vegetation status’ by ‘used source for vegetation status’. Especially ‘approach’ is not the good word here. Also, the last part of the sentence is a bit confusing and could be understood as if the MODIS NDVI products are used to estimate soil moisture directly.

23.    Line 173: ‘and SM measurements’ can be removed because you want to align with the FY-3C VSM (which is why you also interpolated the SM measurements to this resolution).

24.    Line 174 – 176: What is the source for elevation? Furthermore, the ISMN values were also extracted pixel by pixel after the spatial interpolation was applied?

25.    Line 207 – 208: Which one was applied in this study?

26.    Line 213: It is not so nice to use abbreviations in section titles.

27.    Figure 4: Provide a reference if this figure is adopted from another study.

28.    Line 224 – 226: Perhaps it is easier to read if j and k are put in subscript? Also, should there be a k after the second ϴ?

29.    Line 230: Remove ‘Thus, ’. If I understand correctly, this sentence does not directly follow from the previous sentence.

30.    Line 233: Which error limit?

31.    Line 235 – Line 237: This sentence is not clear to me. Please explain more or remove if it is not relevant in the context of this study.

32.    Line 238: The use of ‘After that, ’ is not clear to me (after what?). Also, consider rephrasing this sentence in the context of your study, e.g. ‘the trained BPNN is used to map soil moisture based on the inputs provided’.

33.    Figure 5: ‘Sample’ is pixel number?

34.    Line 242: Are 150 pixels all the pixels available over the USA? Otherwise, why these 150 pixels? It would be nice to see a plot of how the 150 pixels (sample numbers) are distributed over the USA.

35.    Line 252: ‘and’ should be ‘but’ or ‘although’, I think.

36.    Line 263: I do not really understand the explanation ‘since there were lower biases in January and April’. Is that not exactly what the performance metrics indicate? Also, it could be interesting to add the unbiased root mean square error (uRMSE) to Table 2 (see numerous other soil moisture retrieval studies for the definition).

37.    Line 269 – 271: It would be nice to make this sentence more active about what you are going to do. E.g. ‘With an evaluation of the deviations between FY-3C VSM and reference VSM against multiple related factors using the QR model, we aim to correct for the bias and improve the SM inversion accuracy that was described in Section 3.1.’

38.    Line 271 – 274: Repetition of Section 2.2.1. Can be removed.

39.    Line 276: Intercept with/of what?

40.    Line 280 – 282: First, I do not understand how the explanation ‘which suggested … high soil water contents (wet)’ follows from the first part of the sentence and the previous sentence. Second, I wonder if this is true because this is not obvious in Figure 5.

41.    Line 307: How was mean relative error calculated? It would be good to give definitions of all the performance metrics in the form of equations (in methods section).

42.    Line 318 – 324: It would be nice to describe this in the text, possibly already in the methods sections and perhaps in table form (see broad comment 7).

43.    Line 332 – 335: Many sentences in the manuscript can be written simpler. In this case e.g. ‘The results indicate that the application of a nonlinear model with multisource input data, namely FY-3C VSM, MODIS NDVI and geographical location, is a suitable and promising approach for producing SM estimations with high accuracy.’ Remove parts of sentences that are already explained, unnecessarily complex or not relevant. Also, when describing results the present tense is more suitable as the results do still indicate that what you explain.

44.    Line 341 – 343: Where exactly? Because this over- and underestimation are not so obvious to me from Figure 8.

45.    Line 343 – 344: This sentence comes quite in between two parts of text that are related. Could you put this sentence in a different paragraph and extend the explanation (because it is also not so clear what you want to say with it, e.g. what is the reason for the missing pixels).

46.    Section 3.3.3: What are the performance metrics over all observations? (for all months combined)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript proposed a multivariate model for satellite soil moisture estimates. The authors employed quantile regressions to identify land surface variables that are sensitive to soil moisture retrieval accuracy and adopted a neural nets method to improve soil moisture estimates. The manuscript fits the scope of the journal, but the design of the analysis needs major improvement.

 

Major comments:

The authors discuss their results across different samples and across different seasons, e.g. Fig. 5, Table 2, Fig. 8 etc. This provides very limited insights regarding the performance of the proposed methods. The accuracy of microwave soil moisture retrieval algorithm is susceptible to a known set of error sources, e.g., RFI, steep topography, surface water body, snow cover, heavy precipitation, vegetation water content. The performance of the proposed methods should be established by evaluating how they improve soil moisture estimates given different error scenarios. To give a specific example, the performance of soil moisture retrieval is sensitive to the underlying land cover types. Dense vegetation (therefore large vegetation water content) is known to bias satellite soil moisture estimates. For this reason, the evaluation should be carried out for different land cover types. Do the proposed methods improve soil moisture estimate over both sparse and dense vegetation covers? The authors should also show time series that are representative for key land surface characteristics. See, e.g., (Colliander et al., 2017).

 

Another major concern is that it is unclear whether/how the authors define training and validation sets in the study. Without this information, it is difficult to evaluate the robustness of the analysis.

 

Detailed comments:

Line 41. Please clarify what “acceptable” refers to. Perhaps report the mean error statistics for these other months, including RMSE, R, etc.

 

Line 82. It is unclear what these two groups of numbers represent. Are both numbers RMSE? What are the analyzed domain and period?

 

Line 135-137. Please explain how these error estimates are derived.

 

Section 2.1.2 Please clarify if any quality controls are applied to the satellite data.

 

Line 242. Please clarify the criteria of choosing the samples.

 

Line 257. How about unbiased RMSE (ubRMSE), which is commonly reported in studies that evaluate soil moisture products.

 

Table 2. Are these metrics calculated for fixed pixels or across different pixels? For example, is R calculated based on soil moisture time series at a given pixel?

 

Line 300. Is this due to higher vegetation water content in summer?

 

Line 303. Does this approach have an apparent dependency on the seasons?

 

Section 3.3. Are there training and validation sets in this study? How are these sets chosen?

 

References:

Colliander, A., Jackson, T. J., Bindlish, R., Chan, S., Das, N., Kim, S. B., … Yueh, S. (2017). Validation of SMAP surface soil moisture products with core validation sites. Remote Sensing of Environment, 191, 215–231. https://doi.org/10.1016/j.rse.2017.01.021

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors did make significant efforts in replying to my suggestions and in improving the manuscript. However, regarding a few of my suggestions I still do not understand what the authors mean and in some cases the authors properly replied to my suggestion but did not included this in the manuscript (and an explanation why this was not included is missing).

 

Broad comment 7: Which tool did you use in Matlab? (e.g. some neural network package?)

Broad comment 8: The figure along with the explanation are a nice result. Perhaps you could put this figure in an appendix/supplement and refer to it in the manuscript (don’t forget units on the axes).

Broad comment 9:

• There are still no units on the y-axes. If there is no unit, then make this clear by putting (-) or (no unit) in the y-axis label.
• It is not clear which confidence intervals are plotted (e.g. 95%).
• The y-axes labels cannot be correct, as for example NDVI cannot extend below 0 (and have such low values). Probably the figures show another metric for these variables (a quantile regression parameters regarding e.g. NDVI?).

Broad comment 10:

• Please describe the definition of quantiles also in the manuscript.
• If quantile 0.1 refers to the data between percentiles 0% and 10%, and the quantiles 0.1, 0.3, 0.5, 0.7 and 0.9 were used, what happens with the data between percentiles 10% and 20%?
• Regarding elevation: if I understand correctly, elevation could provide information at the lower quantiles, so what not at least test it as an input variable for SM estimation?

 

Specific comment 17: Reference in the reference list is not complete. Check the format of website references.

Specific comment 24: I think that the elevation source should also be mentioned in the manuscript.

Specific comment 31: The lines 235 – 237 (lines 249 – 251 in the new version) are still not clear to me. Do you mean that ‘the training efficiency and accuracy using the GDM method aid/stimulate parameters such as the learning rate, iterations and momentum term’?

Specific comment 34:

• This description of how the sample pixels were selected must also be included in the manuscript.
• It would still be nice to see where the 150 pixels are located. If it is difficult to make such a figure and the 150 samples are from one row, then provide the row’s latitude.
• “The random samples can be used to indicate the whole characteristics of SM over the US”: based on what can you say this? (please also describe this in the manuscript)

Specific comment 39: I still don’t understand this. The intercept of the dependent variable with what?

Specific comment 45: What is the cause for the missing pixels? For example, are they already missing in the FY-3C VSM product (and why then? In Section 2.1.2) or did you filter them?

Author Response

Please see the attachment .

Author Response File: Author Response.pdf