Multispectral Remote Sensing Data Are Effective and Robust in Mapping Regional Forest Soil Organic Carbon Stocks in a Northeast Forest Region in China
Round 1
Reviewer 1 Report
To start with, I would like to say that the general topic of the manuscript is interesting and its practical aspect is to be prized.
Besides the comments that are listed below, I would like to stat my general opinion. I do agree with the authors that a 30 m sampling is not satisfactory for this kind of works. However the biggest issue that I have is that the 30 m for all the attributes is deliver in different ways and not all of them are described properly (see comment below). I would also like to know how was the general uncertainty of the data introduced in to calculations for example STRM data is 30 m but ‘for SRTM missions designed accuracy was as follows: 16 m absolute vertical height accuracy, 10 m relative vertical height accuracy and 20 m absolute horizontal accuracy (Wessling, 1999)’. With the use of Kriging (and I know it is a know method but I think Mr Krieg deserves a citation) and other interpolation/extrapolation methods implemented in ArcGIS some uncertainties should be given. If this was not used, or decided that this information is not significant enough this should be written.
Is it possible to compare this model with other results? (4.2), some published data on either modeling or other samples etc., this would significantly improve the paper, even if this was a descriptive rather than numerical comparison.
Comments
‘The data can have default values and the data types are flexible.’ – this sentence needs further information
‘ The output independent variable contribution and response curve are relatively intuitive, easy to explain, and can also be used as a prediction model,’ – better description needed
I do not know if this is possible but figure 1 and 2 should be larger - the scale is almost invisible in this version
Since the data were obtained from different departments, we resampled them to raster data with a 30m spatial resolution in ArcGIS 10.2. – specify the method or algorithm, what that validated, how were the edges of the area defined? In general how was it done since this is an important initial data for further analysis
Landsat-8 imageries were generated from nine images downloaded from the United States Geological Survey (USGS), which require cloud cover of less than 10% from July to September 2016. – I know what the authors want to say but this needs to be stated clearly, so someone that did not work with those data can understand
In general, the language is clear, and the text is understandable, but it could use a little help. For example.
It is a challenging task to accurately predict the SOC stock in the region. However’Also”?, due to the high cost for data collection, it is difficult to inventory SOC by intensive sampling. With the development of science and technology, remote sensing (RS), global position system (GPS) and geographic information system (GIS) have opened up new approaches for the study of SOC stocks [7]. – The science and technology is not needed it brings nothing new to the sentence. The traditional research on carbon cycle is limited with/by a small number of measuring points The soil collected by Hill and Schütt [14] was taken back to the laboratory for spectral analysis, which/it was found that the characteristics of 0.35 μ m-1.4 μ m spectral band determined the SOC content. The study area is located in Lushun District (38.72°–38.97°N, 121.08°–121.47°E)), Liaoning Province, Northeast China (Figure 1), and the southernmost tip/part/point? of Liaodong Peninsula. NDVI is widely used in the detection of vegetation growth state, and the closer its value is to 1, the better the vegetation growth. Surveying and mapping, only medium and small scale map data can/could be produced in the past. The authors should decide if they want to write 1km or 1 km – this is for all used units BGREEN is a better predictor than 84 NDVI and SAVI in the full variable model, although many scholars considered NDVI was a more 85 effective and powerful predictor [35,37]. – so what is it ?Author Response
Comments and Suggestions for Authors
To start with, I would like to say that the general topic of the manuscript is interesting and its practical aspect is to be prized.
Response: We appreciate your help and your patience. With this submission, we provided a version (marked) of the revised manuscript. Responses to reviewers’ comments on the manuscript are detailed below.
Besides the comments that are listed below, I would like to stat my general opinion. I do agree with the authors that a 30 m sampling is not satisfactory for this kind of works. However the biggest issue that I have is that the 30 m for all the attributes is deliver in different ways and not all of them are described properly (see comment below). I would also like to know how was the general uncertainty of the data introduced in to calculations for example STRM data is 30 m but ‘for SRTM missions designed accuracy was as follows: 16 m absolute vertical height accuracy, 10 m relative vertical height accuracy and 20 m absolute horizontal accuracy (Wessling, 1999)’. With the use of Kriging (and I know it is a know method but I think Mr Krieg deserves a citation) and other interpolation/extrapolation methods implemented in ArcGIS some uncertainties should be given. If this was not used, or decided that this information is not significant enough this should be written.
Response: Based on your comments, we have added this part of the uncertainty analysis in Section 4.2 of the manuscript. L433-435, 440-442.
Is it possible to compare this model with other results? (4.2), some published data on either modeling or other samples etc., this would significantly improve the paper, even if this was a descriptive rather than numerical comparison.
Response: Based on your comments, we have cited some publications to compare with our results. L446-455.
Comments
“The data can have default values and the data types are flexible.”– this sentence needs further information
Response: We have changed to “BRT model can be flexible to deal with linear, polynomial, exponential, logistic, periodic, or general nonlinear problems.”. L92-93.
“The output independent variable contribution and response curve are relatively intuitive, easy to explain, and can also be used as a prediction model,” – better description needed
Response: We have modified to “As a data mining method, BRT model has many advantages, including its limited number of user-defined parameters and ability to model nonlinear relationships. It also can effectively manage both of qualitative and quantitative variables. The model is also robust when there are absence of data and outliers and reduces over fitting”. L94-98.
I do not know if this is possible but figure 1 and 2 should be larger - the scale is almost invisible in this version
Response: Based on your comment, we have modified the figures 1 and 2. See figures 1 and 2.
Since the data were obtained from different departments, we resampled them to raster data with a 30m spatial resolution in ArcGIS 10.2. – specify the method or algorithm, what that validated, how were the edges of the area defined? In general how was it done since this is an important initial data for further analysis
Response: We used a bilinear interpolation method to resample these data to raster format with a 30m spatial resolution in ArcGIS 10.2. The bilinear interpolation method takes the distance from the sampling point to the surrounding 4 neighborhood pixels to calculate the grid value. The grid value of the pixel is obtained by interpolation in Y direction (or X direction) and then in X direction (or Y direction). The boundary cleaning and main filtering tools were used to generalize the edges of the regions in the grid. According to the value of the neighborhood in each position, the edge was smoothed by expanding and contracting the boundary, or increasing or reducing the area. L173-180
Landsat-8 imageries were generated from nine images downloaded from the United States Geological Survey (USGS), which require cloud cover of less than 10% from July to September 2016. – I know what the authors want to say but this needs to be stated clearly, so someone that did not work with those data can understand.
Response: According to your comment, we have changed to “Landsat-8 imageries were acquired from nine one image (Path-row: 120-33) downloaded from the United States Geological Survey (USGS) covering the spatial domain of study area from July to September 2016. In order to ensure image quality and avoid the interference of cloud on model prediction, the cloudiness value error of the downloaded image data is controlled within less 10%.” L167-172
In general, the language is clear, and the text is understandable, but it could use a little help. For example.
Response: According to your comments, we have invited professional English people to polish the whole manuscript to avoid such problems.
It is a challenging task to accurately predict the SOC stock in the region. However’Also”?, due to the high cost for data collection, it is difficult to inventory SOC by intensive sampling.
Response: Based on your comments, we have changed “however” to “also”. L53
With the development of science and technology, remote sensing (RS), global position system (GPS) and geographic information system (GIS) have opened up new approaches for the study of SOC stocks [7]. – The science and technology is not needed it brings nothing new to the sentence.
Response: Following your comment, we have deleted the useless phrase "science and technology". L55
The traditional research on carbon cycle is limited with/by a small number of measuring points
Response: We replaced “with” with” by”. L60
The soil collected by Hill and Schütt [14] was taken back to the laboratory for spectral analysis, which/it was found that the characteristics of 0.35 μ m-1.4 μ m spectral band determined the SOC content.
Response: We changed “which” to “it”. L74
The study area is located in Lushun District (38.72°–38.97°N, 121.08°–121.47°E)), Liaoning Province, Northeast China (Figure 1), and the southernmost tip/part/point? of Liaodong Peninsula.
Response: We have replaced “tip” with “part”. L119
NDVI is widely used in the detection of vegetation growth state, and the closer its value is to 1, the better the vegetation growth.
Response: We have changed “it” to “its value”. L190
Surveying and mapping, only medium and small scale map data can/could be produced in the past.
Response: We have replaced “can” with “could”. L202
The authors should decide if they want to write 1km or 1 km – this is for all used units
Response: We checked the entire manuscript and revised it to "1 km". L216
BGREEN is a better predictor than NDVI and SAVI in the full variable model, although many scholars considered NDVI was a more effective and powerful predictor [35,37]. – so what is it ?
Response: We have added “Therefore, BGREEN should be recommended as an influential factor in the future studies, especially in the areas with dense vegetation coverage” in the manuscript. L237-238
Author Response File: Author Response.docx
Reviewer 2 Report
Wang et al. used 12 predictor variables comprising topographic, climatic and Landsat-derived spectral dataset to run three boosted regression tree (BRT) models to predict soil carbon stocks (SOC) in vegetated environments of Liaoning Province in Northeast China. Observed SOC stocks for their BRT models were based on 236 field plots purposively sampled from the study area. Based on four validation metrics, the authors found that the full model containing all the 12 variables performed best but also noted that the BRT model comprising only the Landsat-derived spectral data performed nearly well as the full model.
Comments
While this is an interesting study indicating how multispectral remotely sensed data can be used to map an important ecosystem attribute as SOC, it raises several questions below.
Introduction
Authors failed to provide line numbers from the beginning, making the manuscript difficult to review. In second paragraph: What is “3S”?Methods
Section 2.3.1: What level of Landsat data did you download? What path-rows do your Landsat data cover? What method of atmospheric correction did you use? If indeed you needed atmospheric correction, why didn’t you use Landsat data with atmospheric correction already provided by the USGS? In what units were your Landsat data before and after your atmospheric correction? In equation 2: Let readers know what I and L represent?Section 2.3.3: What period (years) does your climate data represent?
Section 2.4: On what basis did you determine the optimal parameter settings for your BRT models?
Equations numbering: You have equations numbered 1 to 2, then 5 to 8. No equations 3 and 4?
Format for references in the text are inconsistent in the introduction and in several places in the manuscript (annotations mixed with other formats). Correct these citations in accordance with the journal recommended format.
Results
How did you ensure that pixels geometrically match location of sample sites whose summaries are provided in Table 1? In Table 1: How can SAVI values be zero throughout? In Table 2: What level of significance does the asterisks (* or **) correspond to? Section 3.1: Second paragraph - How can MAP be positively correlated and in the next sentence become negatively correlated with the same variable, SOC?Section 3.2: Line # 5 – 8: Sentence is incorrect with the facts. Accuracy values do not respectively match the order in which the models were mentioned in early part of the same sentence.
Authors should also indicate relative importance of model C. It will be interesting to see relative importance plot of the spectral data in model C.
Section 3.4: Line # 65: Figure 6 should read Figure 7.
Discussion
Line # 94: Tanzan??? Do you mean Tanzania?Authors are advised to fix several of such spelling and grammatical errors which are present throughout the manuscript.
Line # 101: What is DSM? Line # 102: Spell out RI. Several readers may have forgotten what it is at this point.Author Response
Comments and Suggestions for Authors
Wang et al. used 12 predictor variables comprising topographic, climatic and Landsat-derived spectral dataset to run three boosted regression tree (BRT) models to predict soil carbon stocks (SOC) in vegetated environments of Liaoning Province in Northeast China. Observed SOC stocks for their BRT models were based on 236 field plots purposively sampled from the study area. Based on four validation metrics, the authors found that the full model containing all the 12 variables performed best but also noted that the BRT model comprising only the Landsat-derived spectral data performed nearly well as the full model.
Response: We appreciate your help and your patience. With this submission, we provided a version (marked) of the revised manuscript. Responses to reviewers’ comments on the manuscript are detailed below.
Comments
While this is an interesting study indicating how multispectral remotely sensed data can be used to map an important ecosystem attribute as SOC, it raises several questions below.
Response: Thank you for your professional comments and significant improvement to our manuscript. The specific responses are as follows.
Introduction
Authors failed to provide line numbers from the beginning, making the manuscript difficult to review. In second paragraph: What is “3S”?Response: In this revision, we have provided the full names for these acronyms. “3S” are referring to remote sensing (RS), global position system (GPS) and geographic information system (GIS). In order to avoid confusion, we added these to the manuscript in this revision. L57
Methods
Section 2.3.1: What level of Landsat data did you download?Response: The downloaded data is the level of Landsat data (L1T level). Data products obtained are from geometric precision correction of radiometric correction data (using ground control points and digital elevation model data). We added “the downloaded data are L1T. “L1T Landsat-8 imageries” to L167 in this revision.
What path-rows do your Landsat data cover?
Response: The path-row is 120-33. For clarity, we added this information to the manuscript. L167
What method of atmospheric correction did you use?
Response: We used the ENVI FLAASH method to correct the data. See L171-172 in this revision.
If indeed you needed atmospheric correction, why didn’t you use Landsat data with atmospheric correction already provided by the USGS?
Response: Landsat 8 surface reflection code (LaSRC) is an atmospheric correction program specially designed by USGS for Landsat 8 data. It is considered to be the most accurate Landsat 8 OLI atmospheric correction program, which was recently updated to December 2017. LaSRC has a lot of technical improvements, including the application of short wave blue light and the use of day-to-day atmospheric environment auxiliary data. However, USGS only provides LaSRC products in the United States, and most of the landsat8 data provided overseas are L1T or L1TP. Thus, we used LaSRC for atmospheric correction.
In what units were your Landsat data before and after your atmospheric correction?
Response: The unit of Landsat data before and after your atmospheric correction is digital number.
In equation 2: Let readers know what I and L represent?
Response: Based on your comments, we have added this description in the manuscript. L197-198
Section 2.3.3: What period (years) does your climate data represent?
Response: The period was from 1980 to 2010, we have added it to the manuscript. L215
Section 2.4: On what basis did you determine the optimal parameter settings for your BRT models?
Response: The optimal parameter combination was the one that provided the minimum predictive deviation. We have added this to the manuscript. L237-238
Equations numbering: You have equations numbered 1 to 2, then 5 to 8. No equations 3 and 4?
Response: We have corrected this mistake. L249-252
Format for references in the text are inconsistent in the introduction and in several places in the manuscript (annotations mixed with other formats). Correct these citations in accordance with the journal recommended format.
Response: According to your comments, we have checked and revised the manuscript. L101, 220, 228
Results
How did you ensure that pixels geometrically match location of sample sites whose summaries are provided in Table 1?Response: In the field, we recorded the sampling points by a hand-held GPS, and then imported the point information into ArcGIS 10.2 in the laboratory, and then extracted the pixel value by using “Spatial Analyst Tools—>Extraction—>Extract Value to Points” in ArcGIS 10.2. In order to avoid this ambiguity, we have modified the table name to make it clearer. L285-286
In Table 1: How can SAVI values be zero throughout?
Response: We appreciate your help and your patience. We have corrected this mistake. See Table 1.
In Table 2: What level of significance does the asterisks (* or **) correspond to?
Response: p < 0.05 shown in “*”; p < 0.01 shown in “**”. See table 2.
Section 3.1: Second paragraph - How can MAP be positively correlated and in the next sentence become negatively correlated with the same variable, SOC?
Response: SOC stocks were negatively correlated with MAT (r=-0.35), and we've corrected that. L269
Section 3.2: Line # 5 – 8: Sentence is incorrect with the facts. Accuracy values do not respectively match the order in which the models were mentioned in early part of the same sentence.
Response: According to your comment, we have modified this incorrect sentence. L303
Authors should also indicate relative importance of model C. It will be interesting to see relative importance plot of the spectral data in model Cs.
Response: Based on your comments, we have added this part to the manuscript. L352-358, Figure 6b
Section 3.4: Line # 65: Figure 6 should read Figure 7.
Response: We have modified “Figure 6” with “Figure 7”. L364
Discussion
Line # 94: Tanzan??? Do you mean Tanzania?Response: There is Tanzania, we have corrected it. L395
Authors are advised to fix several of such spelling and grammatical errors which are present throughout the manuscript.
Response: We have revised the whole manuscript by professional English speakers.
Line # 101: What is DSM? Line # 102: Spell out RI. Several readers may have forgotten what it is at this point.
Response: DSM is “digital soil mapping”. RI is “relative importance”. We have revised the whole manuscript according to your comments. L402-403
Round 2
Reviewer 1 Report
Some of my previous comments were not included in this version for example the comparison with results of similar studies for this area. However, I am generally happy with all additional information that was given and I think this manuscript looks much better than it did previously. Therefore I will suggest publication
Comments, to fix for final version:
Line 53 - calling 3S a technology is a stretch – I would just remove this word all together
Line 133 - images still have non-readable descriptions, also they all should be in English (or at least have both Chinese and English). The two images on the left have scales that I cannot read.
Author Response
Comments and Suggestions for Authors
Some of my previous comments were not included in this version for example the comparison with results of similar studies for this area. However, I am generally happy with all additional information that was given and I think this manuscript looks much better than it did previously. Therefore I will suggest publication
Response: We appreciate your help and your patience. We have added the research results of this area for comparison. L437-445
Comments, to fix for final version:
Line 53 - calling 3S a technology is a stretch – I would just remove this word all together
Response: Following your suggestion, we have removed this word. L57
Line 133 - images still have non-readable descriptions, also they all should be in English (or at least have both Chinese and English). The two images on the left have scales that I cannot read
Response: We have revised the figure and descriptions. L128-130
Author Response File: Author Response.docx
Reviewer 2 Report
I appreciate the authors effort to improve the manuscript by addressing the review comments. However, their responses regarding their use of the Landsat data and the image pre-processing they claimed to have performed raise more questions than answers. In one instance they mentioned that they performed atmospheric correction on Landsat 8 Level 1 data using FLAASH. In another instance they mentioned that they used Landsat 8 data that had been generated using the Land Surface Reflectance Code (LaSRC). To be clear, LaSRC generates a surface reflectance Landsat 8 imagery which is a Level 2 science product. These obvious inconsistencies raise doubt about their image processing methods. In any case an image that has been atmospherically corrected is not in DN values. As Landsat data is a key data for this paper, authors have the responsibility to clarify these inconsistencies and make the necessary corrections in their methods before this manuscript may be considered.
Also check formatting of figures. For instance, some parts of Figure 2 are missing.
Author Response
Comments and Suggestions for Authors
I appreciate the authors effort to improve the manuscript by addressing the review comments. However, their responses regarding their use of the Landsat data and the image pre-processing they claimed to have performed raise more questions than answers. In one instance they mentioned that they performed atmospheric correction on Landsat 8 Level 1 data using FLAASH. In another instance they mentioned that they used Landsat 8 data that had been generated using the Land Surface Reflectance Code (LaSRC). To be clear, LaSRC generates a surface reflectance Landsat 8 imagery which is a Level 2 science product. These obvious inconsistencies raise doubt about their image processing methods.Response: We appreciate your help and your patience. This is our error in explanation. In our manuscript, we used the ENVI FLAASH method to correct the atmosphere of remote sensing data. In your previous comments, you pointed out why we didn't use Landsat data with atmospheric correction already provided by USGS. Thus, in the revision, we pointed out that only the United States has been provided with the data that had been generated using the Land Surface Reflectance Code (LaSRC), while other regions were only provided with the data of L1T or L1TP, so we need to carry out atmospheric correction. It is not that we use both methods to correct the atmosphere. Hope this is clear now. L168-169
In any case an image that has been atmospherically corrected is not in DN values.
Response: Thank you for your professional comments. We checked the manuscript to avoid these mistakes.
As Landsat data is a key data for this paper, authors have the responsibility to clarify these inconsistencies and make the necessary corrections in their methods before this manuscript may be considered.
Response: We carefully examined the methods of the manuscript to avoid such similar mistakes. L168-169
Also check formatting of figures. For instance, some parts of Figure 2 are missing.
Response: We have checked all the figures and tables in the manuscript to avoid this error. See the figure 2.
Author Response File: Author Response.docx