Mapping of Soil Organic Carbon Stocks Based on Aerial Photography in a Fragmented Desertification Landscape

Round 1

Reviewer 1 Report

The English in the present manuscript is not of publication quality and requires significant improvement. My suggestion is to please carefully proofread spell check to eliminate grammatical errors.

For instance:
-- lines from 24 to 31: Our results could provide soil information with better patchiness and accuracy to help "manager identify" ??? the future LDN status in "a" fragmented desertification landscape. As UAV technology becomes more available, it will support "a" more refined prediction of soil properties.

or

-- line 46: "Land and soil degradation affects" the singular verb appears to disagree with the plural compound subject Land and soil degradation.

Moreover, the sentences are sometimes hard to read to a knowledgeable audience, and my advice is to split them into shorter sentences or use simpler synonyms.

These are but a few examples. More are in the manuscript.

The metrics from number 4 to number 7 are redundant and can be removed.

Section 2.6.3 can be eventually expanded, improving richer explanatory notes to the advantage of readers' comprehension.

Author Response

Responses to Reviewers 1

The English in the present manuscript is not of publication quality and requires significant improvement. My suggestion is to please carefully proofread spell check to eliminate grammatical errors.

Thank you for your efforts to improve our paper. We hope that our responses and the resulting changes will be acceptable, but we will be happy to work with you to resolve any remaining issues.

We have made extensive English language edits and you can see specific revisions in the marked version.

 

Q1: lines from 24 to 31: Our results could provide soil information with better patchiness and accuracy to help "manager identify" ??? the future LDN status in "a" fragmented desertification landscape. As UAV technology becomes more available, it will support "a" more refined prediction of soil properties.

A1: We revised it to “help policymakers determine the future LDN status in this fragmented desertification land-scape” (Line 28-29); “it will fully support digital mapping of soil properties”

 

Q2: line 46: "Land and soil degradation affects" the singular verb appears to disagree with the plural compound subject Land and soil degradation.

A2: We revised it to “Land and soil degradation affect” (Line 44).

 

Q3: Moreover, the sentences are sometimes hard to read to a knowledgeable audience, and my advice is to split them into shorter sentences or use simpler synonyms. These are but a few examples. More are in the manuscript.

A3: In the revision, we have noticed this problem and try to split long sentences into shorter sentences.

 

Q4: The metrics from number 4 to number 7 are redundant and can be removed.

A4: Yes, we have put the metrics from number 4 to number 7 in the supplementary material.

 

Q5: Section 2.6.3 can be eventually expanded, improving richer explanatory notes to the advantage of readers' comprehension.

A5: We have explained the abbreviations and adjusted the logic of the statement (Line 257-260)

Author Response File: Author Response.DOCX

Reviewer 2 Report

This study includes two steps in mapping the SOC stocks: 1) use aerial photography from UAV and NDVI values from MODIS to map the fractional vegetation cover (FVC) of the study area and reclassify the desertification degree, and 2) use the FVC map together with other environmental variables to map SOC stocks at two depth (0-40 cm and 0-100 cm) using regression kriging method. The issue raised in this study is worth investigating and the study fits into the scope of the journal. I found some questions that should be addressed.

 

Figure 1, (c) Locations of the soil samples and UAV photographs, do the blue, purple, and red dots indicate the locations where the UAV photographs were taken? Are the slight, moderate, severe and extremely severe types of these dots classified based on the 2010 desertification map or reclassified based on the calculated FVC?

Figure 2 only shows the flowchart of UAV flight settings and FVC extraction method. A comprehensive flowchart of the whole study involving FVC calculation and SOC stock mapping is needed to make the paper easier to follow.

Line 150, 138 UAV plots. Which method did you use to select the locations of the 138 plots?

Line 155, 231 field survey locations out of 1465 sampling locations. How did you select these 231 locations?

Line 203, “we produced a prediction map of the desertification degree in 2014 to 2017 based on the maximum FVC.” You only used the FVC map in SOC stocks prediction, what was the purpose to produce the desertification degree map? For comparison with the 2010 map or only for display the desertification degree in this study area?

Line 209, where did you download the ET data?

Line 229, “?(?) and ε′′(?) represent the predicted value and the unexplained variability”, the ?(?) should be measured value instead of the predicted value.

Line 231 to 235, the explanation of regression kriging method should be improved. The empirical Bayesian kriging regression was used (line 230), is this the regression part and residual kriging? Please provide more details about it. “combining empirical Bayesian kriging with explanatory raster variables” (ling 231-232), the same method? How did you get the standard error (line 231) of this model? “The result of this approach can provide more accurate standard errors in each pixel” (line 234), it should be “more accurate prediction” or “smaller standard errors”.

Line 238, can you also provide R² (coefficient of determination) of your model?

Figure 5, are the desertification categories in this figure from 2010 map or reclassified map?

Table 3, the ordinary kriging method should be explained in Materials and methods. For the first row, the ME is 0.001, but MSE is -0.007. I think they should be both either positive or negative.

Discussion 4.1., could you please better explain the relationship between environmental variables and SOC stock? For example, the Ap has a direct and negative effect on SOCD40, but it had a positive and indirect effect on SOCD40 through ET. How would you explain this? “spatial organic carbon density decreased with warming and dune height in desertified soil, because organic matter loss may accelerate with higher temperature and moisture” (line 354-356), could you please explain more about it? Why would SOC losses be increased with higher temperature and moisture? What does the R2c mean in Figure 7?

Discussion 4.2., does it mean SOCDSoilGrid250m overestimated more at the severe and extremely severe regions?

 

The writing should be improved, some minor issues:

Line 18-19, change “To improve the patchiness” to “To improve the patchiness identification”; change “accuracy of predicted results” to “accuracy of SOC prediction”; change “low-altitude aerial photography” to “collected low-altitude aerial photography”

Line 51, change “the series of maps of deserts and desertified land is” to “the series of maps of deserts and desertified land are”

Author Response

Responses to Reviewers 2

 

This study includes two steps in mapping the SOC stocks: 1) use aerial photography from UAV and NDVI values from MODIS to map the fractional vegetation cover (FVC) of the study area and reclassify the desertification degree, and 2) use the FVC map together with other environmental variables to map SOC stocks at two depth (0-40 cm and 0-100 cm) using regression kriging method. The issue raised in this study is worth investigating and the study fits into the scope of the journal. I found some questions that should be addressed.

Thank you for your efforts to improve our paper. We hope that our responses and the resulting changes will be acceptable, but we will be happy to work with you to resolve any remaining issues.

 

Tips: Dear reviewer, there are pictures in the Response letter. If it cannot be displayed normally, please download the file and check it.

 

Q1: Figure 1, (c) Locations of the soil samples and UAV photographs, do the blue, purple, and red dots indicate the locations where the UAV photographs were taken?

A1: The legend of our Figure 1C is confusing, so we have modified the legend of this figure to ensure clear information is conveyed.

The blue, purple and red dots are the soil sampling locations (Figure 1c). They are part of the sampling locations for soil surveys (grey dots) throughout the Horqin area. We did not select the grey points because according to the survey records and locations, they are all in non-desertification plots (e.g., farmland and woodland). In Section 2.4, we have made a corresponding description (Line 150-155).

 

Q2: Are the slight, moderate, severe and extremely severe types of these dots classified based on the 2010 desertification map or reclassified based on the calculated FVC?

A2: You must have noticed a mismatch between the degree of desertification indicated by the color of some points and that indicated by the patches. As you said, the desertification degrees of the basemap is from a 2010 database (Wang, T. Atlas of sandy desert and aeolian desertification in Northern China (In Chinese). Science Press: Beijing, 2014), and the color of the points is determined by the results from drone’s aerial photos.

 

Q3: Figure 2 only shows the flowchart of UAV flight settings and FVC extraction method. A comprehensive flowchart of the whole study involving FVC calculation and SOC stock mapping is needed to make the paper easier to follow.

A3: Thanks for your very pertinent suggestion. We have added Figure 2g to briefly express how the prediction of spatial FVC is done. Of course, this is far from your expectations. Please allow me to explain. Our Empirical Bayesian Kriging regression predictions are done in ArcGIS Pro by simply adding different variables raster in option box during the geostatistical wizard. We've explained the method in detail in Section 2.5 and 2.6.1, as you suggested, and attached a webpage about it. If this part is added to the flow chart, it will mainly be described in text, and our original flow chart is mainly expressed in graphics to assist readers' spatial imagination. On the other hand, there are already many Figures, and Figure 2 also occupies more space.

Hope our answer can satisfy you, if you insist on adding, we will make a flow chart.

 

 

Q4: Line 150, 138 UAV plots. Which method did you use to select the locations of the 138 plots?

A4: The drone photos were collected intensively over two years (2016 and 2017, after the launch of DJI Phantom 3), while the soil collection took several years. We tried our best to visit the original location where soil samples were collected, but it was not perfect due to human, material and financial constraints. Therefore, the UAV plots are partly at the location of soil sampling points and partly at what we consider typical desertification areas. We believe that even if the location of the drone photo and the soil sampling point are completely coincident, the FVC will vary greatly in different years due to the influence of environmental factors. We therefore use the average cover from 2014-2017.

 

Q5: Line 155, 231 field survey locations out of 1465 sampling locations. How did you select these 231 locations?

A5: Please combine Figure 1b and c and Line150-155 to understand. The gridded soil survey in the Horqin area initiated by the Naiman Desertification Research Station covers an area of nearly 120,000 square kilometers, including farmland, woodland, non-desertified grassland and desertified grassland, and wetlands, and so on. The basis of our research is the classification system of desertification degree, so we only selected the locations in desertified patches out of all gray points in Figure 1c. It combined with the records during the investigation and the base map of desertification (2010) by overlay analysis. This is how the 231 sample points come from.

Our predictions are completely unrelated to the grey points, but the accuracy of our results is much higher in desertified regions than in previous studies. This is the original intention of this study.

 

Q6: Line 203, “we produced a prediction map of the desertification degree in 2014 to 2017 based on the maximum FVC.” You only used the FVC map in SOC stocks prediction, what was the purpose to produce the desertification degree map? For comparison with the 2010 map or only for display the desertification degree in this study area?

A6: This sentence you extracted (section 2.5.2, lines 194-195) points to Figure 4b. Although it is an intermediate product, it is also one of the foundations of this study. First, our soil sampling sites and UAV plots are completely within the desertification patches in Figure 4b. Second, the color of soil sampling points in Figure 1c and the random effects of the SEM in the Discussion section are also determined based on Figure 4b.

We chose to reclassify the desertification map on the basis of the original desertification map in 2010, so that our results correspond to the period of 2014-2017 and thus be more rigorous. Because the latter version is produced on a completely quantitative basis.

 

Q7: Line 209, where did you download the ET data?

A7: Because ET and NDVI are both MODIS products, we put their acquisition methods together (lines 183-184). For more details, please refer to the website behind the product code.

 

Q8: Line 229, “?(?) and ε′′(?) represent the predicted value and the unexplained variability”, the ?(?) should be measured value instead of the predicted value.

A8: Yes, so we have modified it (Line 219) according to the reference (Keskin, H.; Grunwald, S. Regression kriging as a workhorse in the digital soil mapper's toolbox. Geoderma. 2018, 326, 22-41).

 

Q9: Line 231 to 235, the explanation of regression kriging method should be improved. The empirical Bayesian kriging regression was used (line 230), is this the regression part and residual kriging? Please provide more details about it. “combining empirical Bayesian kriging with explanatory raster variables” (ling 231-232), the same method? How did you get the standard error (line 231) of this model? “The result of this approach can provide more accurate standard errors in each pixel” (line 234), it should be “more accurate prediction” or “smaller standard errors”.

A9: We have made a thorough revision for Section 2.6.1 (Line 216-236) based on your suggestions. ArcGIS tools can calculate the statistics of Prediction Standard Errors, Quartile and Probability at the same time when predicting the spatial variable Z, as the following picture.

 

We reorganized Table 3 (Line 322-325). The prediction method used in this study is EBK Regression, and there are also EBK methods, and ordinary kriging methods that are more conventional. We made a comparison between the cross-validation of the three in Table 3.

 

 

Q10: Line 238, can you also provide R2 (coefficient of determination) of your model?

A10: ArcGIS Pro provides linear regression of measured and predicted values in the result, as in the following pictures, as well as regression function, but no specific R² values are given.

In order to answer your question, I extracted the corresponding measured and predicted values through overlay analysis, and performed linear regression to calculate R². It is slightly different from ArcGIS's calculation results. The picture above is 0-40 cm, and the picture below is 0-100 cm.

The four indicators (ME, RMSE, MSE and RMSSE) of cross-validation can comprehensively evaluate the cross-validation results (https://pro.arcgis.com/en/pro-app/latest/help/analysis/geostatistical-analyst/performing-cross-validation-and-validation.htm).

 

 

 

 

Q11: Figure 5, are the desertification categories in this figure from 2010 map or reclassified map?

A11: No, the desertification categories were based on Figure 4b, i.e., the reclassified map (Line 297-300). In particular, we changed Figure 5 from a bar chart to a box chart, so that the distribution of the data can be clearly seen.

 

Q12: Table 3, the ordinary kriging method should be explained in Materials and methods.

A12: We have added the description of ordinary kriging (Line 233-236), and also given a URLs. Given that this aspect is relatively basic and simple, and our research did not use ordinary kriging, we have not added too much space to this aspect.

 

Q13: For the first row, the ME is 0.001, but MSE is -0.007. I think they should be both either positive or negative.

A13: Thanks for your caution, we are aware of this issue. All error terms are derived by software automatic calculation tools, not manually calculated (the formula given is only to illustrate the calculation process. Reviewer 1 thinks the formula is redundant and asks us to delete it, we have put it in in the supplementary material). To compare the superiority of EBK regression prediction, we adjusted the Table 3 (Line 322-325). We have provided all the error term values for this interface in the pictures attached to Q10.

 

Q14: Discussion 4.1., could you please better explain the relationship between environmental variables and SOC stock? For example, the Ap has a direct and negative effect on SOCD40, but it had a positive and indirect effect on SOCD40 through ET. How would you explain this? “spatial organic carbon density decreased with warming and dune height in desertified soil, because organic matter loss may accelerate with higher temperature and moisture” (line 354-356), could you please explain more about it? Why would SOC losses be increased with higher temperature and moisture?

A14: We have reorganized the explanation of this section (Line 350-359). We hope the explanation will be acceptable.

 

Q15: What does the R²_c mean in Figure 7?

A15: This is the conditional R² based on the variance of both the fixed and random effects used in Piecewise SEM. This is just a representation to distinguish it from R² in traditional SEM. The explanation is in the caption of Figure 7. According to Lefcheck (Lefcheck, J.S. piecewiseSEM: Piecewise structural equation modelling in r for ecology, evolution, and systematics. Methods Ecol. Evol. 2016, 7, 573-9 and https://jonlefcheck.net/2014/07/06/piecewise-structural-equation-modeling-in-ecological-research/), “There are, however, a few shortcomings to traditional SEM. First, it assumes that all variables are derived from a normal distribution. Second, it assumes that all observations are independent. In other words, there is no underlying structure to the data. These assumptions are often violated in ecological research.” Regarding SEM, we also consulted with statistical professional and reflected in the acknowledgments.

 

Q16: Discussion 4.2., does it mean SOCDSoilGrid250m overestimated more at the severe and extremely severe regions?

A16: It is definitely a serious overestimation, and it does not follow the law of changing with the desertification gradient, we provide Table S1, you can refer to the supplementary material.

 

Q17: The writing should be improved, some minor issues:

Line 18-19, change “To improve the patchiness” to “To improve the patchiness identification”; change “accuracy of predicted results” to “accuracy of SOC prediction”; change “low-altitude aerial photography” to “collected low-altitude aerial photography”

Line 51, change “the series of maps of deserts and desertified land is” to “the series of maps of deserts and desertified land are”

A17: We have revised them and made extensive English language edits and you can see specific revisions in the marked version.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have done a good job revising the manuscript. I found a minor issue: line 51, it should be "latest" instead of "lates".