High-Resolution Digital Soil Maps of Forest Soil Nitrogen across South Korea Using Three Machine Learning Algorithms
Round 1
Reviewer 1 Report
Review of the paper “ High-resolution digital soil maps of forest soil nitrogen across South Korea using three machine learning algorithms “ by Yoosoon An, Woojin Shim, and Gwanyong Jeong.
General comment
The researchers used national soil data to map forest soil total nitrogen concentration in two soil horizons (A and B) by testing three machine learning techniques. Among them, the RF technique yielded the best predictions for both the two soil horizons. Geographic variables, elevation, forest type and slope curvature were among the most important variables explaining soil N concentration variation across the South Korea.
I found that the methods followed were appropriate, well used, and clearly defined. The results are clear and their interpretation is not far-fetched at all.
However, the article could be greatly improved by filling some missing information that would help the reader to appreciate the outcome of such study.
Specific comments
- In the M&M, a general description of South Korea climate would be warranted as no climate variable was used in the DSM exercise, and climate plays a major role in N accumulation and cycling.
In the M&M, a brief description of the forest types encountered in South Korea would be welcome, as forest types are among the main variables explaining soil N variation. What are the main tree species forming deciduous, mixed, and coniferous forests in South Korea?
- In the M&M, a brief description of the forest soil types found in South Korea, for instance according to the Great Soils Groups according to the IUSS World Reference Base for Soil resources (2015, see Table 2 in page 10-11), for example, soils with pronounced accumulation of organic matter in the mineral topsoil, soils, soils mainly dominated by Fe/Al chemistry, etc., would help the reader to figure out what really are made the A and B horizons.
- It is mentioned that for the N17-FHM survey, the 0-10 cm soil depth was used as the A horizon. What is the average depth of the A horizon in the national forest soil map data?
- Similar interrogation regarding soil depth of the B horizon. Can the authors give a least a depth range?
- In Figure 7, maps of the N concentrations in forest soils are somewhat misleading, mainly because we do not know what the forest cover extent in this country is. An example: are the white areas low in soil N or is just because there is no forest there?
- In the supplementary materials, I would prefer see maps of the standard deviation or the prediction error associated with the RF model instead of the prediction given by the two other AI models. Mapping errors can help to find regions where there is a need for further soil sampling or finding covariates that are related to the error pattern.
- In the discussion, the authors mentioned that soil N concentrations slightly increased between the two sampling periods in the A horizon. Improvement of ecological forest health is given as an explanation, but this is not very convincing. Could the increase in N concentrations in the forest topsoil be due to increased atmospheric N deposition reaching >20 kg N/ha-year in Asia as reported by Vet et al. (2014)? Or was there a change in forest type, in forest management activities, in natural disturbances events?
- Partial plots of the most important variables (can be added to the Supplementary materials) would help to understand the trend in N variation associated with these variables, for instance what is the trends of N concentrations in the A horizon associated with DISTC4 (distance from the lower right of the country)? Positive or negative?
References
IUSS Working Group WRB, 2015. World Reference Base for Soil Resources 2014, update 2015. International soil classification system for naming soils and creating legends for soil maps. FAO. World Soil Resources Reports No. 106. ROME. 192 p.
Vet, R., R.S. Artz, S. Carou, M. Shaw, C. Ro, Un, W. Aas, A. Baker, V.C. Bowersox, F. Dentener, Galy, C. Lacaux, A. Hou, J.J. Pienaar, R. Gillett, M.C. Forti, S. Gromov, H. Hara, T. Khodzher, N.M. Mahowald, S. Nickovic, P.S.P. Rao et N.W. Reid, 2014. A global assessment of precipitation chemistry and deposition of sulfur, nitrogen, sea salt, base cations, organic acids, acidity and pH, and phosphorus. Atmos. Environ. 93(0): 3-100.
English quality is OK with me.
Author Response
Point 1: In the M&M, a general description of South Korea climate would be warranted as no climate variable was used in the DSM exercise, and climate plays a major role in N accumulation and cycling.
In the M&M, a brief description of the forest types encountered in South Korea would be welcome, as forest types are among the main variables explaining soil N variation. What are the main tree species forming deciduous, mixed, and coniferous forests in South Korea?.
Response 1: We have added a new section (2.1) that provides a brief description of South Korea's climate, topography, geology, forest types, and forest soil properties related to soil nitrogen. Specifically, we included information on the main tree species forming deciduous, mixed, and coniferous forests in South Korea, as these forest types are among the main variables explaining soil nitrogen variation. We also explained South Korea's unique forest soil classification system, which is highly related to geology and topography. Additionally, we included Figure 1 to help readers interpret Section 2.1 and to provide information on the geographical settings and the region mentioned in the following sections of the manuscript. We believe that these additions will give readers a better understanding of the characteristics of the study area and the research topic.
Point 2: In the M&M, a brief description of the forest soil types found in South Korea, for instance according to the Great Soils Groups according to the IUSS World Reference Base for Soil resources (2015, see Table 2 in page 10-11), for example, soils with pronounced accumulation of organic matter in the mineral topsoil, soils, soils mainly dominated by Fe/Al chemistry, etc., would help the reader to figure out what really are made the A and B horizons.
Response 2: Thank you for your suggestion. As mentioned in our added section 2.1, South Korea has a unique forest soil classification system that is not fully compatible with international standards. Nevertheless, we provided information on the major soil classes in South Korea's forest soil and their similarities to soil classes in the FAO/WRB or USDA system based on previous research by Kim (2020) and Park et al. (2010). We hope that this information will help international readers understand the soil properties of South Korea and the characteristics of the A and B horizons in this study.
Point 3-4: It is mentioned that for the N17-FHM survey, the 0-10 cm soil depth was used as the A horizon. What is the average depth ofthe A horizon in the national forest soil map data?
Similar interrogation regarding soil depth of the B horizon. Can the authors give a least a depth range?
Response 3-4: We mentioned the mean and standard deviation of A and B horizon soil depth from N17-FHM survey data in our added section 2.1. This information could fulfill your these suggestions.
Point 5: In Figure 7, maps of the N concentrations in forest soils are somewhat misleading, mainly because we do not know what the forest cover extent in this country is. An example: are the white areas low in soil N or is just because there is no forest there?
Response 5: Thank you for your suggestion. We completely agree with the possibility of misinterpretation of Figure 7. In our revised manuscript, we have added a legend to clarify that the white areas are "Nodata" mainly because they are not forests. The updated figure can be found in Figure 8 of the revised manuscript. We hope this clarification will prevent any potential misinterpretation of the data.
Point 7: In the supplementary materials, I would prefer see maps of the standard deviation or the prediction error associated with the RF model instead of the prediction given by the two other AI models. Mapping errors can help to find regions where there is a need for further soil sampling or finding covariates that are related to the error pattern.
Response 7: Yes, estimating the uncertainties of the digital soil map is crucial. As such, we made to sure to include this information in our manuscript. For quantifying prediction error, we utilized the resampling method, which yielded a total of 10 models. The standard deviation of 10 predictions from the models was used to account for spatial uncertainty.
Point 8: In the discussion, the authors mentioned that soil N concentrations slightly increased between the two sampling periods in the A horizon. Improvement of ecological forest health is given as an explanation, but this is not very convincing. Could the increase in N concentrations in the forest topsoil be due to increased atmospheric N deposition reaching >20 kg N/ha-year in Asia as reported by Vet et al. (2014)? Or was there a change in forest type, in forest management activities, in natural disturbances events?
Response 8: Thank the reviewer for providing valuable suggestions, as well as for the insightful references. What you mentioned should have been reflected in our discussion. We included this information in the manuscript.
Point 9: Partial plots of the most important variables (can be added to the Supplementary materials) would help to understand the trend in N variation associated with these variables, for instance what is the trends of N concentrations in the A horizon associated with DISTC4 (distance from the lower right of the country)? Positive or negative?
Response 9: The variable importance calculated by the random forest algorithm is agnostic to the directionality of the relationship between soil properties and environmental variables. Thus, the partial dependence plot suggested by the reviewer can certainly aid in interpreting the relationships. However, in our case, we relied on the directionality inferred from the digital soil maps of soil nitrogen. We do appreciate the reviewer’s suggestion, but due to time constraints, we were unable to conduct further analyses. We hope you understand our situation.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Dear authors,
You have used three ML methods for the prediction and digital mapping of soil nitrogen across South Korea. The manuscript has an acceptable structure, with clear objectives and methodology. However, it suffers from major drawbacks which should be considered before any further processing:
1- The main issue with this manuscript is its novelty. There are several studies that have used ML algorithms for the prediction and mapping of nitrogen and other soil properties.
2- There is no discussion on why RF showed the highest performance on the prediction of nitrogen, while DNNs results were not that acceptable.
3- I didn't understand why the authors put Table 5 in the manuscript. How this table can help in reasoning of the obtaining results?
4- The third objective of this study (i.e. the relation between spatial variations of forest soil nitrogen and environmental variables) has not been discussed enough.
5- Half of the conclusion is dedicated to methodology, while the conclusion should mostly be brief results.
6- Table 1: Porphyry and Felsite are rock textures, while others are names of the rocks. The name of the rocks with porphyry texture should be mentioned if possible.
The quality of the English language is fine.
Author Response
Point 1: The main issue with this manuscript is its novelty. There are several studies that have used ML algorithms for the prediction and mapping of nitrogen and other soil properties.
Response 1: Thank you for your thoughtful comments on our study. We appreciate your feedback and would like to address your concern about the originality of our work.
Our study focuses on predicting soil nitrogen content distribution by directly utilizing and integrating geographic variables into a machine learning model. We agree that there have been previous studies that have attempted to predict soil distribution using machine learning models, but we believe our study makes a unique contribution to the field by taking a different approach to incorporating geographic variables.
Specifically, our approach directly reflects the spatial autocorrelation on soil distribution, which is a departure from the indirect approach of regressing kriging that has been commonly used in the field. By incorporating geographic variables into our model, we are able to more accurately capture the spatial relationships between variables and soil distribution. We believe this aspect of our study makes it both novel and significant.
We understand that our study may not be completely original in its goals, but we hope you will agree that our focus on incorporating geographic variables directly into a machine learning model represents a unique and valuable contribution to the literature.
In addition, we have provided evidence of the improved accuracy and realism of our model through validation and testing, which we believe further supports the originality and significance of our work. We hope that this information will help to clarify our study's contributions and address your concerns.
Point 2: There is no discussion on why RF showed the highest performance on the prediction of nitrogen, while DNNs results were not that acceptable.
Response 2: Thank you for taking the time to review our manuscript and providing your valuable feedback. We appreciate your insightful comments and would like to address your concern that there is no discussion on the comparison of RF and DNNs.
We agree for not providing a detailed discussion on why RF showed higher performance than DNNs in our study as you correctly pointed out. However we would like to emphasize that our study was not primarily focused on comparing the performance of different machine learning models. Rather, our main goal was to examine the distribution of soil forecasts at the national scale using field-measured data and new variables, and to gain a deeper understanding of the environmental factors that control nitrogen distribution in forest soils. In other words, we want to say that the comparison between the three models was not a major issue.
Certainly, We recognize that the performance of machine learning models can vary depending on the characteristics of the target variable, and we agree that it is important to consider the purpose and ease of use of each algorithm in practical applications. Therefore, we have made modifications to our manuscript to better highlight the strengths and limitations of RF and DNNs, and to provide a clearer explanation of why we judged RF to be more suitable for our study. Once again, thank you for your constructive comments.
Point 3: I didn't understand why the authors put Table 5 in the manuscript. How this table can help in reasoning of the obtaining results?
Response 3: Thank you for your feedback and for bringing up the issue regarding the necessity of Table 5 in our manuscript.
Table 5 may not be essential for the main argument of our manuscript. However, we are of the opinion that the information presented in this table offers valuable insights on two distinct points.
Firstly, it helps validate the reliability of the nitrogen content data in forest soil that we used for our distribution prediction. As you know Nitrogen content in soil changes slowly over time. Based on these characteristics, we aimed to prove the reliability of our data by comparing data collected for the same area but at different times. Actually, Table 5 demonstrates that the difference between the historical data from 1984-1990 and the data from 2009-2021 is not significant for both the A and B layers, which indicates the reliability of the data.
Secondly, as we mentioned in our Response 2, there are implications for forest resource management. It confirms that the total nitrogen content of the surface layer (A) in South Korean forest soils has improved ecologically with a slight increase in the horizon in Table 5. Additionally, the Coefficient of Variance (CV) shows that the variability of nitrogen content within the same layer has increased compared to the past, providing crucial information for managing forest resources.
Nevertheless, there's still room for improvement. We have added a comprehensive discussion of nitrogen concentrations within the forest soil, achievable through the comparative of historical and contemporary data.
Point 4: The third objective of this study (i.e. the relation between spatial variations of forest soil nitrogen and environmental variables) has not been discussed enough.
Response 4: Thank you for your's concern about the discussion of the spatial distribution of nitrogen content in forest soils and its relation to environmental variables. We appreciate the opportunity to address this point.
In our study, we did indeed examine the spatial distribution of nitrogen content in forest soils and its relationship to various environmental variables. Specifically, in section 3.3, we presented the findings from the digital soil maps, which revealed that the A layer exhibited high nitrogen content in areas characterized by high elevation, convex slopes, andesite rock, as well as in high-elevation deciduous forests. Moreover, the discussion section (section 4) further explores the correlation between environmental variables such as geographic variables, elevation, topographic openness, and curvature, and the distribution of nitrogen. We also specifically discussed the relationship between altitude and soil nitrogen content. In addition, in section 3.2, we tried to deal with that by presenting environment variables with high explanatory power for each machine learning model.
Thank you for bringing this to our attention, and we appreciate the opportunity to address this concern in our manuscript. We apologize if there was any confusion or lack of clarity in our manuscript that may have contributed to the reviewer's perception that this objective was not sufficiently discussed. We will ensure that these points are emphasized more prominently in the revised version of the paper to provide a clearer and comprehensive understanding of the spatial distribution of nitrogen content in forest soils and its relationship with environmental variables.
Point 5: Half of the conclusion is dedicated to methodology, while the conclusion should mostly be brief results.
Response 5: We greatly appreciate the feedback regarding our concluding section. We agree with the reviewer's perspective that the conclusion should provide a concise summary of the research findings and effectively convey their significance to readers. We made a deliberate effort to adhere to the typical structure of a conclusion in our manuscript.
Given the comprehensive scope of our research, which encompasses a national scale and involves the generation and comparison of three machine learning models, incorporating more than 20 variables, we acknowledge that certain aspects related to methodology, including the discussion of variables, may appear prominent in the conclusion.
Nonetheless, taking your feedback into account, the conclusions have been refined by significantly reducing the description of the data and succinctly adjusting the explanatory capacity of the variables.
Round 2
Reviewer 1 Report
The authors did a good job in response to the reviewers' comments. This paper is now excellent with all the necessary details for such study. I appreciate the authors could response positively to my comments.
I recommend acceptation of the msfor publication once the authors checked Table 5 which seems missing in the copy I got for review.
Author Response
Point 1: The authors did a good job in response to the reviewers' comments. This paper is now excellent with all the necessary details for such study. I appreciate the authors could response positively to my comments.
I recommend acceptation of the msfor publication once the authors checked Table 5 which seems missing in the copy I got for review.
Response 1: Thank you sincerely for the positive rating and your valuable feedback. Your opinion has been immensely helpful in shaping the content of our work.
The incorrect location of table5 has been corrected. We greatly appreciate your attention to detail and the opportunity to address any concerns you may have had. Your input has undoubtedly enhanced the quality and reliability of our work.
Reviewer 2 Report
Dear authors,
Thanks for addressing my comments.
Please consider my following minor comments:
Line 15: The sentence should be " Reliable estimation of the forest soil nitrogen spatial distribution is necessary for effective forest ecosystem management"
Line 511: The sentence should be "revealing a high soil nitrogen content..."
Regards
Author Response
Point 1: Line 15: The sentence should be " Reliable estimation of the forest soil nitrogen spatial distribution is necessary for effective forest ecosystem management"
Line 511: The sentence should be "revealing a high soil nitrogen content..."
Response 1: Thank you sincerely for the positive rating and your valuable feedback. Your opinion has been immensely helpful in shaping the content of our work.
We have taken your valuable suggestions into account and made the necessary corrections to address the issues in Line 15 and Line 511.
We greatly appreciate your attention to detail and the opportunity to address any concerns you may have had. Your input has undoubtedly enhanced the quality and reliability of our work.
