Refining Spatial and Temporal XCO2 Characteristics Observed by Orbiting Carbon Observatory-2 and Orbiting Carbon Observatory-3 Using Sentinel-5P Tropospheric Monitoring Instrument NO2 Observations in China
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe study reconstructed XCO2 based on methods using ML (CatBoost, LightGBM, and MLP) from multiple data sources and evaluated their performance. Additionally, the authors proposed incorporating NO2 data as tracers to reconstruct high-resolution XCO2 data to better capture the spatiotemporal characteristics of anthropogenic CO2 emissions. The research methods are reasonable and the analysis is comprehensive. However, there are still some shortcomings, as follows:
1. Please explain the physical significance of the variables chosen for modeling.
2. Line94: "TROPMI" is misspelled.
3. Line245: Missing "°".
4. Lines 243-246: The author suggests that the precision of the 0.5° grid resolution is higher than that of 0.1° due to errors introduced by resampling. This argument is debatable and requires further explanation. This is because, compared to coarse-resolution models, high-resolution models incorporate more refined data (multiple 0.1° resolution input data) and should thus yield more accurate modeling results under the assumption of homogenous data.
5. Lines 243-246, 597-598, 663-665: The author compared the modeling accuracy of various machine learning methods, which is beneficial for the reliability of the research. However, some key issues still need clarification:
(1) The author did not specify the exact parameter settings for the LGB, MLP, and CatB models or whether the parameters were optimally chosen, making it impossible to determine which model is optimal.
(2) Both LGB and CatB are boosting algorithms, while MLP is a neural network algorithm. The modeling results of MLP are heavily influenced by the network structure settings and cannot achieve optimal solutions. Therefore, a more advanced structured deep learning framework should be used to demonstrate this.
(3) The SHAP values based on MLP show significant differences from LGB and CatB in prediction results, requiring an explanation of the reasons rather than just stating the results.
Author Response
We are very grateful for your positive comments and constructive suggestions. We revised the manuscript according to your comments item-by-item. The responses to your comments associated changes in the revised manuscript are presented in the response document.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors propose a method to improve the CO2 emissions that arise from the OCO-2 and OCO-3, establishing a correlation between the NO2 TROPOMI sensor aboard the Copernicus Sentinel 5P and CO2.
I found the idea interesting and relevant to produce side products from data that may be available for different species.
Some comments:
In the abstract it is stated “high spatial resolution”. This is not true according to satellite classifications. Both OCO-2 (1.29 km × 2.25 km) and OCO-3 (2.25 km x 0.7 km) have large pixels and even Sentinel 2 with 10m resolution is not high spatial resolution.
With some acronyms, the authors should be very careful in the writing component. For example, “qa_value < 0.8.” does not mean anything for the reader.
Paragraph of lines 154-179 should be rewritten. The OFFL and GEE are stated after being used. It is also very important for the authors to understand that 5P is not a European Space Agency satellite; neither the data is provided by ESA. The data that they report are from the Copernicus ecosystem. Also the GEE with a resolution of 1.1132 km looks strange.
Eq. 1 is straightforward and does not bring anything new to discussion aside from the index confusion (sub-scripts that after are inline).
The major flaw in the manuscript is the machine learning algorithm. Is not clear at all what was the ML algorithm employed, and the major different data (different time span, height columns, among other particularities) can be accommodated in the algorithm.
On page they discuss CatBoots, Light-GBM and MLP, but no additional features of the models are described.
Also in the abstract and the discussion, it is stated that the total deviation is 0.17+-1.17 ppm and 1.03+-1.15ppm. These values have a large error associated but no discussion is provided.
The authors have established a correlation between NO2 and CO2, but some additional references are missing.
Overall, I found the subject relevant but the manuscript must be presented in better way, even for reproducibility and use by others in the future
Comments for author File: Comments.pdf
Many typos and difficult to understand some parts of the manuscript.
Author Response
We are very grateful for your positive comments and constructive suggestions. Additionally, thank you very much for your detailed revision suggestions and comments on the manuscript. We revised the manuscript according to your comments item-by-item in your revised manuscript. The responses to your comments associated changes in our revised manuscript are presented in the response document.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsIn fact, the author made substantial revisions to the article based on the provided feedback, primarily including language corrections, the addition of necessary experiments, and their explanations. The current issue is that the author provided a brief response to our question 1 but did not cite the necessary references. We hope the author can add the required citations. Overall, the author's revised manuscript has basically addressed the modifications as we suggested.
Author Response
We are very grateful for your positive comments and constructive suggestions. We revised the manuscript according to your comments. The responses to your comments associated changes in the revised manuscript are presented in the response document.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsAlthough the authors improved the manuscript I still find it difficult to understand the ML model that they have developed. “We develop an ML-based method of reconstructing XCO2 from the satellite observations using the multiple parameter variables to generate monthly XCO2 data in 0.1° grid.” This sentence does not bring anything new to the discussion. It states that they have used Eq.1 to fill gaps “by computing the average value of variable each grid while MXCO2 is resampled by the nearest neighbor method.” Unfortunately the authors must properly describe the method, using equations, for example, so it becomes clear where the ML algorithm that they proposed. The answer to my previous comment about statistical errors should also be reanalyzed. The authors state bias, which I agree with, but the associated error seems too large for the useful. Probably they should use different metrics to show the quality of the results.
The manuscript still required extensive english corrections, besides minor typos, many sentences must be re-written. Some examples.
Line 20: “pf” should read “of”.
Line 20: “The prediction model” is not correct “the predicting capabilities of the model”
Line 20: “The prediction model shows well predictive”
Line 23: “the introduction of NO2”, rephrase
Line 26: “the NO2 data as NO2…”
Line 91: “CO2” 2 should be subscript
Comments on the Quality of English LanguageA extensive english should be carried out by the authors or by the editor.
Author Response
We are very grateful for your positive comments and constructive suggestions. We revised the manuscript according to your comments item-by-item. The responses to your comments associated changes in the revised manuscript are presented in the response document.
Author Response File: Author Response.pdf
