Forest Height Mapping Using Feature Selection and Machine Learning by Integrating Multi-Source Satellite Data in Baoding City, North China

Round 1

Reviewer 1 Report

This shows the futility and lack of utility for extensive exercises in empiricism when envisioning global goals. The lack of coherent threads among the multiplicity of methods and unimpressive performance parameters should discourage similar strategies for substantial sectors or subsets of earth environments.  There is confirmation that physiographic, bioclimatic, and cultural conditions are crucial for complex terrain; therefore, underscoring zonation and stratification strategies as opposed to hoping that machine learning will lessen logistics of ecological essentials. Since there are lessons to be learned from absence of advantage, if realistically recognized, these could be conveyed in a considerable condensed version.  Note-- "Filed vs Field" in Figure 2.  Heading of 2.3 contains a non-word.

Author Response

Point: This shows the futility and lack of utility for extensive exercises in empiricism when envisioning global goals. The lack of coherent threads among the multiplicity of methods and unimpressive performance parameters should discourage similar strategies for substantial sectors or subsets of earth environments.  There is confirmation that physiographic, bioclimatic, and cultural conditions are crucial for complex terrain; therefore, underscoring zonation and stratification strategies as opposed to hoping that machine learning will lessen logistics of ecological essentials. Since there are lessons to be learned from absence of advantage, if realistically recognized, these could be conveyed in a considerable condensed version.  Note-- "Filed vs Field" in Figure 2.  Heading of 2.3 contains a non-word.

Response: 

We gratefully thank you for your time making your constructive remarks and useful suggestions on our manuscript. We revised the manuscript according to your each suggested revision and comment. Based on the instructions provided, we uploaded the file of the revised manuscript. Accordingly, we uploaded a copy of the original manuscript with all the changes highlighted by using the“Track Changes” function in Microsoft Office Word. Listed below are the responses to your comments, and the baseline for the reference of the line and page is the revised version.

1. We agree with your viewpoint. Because most machine learning models are black-box models, they are difficult to reflect the mechanism and process between forest parameters and remote sensing information, and the interpretability for reality is weak. The improvement of the accuracy of forest parameter estimation by simply constructing empirical models is indeed limited. Physical geography, bioclimatic and cultural conditions are crucial for the estimation of forest parameters under complex terrain. Therefore, in subsequent studies, we will focus on zoning and stratification strategies or coupling remote sensing data and forest physiological process models to estimate forest parameters. We added this content to the section of Discussion 4.7 Limitations and Prospects in our article. (Line 737-744, page 23)
2. We are very sorry for neglecting this issue. We changed the “Filed” to Field" in Figure 2 (Figure 2, page 4) and “process” to “preprocessing” in the heading of 2.3 (Line 165, page 5).

Author Response File: Author Response.docx

Reviewer 2 Report

1. Some figure of the article are difficult to read. Please use vector graphic or higher analysis images. 

2. If it is possible, please include some comparizons with the folowing recent published works:

[1] Huang, W.; Min, W.; Ding, J.; Liu, Y.; Hu, Y.; Ni, W.; Shen, H. Forest height mapping using inventory and multi-source satellite 804 data over Hunan Province in southern China. For. Ecosyst. 2022, 9, 100006.

[2] Liu, Y.; Gong, W.; Xing, Y.; Hu, X.; Gong, J. Estimation of the forest stand mean height and aboveground biomass in northeast 806 China using SAR Sentinel-1B, multispectral sentinel-2a, and DEM imagery. ISPRS J. Photogramm. Remote Sens. 2019, 151, 277–807 289.

3. Explain better the contribution of this work. 

Author Response

Response to Reviewer 2 Comments

We gratefully thank you for your time making your constructive remarks and useful suggestions on our manuscript. We revised the manuscript according to your each suggested revision and comment.Based on the instructions provided, we uploaded the file of the revised manuscript. Accordingly, we uploaded a copy of the original manuscript with all the changes highlighted by using the“Track Changes” function in Microsoft Office Word. Listed below are the point-to-point responses to your comments, and the baseline for the reference of line and page is the revised version.

Point 1. Some figure of the article are difficult to read. Please use vector graphic or higher analysis images.

Response 1: We are so sorry for our negligence on this issue. We improved the quality of the images used in our articles.

Point 2. If it is possible, please include some comparisons with the following recent published works:

[1] Huang, W.; Min, W.; Ding, J.; Liu, Y.; Hu, Y.; Ni, W.; Shen, H. Forest height mapping using inventory and multi-source satellite 804 data over Hunan Province in southern China. For. Ecosyst. 2022, 9, 100006.

[2] Liu, Y.; Gong, W.; Xing, Y.; Hu, X.; Gong, J. Estimation of the forest stand mean height and aboveground biomass in northeast 806 China using SAR Sentinel-1B, multispectral sentinel-2a, and DEM imagery. ISPRS J. Photogramm. Remote Sens. 2019, 151, 277–807 289.

Response 2: Thank you for your advice. We mainly compared our work with these two works from four aspects: data source, methods, results, similarities, and differences. We condensed the following content and added it to our article. (Line 700-719, page 23) 

1. Data source: Huang et al. employed multi-spectral, radar, and topo variables from Landsat-8, Sentinel-1, and ALOS/PALSAR-2 imagery and three types of in-situ measured tree height data (maximum-, averaged-, and basal area-weighted-tree heights) to estimate three types of forest height. Liu et al. used SAR Sentinel-1 backscattering, vegetation coverage (FVC) variables from multispectral Sentinel-2 images, and field measurement data of forest stand mean height. In our study, the remote sensing data we utilized included Sentinel-1, Sentinel-2, ALOS PALSAR-2, and SRTM DEM, the field data we employed was the National Forest Resources Continuous Inventory (NFCI) data.

2. Methods: Huang et al. evaluated the model parameters and model performance of different models by dividing the extracted feature variables into ten scenarios and using two types of variable selection methods (progressive regression and random forest) and three types of models (multiple linear regression, random forest, and support vector regression). Then, the adjustment model was applied to generate three types of wall-to-wall forest height maps in Hunan Province. Liu et al. used the FCV derived from Sentinel-2 combining the VV and VH backscattering coefficient from Sentinel-1 respectively as the input variables and estimated the FSMH by constructing logarithmic regression models between the input variable and FSMH. Our study designed five data source scenarios and utilized three feature selection methods (stepwise regression analysis (SR), recursive feature elimination (RFE), and Boruta) and six machine learning algorithms (k-nearest neighbor(k-NN), support vector machine regression (SVR), random forest (RF), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost)) to construct forest height estimation models. And we used ANOVA to quantify the effects of three factors, including data source, feature selection method, and modeling algorithm on forest height estimation.

3. Results: Huang et al. achieved the best estimate using a random forest model (R² range from 0.47 to 0.52, RMSE range from 3.8 to 5.3 m, and rRMSE range from 28% to 31%). In Liu’s study, the combination of FVC biophysical variables and backscattering values in VH polarization achieved the result with R² = 0.53414 and RMSE =2.9156m, whereas the FVC and VV polarization backscattering values are characterized by R² = 0.42683 and RMSE = 3.1754m. In our study, the variables selected based on Boruta including Sentinel-1, Sentinel-2, and topography metrics, combined with the XGBoost algorithm provided the optimal model (R² = 0.6701, RMSE = 2.1924m). And According to the ANOVA results, data source, feature selection method, and modeling algorithm all had significant influence on forest height estimation, and data source was the most influential factor.

4. Similarities and differences: The similarity is that these three studies estimating forest height by constructing an empirical model between forest height and multi-source remote sensing information. The difference is that Liu et al proposed a methodological framework to estimate forest height by constructing a simple logarithmic regression based on backscattering coefficients derived from Sentinel-1 data and FVC derived from Sentinel-2 data, while Huang and our study both extracted considerable feature variables and employed different feature selection method and regression algorithms to estimate forest height. Huang's study focused on assessing the significance of multiple spectra, radar metrics on forest height, the impact of different algorithms and different land cover classification products on different forest height types, whereas our study concentrated on exploring the effects of different data sources and their combinations, variable selection methods, and machine learning algorithms on forest height estimation.

Point 3. Explain better the contribution of this work.

Response 3: Thank you for underlying this issue. We adjusted our description of the contribution in the Abstract section. And the modified content is “Overall, we proposed a methodological framework for quantifying the importance of data source, feature selection method, and machine learning algorithm in forest height estimation, and it was proved to be effective to estimate forest height by using freely accessible multi-source data, advanced feature selection method and machine learning algorithm. These findings provide a feasible approach for subsequent national or global scale high-precision forest height mapping.” (Line 35-40, page 1)

Author Response File: Author Response.docx

Reviewer 3 Report

The paper is interesting, very well written. Improvements are needed in some sections especially in the section that concerns the comparisonv between estimated heights and reference map. It is necessary to make this comparison on the same forest mask.

Comments for author File: Comments.pdf

Author Response

Point： The paper is interesting, very well written. Improvements are needed in some sections especially in the section that concerns the comparisonv between estimated heights and reference map. It is necessary to make this comparison on the same forest mask.

Response : First of all, thank you very much for your recognition and affirmation of our work. Secondly, we gratefully thank you for your time making your constructive remarks and useful suggestions on our manuscript. We revised the manuscript according to your each suggested revision and comment. Based on the instructions provided, we uploaded the file of the revised manuscript. Accordingly, we uploaded a copy of the original manuscript with all the changes highlighted by using the“Track Changes” function in Microsoft Office Word. Listed below are the point-to-point responses to your comments, and the baseline for the reference of line and page is the revised version.

Point 1: One digit after the comma for RMSE and two for R

Response 1: Thank you for underlying this issue. We changed the digit of R² and RMSE throughout the manuscript.

Point 2: The discription “LiDAR data is often limited in spatial and temporal coverage” is inaccurate because GEDI it is not completely true.

Response 2: Thank you for underlying this issue. The sentence was modified to “almost all LiDAR data is often limited in spatial and temporal coverage.” (Line 62, page 2)

Point 3: The statement “Compared with LiDAR data, SAR can penetrate the tree crown and obtain information about the structure and function of the forest” is not true and please discuss the limits of SAR and optical data.

Response 3: Thank you for underlying this issue. We modified this phase to “Compared with LiDAR data, although SAR and optical data have limited ability to penetrate the canopy or even cannot penetrate the canopy, the backscattering coefficient of SAR and the rich spectral information of the optical data can also reflect the information about the structure and function of the forest.” (Line 67-70, page 2)

Point 4: Not clear the difference between the planned forest land area and the forest area : 987000ha and 590000 ha! the actual surface is 590000 and the future surface will be 987 000?

Response 4: Your understanding is correct. “The planned forest land area” means the area of forest land planned by the government and not yet reached currently. “The forest area” means the current real forest area. In order not to cause misunderstanding, we deleted the description of the planned forest land area.

Point 5: Please give additional useful information for forest: the range of canopy hight and biomass.

Response 5: Thank you for underlying this issue. We tried to find information on forest canopy height and biomass in Baoding city, but unfortunately, there is no corresponding information. But we found the information about Baoding's forest stock and we added it to the manuscript.(Line 140, page 3)

Point 6: The quality of Figure 1 is bad ==> lat/long is not visible, and please replace the false color image by a land-use/cover map or with a map with two classes: (1) forest areas, (2) no forest areas.

Response 6: Thank you for underlying this deficiency. We adapted Figure 1 according to the comment. (Figure 1, page 4)

Point 7: About Figure 2: Please put lines horizontally to clearly distinguish the blocks of each step. The english of the sentence "C/L band SAR backscatters coefficients dual-polarized indices" is not correct, similar remark for the second sentence. The spell of “Field” is not correct.

Response 7: Thank you for underlying this deficiency. We adapted Figure 2 according to the comment. (Figure 2, page 4)

Point 8: Please give the date (year / years) of field data collection.

Response 8: Thank you for underlying this issue. The date of field data collection is consistent with the date of the field survey in November 2016. (Line 168-169, page 5)

Point 9: Do instead of what you suggest cross validation k-fold: 5-fold for example. It is more relevant to do a cross validation: 5-folds for example

Response 9: Thank you for underlying this issue. We understand and agree with your view. Cross-validation is indeed a commonly used validation method in a small sample dataset. However, to our knowledge, hold-out validation is also used by some researchers in small-sample studies. For instance, Pham et al.[1] combined 105 field plot data and multi-source remote sensing data to estimate the aboveground biomass of mangroves in the Red River Delta region using the hold-out validation method. Wai et al.[2] also employed hold-out validation method to validate the estimate results for evergreen and deciduous forest aboveground biomass with 88 field plots data of evergreen forest and 170 field plots of deciduous forest when they used Sentinel-2 data to estimate aboveground biomass of evergreen and deciduous forests, respectively. Furthermore, due to time constraints, it is difficult for us to use 5-fold cross-validation to complete the relevant experiments in a limited time. Therefore, we are so sorry that we are afraid not to do a 5-fold cross-validation according to your request. However, in the follow-up research work, we will complete and improve the content of this part.

Reference:

• Pham, T.D.; Yokoya, N.; Xia, J.; Ha, N.T.; Le, N.N.; Nguyen, T.T.T.; Dao, T.H.; Vu, T.T.P .; Pham, T.D.; Takeuchi, W. Comparison of Machine Learning Methods for Estimating Mangrove Above-Ground Biomass Using Multiple Source Remote Sens. Data in the Red River Delta Biosphere Reserve, Vietnam. Remote Sens. 2020, 12, 1334.
• Wai, P.; Su, H.; Li, M. Estimating aboveground biomass of two different forest types in Myanmar from sentinel-2 data with machine learning and geostatistical algorithms. Remote Sensing 2022, 14, 2146.

Point 10: Why the use of Sentinel-2 data only one date? ==> discuss this point (not logical!) Why in august 2016 not in another month / year?

Response 10: Thanks for your question. There are two main reasons why we only used the Sentinel -2 data on August 2016.

First, why we only use one date of Sentinel 2 data is because we got seven scenes with very good quality that can cover the area of Baoding at that date. Moreover, considering that too much remote sensing data will reduce the work efficiency in the subsequent work, we selected these seven Sentinel-2 images.

Second, why in august 2016 not in another month/year is that we need to match the collection time of our sample plot data, which was conducted in November 2016. And we considered that the spectral information of plants is more obvious in the growing season from July to September, so we screened the images of Sentinel-2 from July to September in 2016, and selected the seven images with good quality that can completely cover the entire study area on August 2016.  (Line 197-200, page 5)

Point 11: Why the use of Sentinel-1 data only one date? ==> discuss this point (not logical!)Why in October 2016 not in another month / year?

Response 11: Thanks for your question. This reason is similar to Point 9. To match the time of sample plot data collection, we screened Sentinel-1 data from July to December 2016 and selected 10 good-quality images that can cover the area of Baoding, which were captured on October 2016. (Line 212-213, page 5- 6)

Point 12: Please clarify and discuss why you are not using multi-temporal data/indexes for S2 and S1

Response 12: Thank you for underlying this issue. First, we do this to match the collection time of our filed data. Second, considering that the data needs to be preprocessed such as atmospheric correction, too much data will make the preprocessing time of S1 and S2 very long, which will affect our work efficiency. Third, we referred to previous studies showing that it is feasible to estimate forest height using S1 and S2 imagery of a specific time. For example, Liu et al.[1] used Sentinel-2 images (July 05, 2017) and Sentinel-1 images (June 30, 2017, and July 05, 2017) to estimate forest stand mean height in northeast China, and the R² and RMSE was 0.53 and 2.92m, respectively. Lastly, we respect and recognize your opinion very much. In future work, we will consider multi-temporal remote sensing data.

Reference:

• Liu, Y.; Gong, W.; Xing, Y.; Hu, X.; Gong, J. Estimation of the forest stand mean height and aboveground biomass in northeast 806 China using SAR Sentinel-1B, multispectral sentinel-2a, and DEM imagery. ISPRS J. Photogramm. Remote Sens. 2019, 151, 277–807 289.

Point 13: NFCI= xxxx, the sentence “ we employed the ninth NFCI data of Baoding city to test the classification product” is not clear for me: what is the differnce between FROM et NFCI. If NFCI correspond to a reference dataset of forest areas ==> please use it not FROM....

This point is not clear ==> please clarify because if your forest class is accurate at 74.7% the study will be problematic

Response 13: 

1. Thank you for underlying this issue. NFCI means the National Forest Resources Continuous Inventory campaign, which is conducted every five years in China. And the NFCI data of Baoding city means the 1210 sample plots with each area about 0.067ha. It is not a land use classification product, just some square plots that recorded land type attributes.
2. Thank you for the comment. We did verify the FROM-GLC 2017 product using the NFCI data, with an overall accuracy of 74.7%, while the overall accuracy of the FROM-GLC 2017 product was 72.43%. But at the same time, we additionally verified the accuracy of other land-use classification products (Global PALSAR-2/PALSAR Forest/Non-Forest Map and Global 30-m land-cover dynamic monitoring products with fine classification system from 1985 to 2020) and the Potapov-reference forest maps, and we found that the FROM-GLC 2017 product gave the best accuracy. Therefore, we finally chose the FROM- GLC 2017 product as our forest reference map.

Point 14: In Table 2, you use the ratio not the difference ==> please clarify and give the unit used for the radar backscattering coefficients;the meaning of "/" in all these expressions: for eample what is VV/VH/HH/HV

The terms of Table 3 do not correspond to what we read in Table 2: we should read in Table 2 and Table 3 the same terms.

Response 14: We are very sorry for our negligence in this issue. The “/” in VV/VH or HV/HH means ratio between them, and the “/ ” in “VV/VH/HH/HV....” means  “VVorVHor...”, we changed the way we expressed, the “VV/VH” was changed to “V/H”, the “HH/HV” was changed to “H/V”. We added units for the radar backscatter coefficient and adjusted the terms of Table 2 and Table 3 into the same one. (Table 2-3, page 7-8)

Point 15: For you the reference canopy height map is the map generated by Potapov: you confirm? if it is true please discussion the accuracy of this refernce map.

Response 15: Thank you for underlying this issue. We confirmed the reference canopy height map is generated by Potapov et al. They created a 30 m spatial resolution global forest canopy height map for the year 2019 by integrating GEDI and Landsat data. The global forest height map was compared to the GEDI validation data (RMSE = 6.6 m; MAE = 4.45 m, R² = 0.62) And we added the accuracy information results in our manuscript. (Line 435-436, page 13)

Point 16: It's very heavy to follow with so many numbers. The lines 480 to 528 must be redone by going to the essentials ==> the length must be reduced.

Response 16: Thank you for your thoughtful advice. We deleted some content and rewrote this section according to your recommendations. (Line 477-516, page 15-16)

Point 17: Please revisit your description of results according to the following recommendations:

- Boruta seems the best

- s2, S1S2p2, S2to and s1s2p2to give similar results and S1P2to gives the estimate with the lower accuracy (using Boruta)

- XgBoost gives the best accuracy

- The lower accuracy is given by SVR

Response 17: Thank you for your thoughtful advice. We revisited our description of results according to your recommendations.

1. You commented that Boruta seemed to be the best, and we gave a comparison of the estimation accuracy of SR, RFE, and Boruta three feature selection methods (Table 8). We found that the average estimation accuracy of RFE was slightly better than that of Boruta, but the running time of RFE was much longer than Boruta. Therefore, we preferred to that the accuracy of RFE and Boruta was better than SR, and could not directly concludedthat Boruta was the best in the part of this result. (Line 500-504, page 16)
2. We were very agreed with your opinion that s2, s1s2p2, s2to, and s1s2p2 gave similar results and s1p2togave the estimate with lower accuracy. We added this content to our results.(Line 477-482, page 15)
3. We agreedwith this view that “XGBoost gave the best accuracy”, but had some different views about the point that “The lower accuracy was given by SVR”. Because compared with the tree-based algorithms, the accuracy of SVR and k-NN was both low, and the accuracy of k-NN ( R² ranged from 0.08 to 0.48, RMSE ranged from 2.8 to 3.7) was even not good as SVR (with R² raged from 0.09 to 0.53, RMSE ranged from 2.6 to 3.6). (Line 506-513, page16)

Table 8. Average running time and statistical of R², RMSE, and rRMSE for different variable selection methods.

Point 18: For the Figure 3 use the same scale in Y-axis for RMSE: between 2.0 and 3.6 same remark for R²: between 0 and 0.7.

Response 18: We are very sorry for neglecting this issue. We adjusted the X-axis and Y-axis to be consistent in all the subgraphs of Figure 3. (Figure 3, page 17)

Point 19: Please discuss why the use of S1 and/or P2 in addition to S2 does not improve the results: it is not logical

Response 19: Thank you for this comment. We described this question in Discussion 4.1. We thought there were two possible causes. The first was because the C-band has limited penetration of the forest, and is vulnerable to topographic factors in mountainous areas. The second is that the used PALSAR-2 data was not the image at the time of field data collection, but the mosaic image in 2016. The inconsistency between the ground data and the image may cause the result to be not very inaccurate. (Line 604-608, page 20)

Point 20: The description ” while radar variables were as important as optical variables and terrain variables based on SR” is not true: the results obtained in the figures before show that the integration of radar data does not improve the estimation of the height (or very very little).

Response 20: Thank you for underlying this issue. We rewrote this sentence as “When using Boruta and RFE, optical variables and terrain variables were more crucial, while the importance of radar variables increased based on SR compared with Boruta and RFE.” (Line 542-543, page 17-18)

Point 21: What is the meaning about “we found the higher trees distributed almost consistently”.

Response 21: Thank you for underlying this issue. The sentence “The higher trees distributed almost consistently” means “The distribution of higher trees had certain consistency”.

Point 22: Please check if the highest differences between your estimates and the reference map are correlated with the slope because you say that the greatest differences are in mountainous areas: therefore draw the slope in X and the difference in Y.

Response 22: Thank you for your thoughtful suggestion. We checked the high difference value between my map and the reference map. And we found the areas with high differences were mountainous areas with large slopes and steep terrain. After that, we further counted the difference values above the average difference value in the distribution of different slope levels, we found that the high difference values were primarily distributed in the areas with a slope above 15°, accounting for more than 80% of the total number of high difference values. (Line 565-574, Figure 6-7, page 18-20)

Point 23: Please explain why your map was build using data acquired in 2016 not in 2019 for example

Response 23: Thank you for underlying this issue. We are very sorry to tell you that the only field plot data we currently obtained is from 2016, and we are also looking forward to obtaining more updated field data for our future study.

Point 24: Please use the same forest mask for your map and for the Potapova map. The Figure 5 is not visible ==> please improve the quality and add the map with the difference, and add the map with the slope.

Response 24: Thank you for underlying this issue. We changed the forest mask of our forest height map and Potapov’s map to the same. We also improved the image quality of Figure 5 and added the two maps with the difference and slope, respectively (Figure 5-6, page 19).

Point 25: The description “demonstrated that the performance of the combination of optical, radar, and terrain variables was better than that of a single sensor, which indicated the necessity of fusing multi-source remote sensing data to estimate forest height.” is not correct.

Response 25: Thank you for underlying this issue. We added “slightly” between “was” and “better” and removed the sentence “which indicated the necessity of fusing multi-source remote sensing data to estimate forest height” (Line 597, page 20)

Point 26: For this sentence ”so these three factors could not be disregarded”, please add the three factors in brackets

Response 26: Thank you for underlying this issue. We added it based on your suggestion. The modified sentence is “so these three factors including data source, feature selection, and regression algorithm could not be disregarded” (Line 676-677, page 22)

Point 27: Result obtained with only optical data were slightly lower ==> please add this conclusion

Response 27: Thank you for underlying this issue. We added this conclusion to our manuscript. The added sentence is “The combined data of multiple sensors could improve the estimation accuracy to a certain extent compared with a single optical sensor whose accuracy was relatively low.” (Line 752-754, page 24)

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The modified version is more circumspect, but I would contend that the last sentence of the abstract is still not justified and should be deleted.  To justify that statement would require consideration of confounding between biomass and height.  Consideration would need to be given to what I will call sensing of "forestness".  Nothing has been given about correlation in the field data between height and biomass.  Height is a factor in biomass to varying degrees, and many of the sensors involved obviously respond to density of foliar material.  Stand structures in the study area would need to be addressed to help differentiate between foliar density, vertical structure, and horizontal structure.  Absent this, statements regarding global extensibility seem unwarranted.

Author Response

Point: The modified version is more circumspect, but I would contend that the last sentence of the abstract is still not justified and should be deleted.  To justify that statement would require consideration of confounding between biomass and height.  Consideration would need to be given to what I will call sensing of "forestness".  Nothing has been given about correlation in the field data between height and biomass.  Height is a factor in biomass to varying degrees, and many of the sensors involved obviously respond to density of foliar material.  Stand structures in the study area would need to be addressed to help differentiate between foliar density, vertical structure, and horizontal structure. Absent this, statements regarding global extensibility seem unwarranted.

Response: Thank you very much for your recognition and thoughtful advice. We deleted the last sentence of the abstract according to your advice. (Line 39-40, page 1)

Author Response File: Author Response.docx

Reviewer 2 Report

The paper can be accepted. 

Author Response

Point: The paper can be accepted.

Response: Thank you very much. We are very grateful for your effort in reviewing our paper and your positive feedback. 

Author Response File: Author Response.docx

Reviewer 3 Report

Some comments have not been taken into account.

 Lines 63-65: I disagree with the edit. Please modify the text according to my comment in round 1:

“Waveform lidar data can penetrate the forest canopy to reach the ground. This is also the case for SAR data but only at very high radar wavelength (beyond the L band). The C-band penetrates the forest canopy only slightly. L-band penetrates the forest canopy to reach the ground only in certain cases... This whole part is wrong and must be completely redone”

Lines 66-72: “please discuss the limits”

 Section 2.1: please give additional useful information for forest: the range of canopy hight and biomass

 one digit after the comma for RMSE and two for R

 Result obtained with only optical data were slightly lower ==> please add this conclusion

Author Response

Point: Some comments have not been taken into account.

Response: Thank you for your time making your constructive remarks and useful suggestions on our manuscript. We are very sorry for our negligence in these issues. We revised the manuscript according to your each suggested revision and comment. Based on the instructions provided, we uploaded a copy of the original manuscript with all the changes highlighted using the “Track Changes” function in Microsoft Office Word. Listed below are the point-to-point responses to your comments.

Point 1: Lines 63-65: I disagree with the edit. Please modify the text according to my comment in round 1:

Response 1: Thank you for underlying this issue. We modified the text in our manuscript. “Terrestrial laser scanning (TLS) and airborne laser scanning (ALS) are typically limited by their high application costs, and it is difficult to generate wall-to-wall forest height maps in large areas due to the sparse measurements in the space of satellite LiDAR.”(Line 62-66, page 2)

Point 2: “Waveform lidar data can penetrate the forest canopy to reach the ground. This is also the case for SAR data but only at very high radar wavelength (beyond the L band). The C-band penetrates the forest canopy only slightly. L-band penetrates the forest canopy to reach the ground only in certain cases... This whole part is wrong and must be completely redone”

Response 2:Thank you for underlying this issue. We deleted the description of this whole part. (Line 72-73, page 2)

Point 3: Lines 66-72: “please discuss the limits”

Response 3: Thank you for underlying this issue. We replenished this content into our manuscript “Compared to lidar data, optical data is more susceptible to the influence of weather conditions and has issues like limited sensitivity and low saturation in dense vegetation areas, SAR data is susceptible to terrain and speckle noise, and there is a problem of backscatter signal saturation in high vegetation coverage areas as well as optical data”. (Line 68-72, page 2)

Point 4: Section 2.1: please give additional useful information for forest: the range of canopy hight and biomass

Response 4: Thank you for underlying this issue. We tried our best to find the information on forest canopy height and biomass in Baoding city, but there is no corresponding information. And according to the National Forest Resources Continuous Inventory data which contained 128 forest sample plots in Baoding city, the forest height ranged from 3.0m to 24.5m.

Point 5: one digit after the comma for RMSE and two for R

Response 5: Thank you for underlying this issue. We changed the digit of R² and RMSE throughout the manuscript.

Point 6: Result obtained with only optical data were slightly lower ==> please add this conclusion

Response 6: Thank you for underlying this issue. We added this sentence “The accuracy with optical data alone was slightly lower than the combined data of multiple sensors” to our manuscript according to your advice. (Line 759-760, page 24)

Author Response File: Author Response.docx