Evaluating the Sensitivity of Polarimetric Features Related to Rotation Domain and Mapping Chinese Fir AGB Using Quad-Polarimetric SAR Images

Round 1

Reviewer 1 Report

This research evaluates the sensitivity of the rotated polarimetric features with forest parameters and AGB, compares the features screening capability of SI and Pearson correlation coefficient, and map the forest AGB based on MLR model using ALOS PALSAR-2 images. It is an interesting topic to explore the potentials of the rotated polarimetric features for mapping forest AGB. The manuscript is well written.

1.         Title: the tree species is Chinese fir, so it is appropriate to add the research object, Chinese fir, to the title.

2.         Line 46: ‘in planted forest t’?

3.         Line119: How to measure DBH and tree height？adding the instruments used.

4.         Formulas (1) and (2): It’s very likely that the relationship between biomass and d,h for every component is different, so if using the same models to estimate AGB of the different components, it will produce errors.

5.         Line 196 & 216: The same headings

6.         Line 284-287: grammar mistake

7.         3.5 Feature selection method with SI: Does the order in which variables are added affect the results? Is there a difference between not grouping and grouping when adding features?

8.         4.3 The results of feature selection based on SI: in this part, all alternative features were grouped in to three categories, and figure 9 just illustrates the results of the three sets. How about the results when all the features are screened together?

9.         Discussion: It’s basically a repeated description of the methods and results. It’s better to add some in-depth interpretations of the physical meanings of the features used, comparisons with similar studies, and the limitations of this study.

Author Response

Response to Reviewer 1 Comments

 

Thank you for your insightful comments and suggestions. Please find the answers to each of your questions below.

 

Point 1: Title: the tree species is Chinese fir, so it is appropriate to add the research object, Chinese fir, to the title.

 

Response 1: The title has been changed to ” Evaluating the sensitivity of polarimetric features related with rotation domain and mapping Chinese fir AGB using quad-polarimetric SAR images”.

 

Point 2: Line 46: ‘in planted forest t’?

 

Response 2: The excess "t" in line 46 has been removed.

 

Point 3: Line119: How to measure DBH and tree height？adding the instruments used.

 

Response 3: DBH and tree height were measured using diameter tape and laser altimeter. It has been added in line 119.

 

Point 4: Formulas (1) and (2): It’s very likely that the relationship between biomass and d,h for every component is different, so if using the same models to estimate AGB of the different components, it will produce errors.

 

Response 4: This formula is derived from the empirical formula of Chinese fir AGB technology, which is the same as the formula in reference 44. The various components of the AGB in the formula are determined by the coefficients "a" and "b".

 

Point 5: Line 196 & 216: The same headings.

 

Response 5: Title in line 216 has been changed to "3.3.2. Polarimetric coherence features in rotation domain".

 

Point 6: Line 284-287: grammar mistake

 

Response 6: Syntax errors have been fixed for line 284 through 287.

 

Point 7: 3.5 Feature selection method with SI: Does the order in which variables are added affect the results? Is there a difference between not grouping and grouping when adding features?

 

Response 7: The order in which variables are added affects the result. For polarization feature, the higher SI, the higher accuracy of the AGB results. When a higher SI feature is combined with a lower SI feature, even if the accuracy is slightly improved, the accuracy will be dragged down by the previous lower SI feature after adding a higher SI feature, and the accuracy will increase very little. Combining directly in the order of SI value can quickly improve the accuracy to a higher value, and then adding low SI features can quickly screen out the features that do not improve the accuracy.

 

The purpose of grouping features in this study is to compare the ability of different types of polarization features in AGB estimation. When SI values are determined and combined in the order of SI values, the results should be similar with and without grouping. However, the contribution of different types of polarization characteristics to the AGB estimation model cannot be determined without grouping.

 

Point 8: 4.3 The results of feature selection based on SI: in this part, all alternative features were grouped in to three categories, and figure 9 just illustrates the results of the three sets. How about the results when all the features are screened together?

 

Response 8: Part 4.3 is to demonstrate that the SIS method is superior to the PSS method. When all features are filtered together, the results are consistent with those of the three feature sets shown in Figure 9. Because in this experiment, the sensitivity index is better than the commonly used Pearson correlation coefficient.

 

Point 9: Discussion: It’s basically a repeated description of the methods and results. It’s better to add some in-depth interpretations of the physical meanings of the features used, comparisons with similar studies, and the limitations of this study.

 

Response 9: Changes have been made to the discussion section.

 

Author Response File: Author Response.docx

Reviewer 2 Report

This paper presents a new method to map forest AGB with several rotated polarimetric features, which is proved to be effective. And sensitivity of rotated polarimetric features with forest parameters were evaluated by Pearson correlation coefficient, SI and saturation levels. The result shows that the accuracy is better than previous researches. The idea of the article is good, but there are still the following problems:

 

Major points:

1) The parameters selected in this paper are not tested for independence.

2) The results need to be verified on a large scale.

3) This article mainly applies the existing methods, but is not innovative enough.

4) The title of 3.3.2 is the same one as 3.3.1. Please check it.

5) No significance test was performed when using Pearson correlations.

6) The title of this paper is about mapping forest AGB and point out the advantages of mapping planted forest, lacking of evidence for the superiority of other forests.

Figures and Tables: 1) The legend in figure 1 is not clear. Band_1 to Band_3 of which image is unknown. Besides, there is an obvious error in the scale. Please check it. 2) The titles and notes of some figures and tables make the readers confused. For example, it’s better to tell the readers the meaning of number and the number of min\max\mean. 3) The typesetting of Fig 4 is not so comfortable. Please use figure with high definition and add the process of Step 3. 4) It is not persuasive to use Fig 13 to express relationship between SI and features.

Methods and Results: 1) Line 338 and Fig 7 deem that BC and C4 have a good response relationship with AGB. However, there is no function or quantitative method to describe it. How to explain this view? 2) Line 368, what does “appropriate feature selection method” mean? Comparing Pearson correlations + MLR with SI + MLR using features sorting by SI is not convincing. 3) The number of features in each feature set is not mentioned in the article. Choose the same number or the best number? 4) Fig7 and Fig10, how to determine which values are saturation values?

References: Line 203, Incorrect citation format of references.

 

 

Minor points:

Line 46: Redundant “t”

Line 53: various sizes

Line 82: “is become” may be wrong

Line 88: full name of GSV

Line 162: may be “different sizes”

Fig 5 and Fig 6: please change the sizes and position of words

Author Response

Response to Reviewer 2 Comments

 

Thank you for your insightful comments and suggestions. Please find the answers to each of your questions below.

 

Major points

 

Point 1: The parameters selected in this paper are not tested for independence.

 

Response 1: The three polarization features extracted in this experiment are extracted from the physical significance of polarimetric SAR data. For example, the backscattering coefficient represented by BC feature reflects the intensity information after radar reflection on the target. The polarization decomposition feature represents the scattering information distinguished by the radar for different scattering types of targets. The correlation between these features and AGB was calculated after extraction. Only the features with high correlation were retained for estimation, while the features with low correlation were abandoned. The purpose of this paper is to explore the response relationship between these polarization characteristics and AGB, and find out the polarization characteristics suitable for estimating AGB.

 

Point 2: The results need to be verified on a large scale.

 

Response 2: In this experiment, there are 50 sample sites for field investigation, and the estimation accuracy of AGB is good according to the current results. Therefore, we will further expand the range and number of sample sites to verify the AGB estimation effect of this method in a large area.

 

Point 3: This article mainly applies the existing methods, but is not innovative enough.

 

Response 3: These methods are relatively novel at present. We are very interested in these methods. For example, the polarization rotation feature is only used in the classification problem in the previous research, and we try to apply it to the regression problem. Previous studies on the sensitivity index were used to study the sensitivity of SAR data of different bands to AGB, but we tried to use the sensitivity of different types of polarization characteristics and AGB. These methods provide us with more ideas, and we also hope to make further innovations on the basis of these methods.

 

Point 4: The title of 3.3.2 is the same one as 3.3.1. Please check it..

 

Response 4: Title in line 216 has been changed to "3.3.2. Polarimetric coherence features in rotation domain".

 

Point 5: No significance test was performed when using Pearson correlations.

 

Response 5: In this experiment, the significance test of Pearson correlation coefficient of features was carried out. The results are not marked because they are presented in the form of graphs. Among them, the features selected by the AGB model constructed by PSS method were all significantly correlated at the level of 0.05 or 0.01. In conclusion 4.1, supplementary explanations were also made for significance test.

 

Point 6: The title of this paper is about mapping forest AGB and point out the advantages of mapping planted forest, lacking of evidence for the superiority of other forests.

 

Response 6: The Huangfengqiao Forest Farm in the research area of this experiment is mainly a forest farm of Chinese fir plantation, so the research objects of this experiment are all Chinese fir plantation, and other forest types and tree species will be further verified by the method in this paper. At the same time, the title of the paper is changed to "Evaluating the sensitivity of polarimetric features related with rotation domain and mapping Chinese fir AGB  using quad-polarimetric SAR images".

 

 

Figures and Tables

 

Point 1: The legend in figure 1 is not clear. Band_1 to Band_3 of which image is unknown. Besides, there is an obvious error in the scale. Please check it.

 

Response 1: Figure 1 has been reuploaded in higher definition. RGB image is the optical image of Huangfengqiao Forest Farm during the corresponding period. At the same time, a supplementary explanation is also made in the article figure 1. The scale error has also been corrected.

 

Point 2: The titles and notes of some figures and tables make the readers confused. For example, it’s better to tell the readers the meaning of number and the number of min\max\mean.

 

Response 2: The titles and notes of tables and figures in the article have been modified and supplemented.

 

Point 3: The typesetting of Fig 4 is not so comfortable. Please use figure with high definition and add the process of Step 3.

 

Response 3: Figure 4 has been added and modified and reuploaded.

 

Point 4: It is not persuasive to use Fig 13 to express relationship between SI and features.

 

Response 4: What Figure 13 intends to express is that the SI level can be improved from two sources: continuous growth within the range of AGB, i.e. low saturation level, or low fitting error, i.e. RMSE. It can be clearly seen that these two features have similar SI, but feature a is due to high saturation level, while feature b is due to low RMSE. Thus, when the characteristics of high saturation and low RMSE are satisfied at the same time, the SI will be higher.

 

Methods and Results

 

 

Point 1: Line 338 and Fig 7 deem that BC and C4 have a good response relationship with AGB. However, there is no function or quantitative method to describe it. How to explain this view?

 

Response 1: The problem you mentioned is exactly the saturation problem mentioned in the article. How to quantify and express the saturation problem is also a hot issue in AGB estimation. So far, there are few methods or models that can accurately describe or quantify the saturation problem. At present, it is generally believed that the more understandable and intuitive method is to observe the scatter plot between features and AGB. As shown in Figure 7, the value of the three types of features increases with the increase of AGB value in the range of 0-200 m³/ha, that is, there is a good correlation mentioned in the paper, while the characteristic value of BC feature and C4 feature tends to be stable and does not increase with the increase of AGB when AGB is greater than 200 m³/ha. So this phenomenon is used to determine the saturation problem that occurs.

 

Point 2: Line 368, what does “appropriate feature selection method” mean? Comparing Pearson correlations + MLR with SI + MLR using features sorting by SI is not convincing.

 

Response 2: Line 368 expresses the meaning that optimal selection needs to be obtained by using feature selection method. "appropriate " has been removed to avoid ambiguity. We believe that Pearson correlation coefficient only considers the statistical response relationship between the feature and AGB, ignoring the specific situation in AGB estimation. SI considers the two key indicators in AGB estimation, saturation and error, and is more suitable as the feature screening criteria in AGB estimation. The MLR model is used in consideration of the fact that it is generally selected as the benchmark model for regression problems. We believe that it is reliable to compare feature screening methods with MLR model.

 

Point 3: The number of features in each feature set is not mentioned in the article. Choose the same number or the best number?

 

Response 3: In order to fully compare the two feature screening methods, we used the best number to construct the AGB estimation feature set. Table 5 in the paper also marks the number of features in the optimal feature set selected under each type and feature selection method.

 

Point 4: Fig7 and Fig10, how to determine which values are saturation values?

 

Response 3: Figure 7 and Figure 10 cannot accurately determine the saturation value. Only by observing the general range of saturation phenomenon, how to accurately calculate the specific saturation value is still a major problem that needs to be broken through in AGB estimation, and we are also exploring and making efforts on this problem.

 

 

References

 

Point 1: Line 203, Incorrect citation format of references.

 

Response 1: The reference on line 203 has been incorrectly formatted.

 

 

Minor points

 

Point 1: Line 46: Redundant “t”.

 

Response 1: The excess "t" in line 46 has been removed.

 

Point 2: Line 53: various sizes.

 

Response 2: The syntax errors in line 53 have been fixed.

 

Point 3: Line 82: “is become” may be wrong.

 

Response 3: The syntax error on line 82 has been fixed.

 

Point 4: Line 88: full name of GSV.

 

Response 4: The full name of GSV (growing stem volume) has been added on line 88.

 

Point 5: Line 162: may be “different sizes”.

 

Response 5: The syntax error on line 162 has been fixed.

 

Point 6: Fig 5 and Fig 6: please change the sizes and position of words.

 

Response 6: Fig 5 and Fig 6 have been modified.

 

 

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript presents an evaluation of the sensitivity of SAR features related with rotation domain and the use of the optimal features to mapping forest AGB. I found the paper to be well written and very much of it well described. The design of the experiments makes the datasets seem quite useful for the purpose of the research. Also, the current study is on a topic of relevance and general interest to the readers of the journal. Therefore, I would recommend the  manuscript to be considered for publication. 

Author Response

Response to Reviewer 3 Comments

 

Thank you for your insightful comments and suggestions. Please find the answers to each of your questions below.

 

Point 1: The manuscript presents an evaluation of the sensitivity of SAR features related with rotation domain and the use of the optimal features to mapping forest AGB. I found the paper to be well written and very much of it well described. The design of the experiments makes the datasets seem quite useful for the purpose of the research. Also, the current study is on a topic of relevance and general interest to the readers of the journal. Therefore, I would recommend the  manuscript to be considered for publication.

 

Response 1: Thank you very much for your valuable advice during your busy schedule. The purpose of this study was to investigate the potential of polarization rotation features extracted from L-band polarimetric SAR data for AGB estimation. The present results show great potential, so we will further combine these features with more data to further improve the accuracy of AGB estimation.  

 

 

Author Response File: Author Response.docx

Reviewer 4 Report

Essentially, the study is using L-band SAR data with several modifications on the backscatter coefficients to predict aboveground forest biomass which is interesting. The study introduces a novel idea to use SAR data; however, I would love to see them incorporating optical data as well (not for this research but could write a paragraph on its potential). There are several related researches on the use of SAR and optical data for similar studies. A recent paper on estimating leaf area index and basal area using SAR and optical data as follow could be compared with this research since the mentioned article claims SAR data (C-band) to be not very effective for basal area estimation:

Furthermore, I would suggest the authors to include some site variables (refer above article) in their study and see if their inclusion makes any modeling improvement (may be include a paragraph on it for now).

Some specific comments are: 

line 120-121: please clarify how you obtained mean range. I suspect authors are just trying to mention diameter and tree height range.

Page 125: I am having a hard time to understand Wi. What is it's significance? 

Page 144: How about using a smaller filter window? 

Page 263: Please mention the previous studies.

Page 271: MLR is not defined anywhere in the text. Please consider defining it before using.

Page 337-339: Would incorporating other datasets (optical) solve this problem? 

Page 388: I believe authors need to simplify and explain the variable selection procedure they used.

I also would love to know which software or package authors used to obtain SAR polarimetric decomposition features. 

Further grammatical and spelling errors are highlighted in the pdf file attached. 

 

 

 

 

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 4 Comments

 

Thank you for your insightful comments and suggestions. Please find the answers to each of your questions below.

 

Point 1: Essentially, the study is using L-band SAR data with several modifications on the backscatter coefficients to predict aboveground forest biomass which is interesting. The study introduces a novel idea to use SAR data; however, I would love to see them incorporating optical data as well (not for this research but could write a paragraph on its potential). There are several related researches on the use of SAR and optical data for similar studies. A recent paper on estimating leaf area index and basal area using SAR and optical data as follow could be compared with this research since the mentioned article claims SAR data (C-band) to be not very effective for basal area estimation:

 

Furthermore, I would suggest the authors to include some site variables (refer above article) in their study and see if their inclusion makes any modeling improvement (may be include a paragraph on it for now).

 

Response 1: Thank you very much for your valuable advice during your busy schedule. The purpose of this study was to investigate the potential of polarization rotation features extracted from L-band polarimetric SAR data for AGB estimation. The present results show great potential, so we will further combine these features with optical data to further improve the accuracy of AGB estimation. We will also consider adding the site variables you mentioned to further experiments, hoping to bring better results.

 

Point 2: line 120-121: please clarify how you obtained mean range. I suspect authors are just trying to mention diameter and tree height range.

 

Response 2: The diameter at breast height (d) and tree height (h) mentioned here refer to the diameter at breast height and tree height of each of the 50 plots in the field survey. This range refers to the range of the diameter at breast height and tree height from small to large in the 50 plots.

 

Point 3: Page 125: I am having a hard time to understand Wi. What is it's significance?

 

Response 3:  refers to the biomass of each part of Chinese fir. The biomass of Chinese fir is calculated in four parts, including branches, trunks, leaves and bark. Converting to Wi is Wb, Ws, Wl, and Wp.

 

Point 4: Page 144: How about using a smaller filter window?

 

Response 4: For two SAR images, we experimented with both larger and smaller windows, and the results showed that a 7x7 window was the most suitable. In addition, according to the studies of other scholars, such as Reference 15, the 7X7 window of the experiment in forest area is more appropriate.

 

Point 5: Page 263: Please mention the previous studies.

 

Response 5: Reference 42 has been cited at line 263 to indicate the previous studies.

 

Point 6: Page 271: MLR is not defined anywhere in the text. Please consider defining it before using.

 

Response 6: The full term of MLR(multiple linear regression) has been added in line 271.

 

Point 7: Page 337-339: Would incorporating other datasets (optical) solve this problem?

 

Response 7: In general, theoretically speaking, optical data as passive remote sensing is affected by various complex factors, which will lead to low saturation between the optical data and AGB. As active remote sensing data, SAR data is less affected by environment, so it is generally believed that the saturation between SAR data and AGB is relatively high. Many studies have shown that the combination of optical data and SAR data can improve the saturation, so we will further try to combine optical data to further improve the saturation in AGB estimation.

 

Point 8: Page 388: I believe authors need to simplify and explain the variable selection procedure they used.

 

Response 8: The variable selection process has been further explained.

 

Point 9: I also would love to know which software or package authors used to obtain SAR polarimetric decomposition features.

 

Response 9: In this experiment, “PolSARpro_v6.0.3_Biomass_Edition” software was used to extract the polarization decomposition features. Matlab was used for partial processing and extraction.

 

Point 10: Further grammatical and spelling errors are highlighted in the pdf file attached.

 

Response 10: The grammar and spelling errors you noted in the pdf have been corrected.

 

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

 There are still the following suggestions：

1）  The significance test results are suggested to be marked in the diagram.

2）  Fig7： It may be possible to use regression analysis to represent the results like Fig10，or using different colors to represent saturation values.

Author Response

Response to Reviewer 2 Comments

 

Thank you for your insightful comments and suggestions. Please find the answers to each of your questions below.

 

Point 1: The significance test results are suggested to be marked in the diagram.

 

Response 1: The results of the significance tests have been added in Figure 5 and Figure 6 with additional annotations.

 

Point 2: Fig7： It may be possible to use regression analysis to represent the results like Fig10，or using different colors to represent saturation values.

 

Response 2: Figure 7 shows a scatter plot between typical features and AGB in order to compare the three different types of polarization features. The most popularly presented and intuitive method for the observation of saturation phenomena is currently the observation of a scatter plot between a single feature and the AGB. As mentioned in Ref. 3, Ref. 15 and Ref. 48. Of course there are also methods to quantify saturation by regression results, but they are affected by the combination of features and regression models.  Also, the results of the SI in Figure 8 and the R² of the semi-exponential model can be seen as regressions for comparison. We just want to compare the saturation of the three polarization features here using the method that is most acceptable in the field to evaluate the saturation phenomenon. The specific quantification of saturation to obtain an accurate saturation method we are also working for.