Integrating Remote Sensing Data and CNN-LSTM-Attention Techniques for Improved Forest Stock Volume Estimation: A Comprehensive Analysis of Baishanzu Forest Park, China

Round 1

Reviewer 1 Report

The article combines remotely sensed imagery and modeling to estimate forest storage, but there are several suggestions as below.

1. Most of the figure requirements are not up to standard. Figure 1 does not have a legend. Suggest Figure 2 adding different colors to each sub-figure in order to distinguish and show the details of their respective contents. Figure 3 is totally unacceptable. The size of Figure 8 is too small and unclear.

2. In section 3.4 the reason why 5 types of parameters were chosen is missing.

3.The article should have focused on analyzing the spatial and temporal variation and distribution of forest stock in section 4.3 and lacked a comparative discussion with related studies.

Extensive editing of English language is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript titled "Estimation of Forest Stock Volume based on CNN-LSTM-Attention from Landsat 8 and SRTM: A Case Study of Baishanzu Forest Park, China" introduces a novel method for estimating Forest Stock Volume (FSV). The novelty lies in the methodology employed for FSV estimation, wherein the authors utilize various datasets and compare different methods, assessing their goodness of fit. The manuscript incorporates significant changes to the methods, including modifications to the representation of graphs and reporting of outcomes. Please carefully review all the comments provided to enhance the manuscript before submitting the revised version.

Comments

1.     Authors only suggest about the passive data, not includes about the other format of data in line 43-48.

2.     Line no 106, 109 the text R2 should be replaced with ‘Coefficient of Determination (R²)’ and other places in the MS.

3.     Author should measure the Goodness of fit not only with Coefficient of Determination and MAE, they also used the Chi square, RMSE, AUC and other methods. The coefficient of determination doesnot provide the complete over view of the correlation.

4.     The summary explained in the introduction section (lno 111-118), should be changed with the inference of literatures or hypothesis, the research on FSV used the mentioned datasets. Any summary and conclusions should be moved to the conclusion or discussion section of the MS.

5.     Write about the modelling factors (lno:113-118), what all the factors is it some kind of black box or is it a language error?

6.     Table 1 the Landsat8 band B10-TIR, spatial resolution should be check, is it of 30/100m. if it is for 100m, then explain about how the thermal band of the landsat8 image is used with other 30m bands for PCA and texture based product derivation, Pearson correlation in between the variables and how it is correlated for the goodness of fit?

7.     In PCA factor (3.4.2), After three components have great noise??, who is it so, and kindly take look in the process of PCA, how the eigen value was calculated between the components. Author should correctly describe about why they are using PCA in place of lansat8 image.

Author suggest the MS to read more about the PCA

A machine learning-based classification of LANDSAT images to map land use and land cover of India. Remote Sensing Applications:Society and Environment.

8.     PCA graph (Figure 3:) I hope it should be corrected as graph between ‘components’ (X axis) and ‘eigen value’ (Y axis). It caption should also be corrected accordingly.

9.     Author should correct the flow diagram (Figure 6), direction of flow at every point and explain it with basic English what is input, what is intermediate(hidden) and what is predicted outcome. Clearly show the flow direction at every point function.

10.   Author should try to use the PCA in Data 5/6 data set and tried to predict the model outcome. It reduces the noise and data dimensionality between all kind of texture, vegetation index products, srtm data.

11.  In Dataset 3/4/5/6, why image band and its PCA together used for the analysis. PCA or image band should be used analysis. It also affected the correction between the products and goodness of fitness model.

12.   Figure 10 representing absolute error between real and predicted values from different models, kindly check the point value representation for all the blue color points (difference value between 50 and 100 unit value)

13.  The result section should add the discussion about the data, methods selection and Spatial map which are representing the predicted FSV outcome from the CNN-LSTM model.

General comments

1.     Check English in line 33 to 36.

2.     Heading 3.6 what is ‘FSV Rrediction Model’, kindly correct it

3.     Figure 8 the label fonts should be readable. Other wise use it notation in the graph label. It caption should also be corrected.

4.     Font size of all the diagram label should be corrected.

 Suggestion to change the topic of the MS-

“Integrating Remote Sensing Data and CNN-LSTM-Attention Techniques for Improved Forest Stock Volume Estimation: A Comprehensive Analysis of Baishanzu Forest Park, China”

1.     Check English in line 33 to 36.

2.     Heading 3.6 what is ‘FSV Rrediction Model’, kindly correct it

3.     Figure 8 the label fonts should be readable. Other wise use it notation in the graph label. It caption should also be corrected.

4.     Font size of all the diagram label should be corrected.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript explores approaches to forest stock volume estimation based on the combined use of Landsat 8 and SRTM data using quantitative spectral reflectance analysis techniques. The topic of the article is quite relevant, given that forest stock volume is one of the key parameters of forests. When performing the study, the authors use modern methods. However, a number of corrections need to be made to the manuscript.

1)      The annotation must provide a definition of all abbreviations used.

2) Lines 23-24. FSV values must be rounded to one or two digits after the integer part of the number.

3) Section 3.1. (Forest Survey Data): information about the species composition of forests must be provided. It is necessary to indicate how many objects were used for the training and test sets.

4) All figures in the manuscript must be presented in higher quality.

5) Table 1, bottom line. It is necessary to correct the spatial resolution of the TIR band (B10) from 30 to 100 m. The title of Table 1 needs to be adjusted, since not all Landsat 8 bands are represented in the table.

6) Section 3.2 (Landsat 8 Data): the acquisition dates of Landsat images must be specified.

7) Section 3.4.1. No reference to Figure 2.

8) Figure 2. It is necessary to show a scale bar in the figure.

9) Line 188. In Figure 3, you need to enter the designation of units of measurement along the vertical axis and remove unnecessary text in the title.

10) The introductory part to section 4 (lines 330-335) is unnecessary. It is recommended to remove it.

11) Line 345. Figure 8 needs to be edited so that it does not contain small, hard-to-read symbols along the horizontal and vertical axis.

12) Line 380 and Line 414. It is necessary to provide an explanation of the column names in tables 4-5.

13) Line 426. In Figure 11 it is necessary to provide a scale bar.

14) The “Conclusion” section should be shortened and edited so that it briefly summarizes only the results obtained and their small analysis.

15) In the text of the manuscript it is necessary to correct “LandSat8” to “Landsat 8”.

The manuscript should be checked for frequent repetitions of the same words, for example: Line 182. It is also necessary to check the manuscript for the presence of erroneously combined words and symbols, for example: Line 124 (2000m), Line 126 (2341.8mm).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Thank you for the opportunity to review the manuscript tiled “ “. The manuscript predicts forest stock volume using Landsat and SRTM data derived variables and CNN-LSTM attention algorithm. The study derives a set of variables based on band reflectance, indices, texture variables from Landsat data and three variables from SRTM data. Then the variables group in to 6 categories and apply them through one dimensional CNN followed by LSTM to predict the forest stock volume derived using field data. The performance of the new model is compared against commonly used three other models. The final results shows that incorporation of all predictor variables provide the highest accuracy among the data grouping. Compare to the other methods, this proposed method shows higher performance. The study is very interesting and addresses a critical topic; the use of optical data to estimate FSV in larger scales. However, there need some modifications before the publication. Please refer below detailed comments and suggestion that are critical to better understand the quality and the improvement provided to the method described in this manuscript.

The manuscript has clear sentences and a great introduction. However, the data, figures, and the discussion must be improved.  In the methods, requires following information regarding the field data and Landsat data processing. How many samples were collected (please make sure to include the number provided here in the manuscript is the number of data points or number of plots.). If they are plots, what size the plots are, where those sample/plot locations in the study site image in figure 1, what methods were used to collect the data, what year range was the data collection happened or were those from a single year. If they are from multiple years, what years and how the Landsat data were processed to accommodate the year of data collection.  There is no explanation on how the time series of Landsat data were used or only used one time image. What methods were used to accommodate date/year differences in field data and Landsat data collection. Also provide a summary of data collected in a table, especially the distribution of species, vegetation density, height, DBH variation (min/max/mean) so the reader can better understand the data. If the plot sizes are different from Landsat pixel sizes, how you resolve that issue when derive the variables?

Please provide how the principal component analysis was done? What were the inputs?

All figures need descriptive captions. The axes in all figures are not readable. Please make them little bigger.

Needs some methods description of other methods used here to compare the data and how they were implemented. What software used and what variables used. Did you test any variable selections and used all the variables those were in the best method in the main method describe in this manuscript?

Some metrics of uncertainty have been measured but no uncertainty evaluation within the study area t see how well the proposed method perform overall and what factors may influence that variability.

There is no discussion on method performance and potential underline reasons on what or why not these methods, variable selections work. Also, there are some points for future studies based on what factors might need to further investigate, however why this method would need to apply and how this method would work in other areas (or whether it would work only in this area) has not been discussed. Also, the value, implications, and novelty of the study and how this study superior or good compared to other methods have not been discussed (performance or computational ease). There need to have a discussion on above areas.

L 196: what is the moving window size for texture variable calculations.

L202: topographic factors three or two. Three mentions here. And how the aspect was calculated.

L307-313: better and a workflow diagram for this section.

L471: be consistent on writing the name most established datasets like Landsat.

Figures:

Figure 1, need to include whare this location compared to the mainland of China and where are the locations of field data. Why there is one tiny island in the south area. Why that was considered and not the land in between or was this little tiny island excluded?

Figure 2: Prefer color figures. Also need to add legend for all figures and also what each index abbreviation stands for.

Figure 3: figure title on top has cut off.

Figure 4: What is DEM? Legend is not readable.

Figure5: Add more description on layer structures were implemented here

Figure 6 &7:  What those letters represent. They have been explained in the text but needs to add little description in the figure cation too.

Figure 8: Axis labels cannot read. May be better to add only variables that have higher correlation with FSV.

Figure 9: axis labels are not readable.

Figure 10. Axis labels and legend are not readable.

Figure 11. Needs to dd uncertainty map as well to see how well the prediction varies across the study region.

Tables:

Table 2: what each of these index’s abbreviations stands for? Any references for these indices?

Table 3: what is filtered bands?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The manuscript imporved after the revision, so I think it should be considered for publication.

The English writing should be improved.

Reviewer 2 Report

The updated manuscript, entitled "Integrating Remote Sensing Data and CNN-LSTM-Attention Techniques for Improved Forest Stock Volume Estimation: A Comprehensive Analysis of Baishanzu Forest Park, China," is set for publication in the journal. All comments raised, including those regarding the manuscript's title, have been addressed in the revised version.

Reviewer 3 Report

The authors took into account the comments. The manuscript may be recommended for publication.

It is recommended to check the manuscript for minor inaccuracies.