Mapping Forest Aboveground Biomass Using Multisource Remotely Sensed Data

Round 1

Reviewer 1 Report

I am very glad to see that authors have optimized their manuscript according to the advice from reviewers. There are some details that still need to be optimized before publication:

• The format of paragraphs on Page 8.
• Please unify the format and size of all equations shown in this paper.
• Sentinel-2 (not Sentineal-2) in Figure 2.
• The format of Table 2.
• Please unify the format of references. It may be necessary to mark DOI. Some examples are shown in the Reference List and Citations Style Guide.

Author Response

Thanks for your review comments. Addressing these comments improved the quality of the manuscript. Thanks for identifying these minor issues. We addressed them all. We reformatted the paragraph on page 8; unified the size of equations, corrected the spelling in Figure 2; reformatted Table 2. As for the reference styles, some references only provided DOI, whereas other older references do not have DOI. Thus, we cannot unify them. But each reference can be identified with the information provided.

Reviewer 2 Report

The manuscript presents a conceptual model to map forest aboveground biomass using multiple remote sensing data. The results revealed, as expected, the LiDAR height metrics contributed the most in forest biomass estimation. The conceptual model is robust through an assessment of the effects of sample size. Although I cannot find anything more interesting in this manuscript. I believed the content of the article is informative and corrected. It still makes sense to lay the groundwork on a broad scaleThe manuscript presents a conceptual model to map forest aboveground biomass using multiple remote sensing data. The results revealed, as expected, the LiDAR height metrics contributed the most in forest biomass estimation. The conceptual model is robust through an assessment of the effects of sample size. Although I cannot find anything more interesting in this manuscript. I believed the content of the article is informative and corrected. It still makes sense to lay the groundwork on a broad scale.

Specific comments:

1. Comparisons with previous studies should be included in the discussion part rather in the results, e.g., Lines 415, 424.
2. The relative RSME provided for comparison purposes.
3. Line 554, RMSE of the model was 18.5 Mg/ha, please check the figures for correctness.

Author Response

Thanks for recognizing the value of the conceptual model. This model provides guidance for input data selectionm to estimate forest aboveground biomass using remotely sensed data. We moved the comparisons of model results with previous studies to the discussion section as suggested. We provided the relative RMSE. We double checked the RMSE of the model and that in the figure. The RMSE in the figure is based on the regression only, while the model RMSE is calculated based on model predictions for the out-of-bag samples, thus should be bigger. We added explanation in the manuscript to point this out so that readers will not get confused. 

Reviewer 3 Report

The authors took into account the comments in the first review and calculated the accuracy of the model for different sizes of simulated areas and showed that the model is robust to the choice of sizes of simulated forest areas. This result increases the credibility of the model and allows me to recommend the article for publication.

Author Response

Thanks for your suggestion for generating the effect of sample size on the model accuracy. The suggestion helped improve the quality of the paper

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

In this study, authors used multi-temporal RS datasets collected by multi-sensors to extract attributes of forest canopy for evaluating aboveground biomass of study area. Similar studies rarely use multi-temporal data. However, authors listed the importance of parameters to forest biomass estimation in study sites. The model description and results were not clear. More importantly, I am unsure whether the importance evaluation results of selected parameters to biomass estimation are universal. Besides, authors need to check figures and abbreviations before submission them carefully. As many sections need to be modified and added, I suggest authors carefully revise their manuscript and resubmit it for review. Here are some detailed comments:

• In the Abstract, what is the “EVI”, “VV”, “VH”? In my opinion, some details about remote sensing datasets and study areas were not required in the Abstract. You need to pay attention to some details, such as R2.
• In the introduction, the author described the importance and related concepts of aboveground biomass analysis in detail. Some paragraphs are too long. It is suggested to reduce the content of some sections and clarify the theme of each paragraph.
• Line 122 I think this sentence is not clear. What is the principle? It seems that allometric relationships are not necessarily required for biomass assessment. I think some TLS-based tree structure metrics can be used to estimate biomass, and remote sensing data can reveal the allometric relationships for biomass.
• Line 150 Please unify “LiDAR” and “lidar” throughout the manuscript. I think in this paragraph, you only described some research about ALS.
• Line 173 Texture from high-resolution optical images is highly correlated to tree crown size, which is beneficial to derive DBH. However, this scheme seems to be applicable only to open woodland.
• Line 175 “These remotely sensed data capture information about height, DBH, and species composition for the forests.” The data obtained by active remote sensing technology can extract some related information. But this sentence contradicts some content in Paragraph 3 (Line 117).
• In section 2.1, please list some details of your RS datasets and show the download website of your data in Figure 1, such as the images or other RS datasets. It is not enough to provide the location of your study area on a map. Moreover, the locations of sample plots should also be shown in Figure 1.
• Line 206 “the ninety-fifth percentile height”à“the 95th percentile height”
• Line 225 “a resolution of 2 points per square meter.” I worried that it was difficult to accurately characterize the structure features of canopy using low-density point cloud data. I suggest you verify the accuracy of extracted canopy structure features by field data before estimating FAGB.
• When introducing data source, the download links of open-source data is also necessary.
• Lines 264-266 This sentence can be deleted.
• Punctuation needs to be added after the formula. It is clearer to mark the band number directly in the equation (1) to (3).
• Lines 283-285 Using principal component transformation of the NAIP imagery is beneficial to maintain some helpful information and refine data noise. Here, you need to explain the physical meaning of the first principal component.
• Line 289 “calculate” à“calculated”
• Lines 289-293 I did not understand the proposed method. Even if the ratio-based parameters you extracted were related to the mean stand crown diameter, how to deduce the DBH of individual trees without the species information.
• Figure 2 “Airborne” “LiDAR datasets”. This figure should be remade as it is incomplete.
• Line 330 “derived from different sensors”? This sentence is not clear.
• In Table 1，the sensor/data source column should be added. Besides, the abbreviations listed in this table must be consistent with Figure 2.
• Lines 344-348 Using the abbreviations shown in Table 1 would better describe your results.
• Line 366 I did not understand “the number of variables considered at each split within a tree”.
• In Figure 4, please show the meaning of “IncMSE” to make this figure more clearly.
• In the results section, I did not find any description and related equation of your model but selected parameters.
• Line 460 “very high-resolution data”? I suggest you carefully propose the conclusion related to “first study”.
• For evergreen forests, I think it is not essential to use metrics from multitemporal images to estimate biomass.

Reviewer 2 Report

This paper developed a random forest model for AGB mapping using multisource remotely sensed data, which is a popular topic in the forest and ecology remote sensing field. I understand the author's workload for remote sensing data processing and AGB production mapping for such a large scale in Eastern North Carolina, but I found that this study seems to be simple and common in this field, at least it brings less novelty for me. So I have some major comments on this manuscript.

       The main concern was the evaluation of predictors from multisource data. The importance of variables in a random forest model may be influenced by the collinearity between variables. Although it seems to have little impact on the generalization of the model, it affects readers' understanding of the contribution of multi-source data to biomass mapping. Here I recommend a backward elimination method in random forest variable selection for predictor evaluation. In addition, does the author consider the impact of topography and spatial autocorrelation on AGB modeling?

       Secondly, I considered it will be a good choice to evaluate the prediction performance of multisource remotely sensed data separately, which can guide readers to estimate AGB when some auxiliary data is not available.

       The result part should be a quantitative description of the results of the article, rather than the comparison with others and the speculative expression (e.g. “may be” in Line 357). Some sample points with large deviation in Figure 3 should be further analyzed.

Other specific comments:

1. Scaling bar and compass are missing in Figure 1.
2. Figure 2 is incomplete.
3. Please provide the RMSE and relative RMSE between measured and predicted AGB rather than the evaluation of linear regression between them.

Reviewer 3 Report

The article is generally suitable for publication in the journal Remote sensing. In fact, the authors use the allometric equation for calculating the phytomass, which connected with the average stand height. Additional variables are needed to adjust the model depending on the coefficients of the allometric equation. This, in my opinion, is the correct idea in article.

However, there are the following remarks.

1. Random forests model use methods of supervised learning and the accuracy of the model depends on the amount of data used. For modeling, the authors used inventory data for more than two hundred units and obtain a model with a of determination coefficient about 67%. It is unclear whether the accuracy of the model will increase with increasing training sample sizes. It would be interesting to check the reliability of the model when using a smaller training sample.
2. The scatter of the results of calculations of phytomass values ​​is possibly related to the heterogeneity of the species composition and structure of the analyzed forest stands. A way to test this hypothesis can be the use of a training sample that is more homogeneous in inventory characteristics. Will the accuracy of the model increase with this choice?
3. The authors rightly conclude thatmore studies are needed to further test the conceptual model for above-ground biomass mapping in areas with a broader biomass range and a more diverse species composition. However, it is unclear how accurate the model will be under these conditions. It seems possible to choose a heterogeneous training sampleto test this hypothesis.

 It would be nice if the authors discussed these aspects of the model in the text. Perhaps you should still publish the article as it is without additional analysis, and just mention the above problems in the conclusion.

Reviewer 4 Report

The manuscript submitted by Ehlers et al. represents a state of the art application of a machine learning approach to model biomass using multiple sources of remotely sensed data. It is an interesting approach to biomass mapping with regards to the use of image texture from NAIP imagery in the model and LiDAR derived metrics. I believe this article represents a standard research article that has merits to be published in Remote Sensing, after addressing the following issues:

• The Introduction could be greatly reduced. I found it to long with repetitive sentences throughout stating the same idea sometimes (especially in the beginning when discussing CO2)
• In line 93, change ‘unlike’ to ‘except for’.
• Line 98 to 103 – this part lacks of any citation to important statements.
• Figure 1 needs to have a map of the US or North America showing the relative location of the study area.
• Lines 283-285 – this section needs to be better explained. Include the formula used to calculate the ratio.
• The manuscript falls shorts in comparing their results to other attempts to map FAGB using the same RF-based approach.
• The author refers to the approach as a new algorithm for FAGB mapping, which is not correct. This is a statistical analysis and not the development of an operational algorithm.
• The results show the importance of LiDAR as a predictive variable and disregards the value of using radar and optical imagery in the process, but then reaffirms the need to use multiple sources of data. This is contradictory. I would think, based on your results, that using LiDAR only will be sufficient, and given your discussion on the disadvantages of using optical and radar, I would say there is no need for multiple sources. Please check the narrative of the manuscript and provide better justification for using multiple sources.