Prediction of Aboveground Biomass from Low-Density LiDAR Data: Validation over P. radiata Data from a Region North of Spain
, Javier Maria Sanchez Espeso 4, Jose Manuel Lopez-Guede 2 and Manuel Graña 2Round 1
Reviewer 1 Report
Despite the use of different tests tot es for linear regression models, all tests yielded the same results for all variable combinations, especially r2, adjusted r2, standard error (SE) and Bonferroni outlier test. A comparison in the Discussion (lines 464-470) does not make sense. Such comparisons would have sufficed with results obtained from other allometries for the same area. An alternative would have been from the review of AGB values of the study area from previous studies. An earlier comparison with HAZI foundation sounds more empirical (line 439 - ). The eco-climatic conditions and spp might be different and so their effects of modelled biomass are not factored in. From the linear regression results in Table 2, the inferences from prediction quality shows an almost similar results. There are 3 variables (regressors) derived from 95% percentile – CD corresponding to the 2nd and 3rd layer (p95-tr_2, p95-tr_3) and that from all above mean (p95-allabovemean). From Line 378, the authors have not explained which statistical inferences or the model adopted in selecting the p95-tr_3 apart from this regressor having the highest Breustch-Pagan value of 0.19. With technological advancements in (3D) Lidar data acquisition and the need by research community to improve on the accuracy of reference data for vegetation structure variables, biomass inclusive, the reader would want to know the rationale behind the current study opting for low density Lidar data. This motive is not clearly explained by the authors. A point at hand, the authors in Lines 526 – cites a study finding that “…R2 increased as the point density increased.” The difference between the modelled biomass in this study and the validation dataset from HAZI foundation can be attributed to several errors cited by authors in Lines 537 onwards. How has the current study compensated for these sources of error? Allometric errors are propagative. With additional modelling stage, the error is increased and transferred, unless a method or assumptions are laid down on how to handle these errors. This is important for carbon account purposes especially to ecologists and conservationists in climate related policy formulation. The question would be: Whether factoring in lower Lidar points (2 m) improved the current model From the conclusion, the most relevant biomass predictor variables are found to be height-related. One obvious reason would be because canopy density is not only a variable in the allometric development in Equation 2, but is never a direct variable for most biomass allometries. Does leaving out tree volumetric variable in the current model affect the results?Author Response
Prediction of Aboveground Biomass from Low Density LiDAR data: Validation Over P. Radiata Data From a Region North of Spain
Leyre Torre Tojal, Aitor Bastarrika, Brian Barret, Javier Maria Sanchez Espeso, Jose Manuel Lopez Guede and Manuel Graña
We would like to thank the editor for the careful and timely handling of our paper, and the anonymous reviewers for their valuable comments and suggestions. All of the comments and suggestions have been addressed, as a result of which we believe that our submission has improved considerably.
Reviewer 1
Comments and Suggestions for Authors
Despite the use of different tests tot es for linear regression models, all tests yielded the same results for all variable combinations, especially r2, adjusted r2, standard error (SE) and Bonferroni outlier test.
Response: the phrase is difficult to understand. We believe that the reviewer refers to Table 2, where all values of the coefficient of deviation R2 little variation, in fact are almost identical after truncation. We remind the reviewer that this table contains the test for the 10 best feature subsets, the entire set of combinations of diverse sizes show wide differences of values but we thought the reproduction of the results for all combinations is not necessary and uses too much space. In fact, the important information in the Table 2 are the results of tests with some value diversity across feature combinations.
A comparison in the Discussion (lines 464-470) does not make sense. Such comparisons would have sufficed with results obtained from other allometries for the same area. An alternative would have been from the review of AGB values of the study area from previous studies.
Response: the comparison is meant to show that our work gives results that are within the state of the art ranges reported in the literature. Therefore it is not a direct comparison in the sense of trying to show that we under-perform or over-perform. Specifically, we are not competing with the references. We have changed the writing in order to avoid confusion.
An earlier comparison with HAZI foundation sounds more empirical (line 439 - ).The eco-climatic conditions and spp might be different and so their effects of modelled biomass are not factored in.
Response: the phrase is difficult to understand. It appears to refer to comparisons done by researchers before our own work. There is none that we are aware of. Maybe the meaning of the phrase is that we have not taken into account some effects in section 3.2 that may account for the 8% difference. Thanks for the observation. We think that eco-climatic conditions are longer term that the time scope of our work, so they are not considered, besides the entire modeling approach would need to be remade to account for them. We do not know the meaning of “spp”.
From the linear regression results in Table 2, the inferences from prediction quality shows an almost similar results. There are 3 variables (regressors) derived from 95% percentile – CD corresponding to the 2nd and 3rd layer (p95-tr_2, p95-tr_3) and that from all above mean (p95-allabovemean). From Line 378, the authors have not explained which statistical inferences or the model adopted in selecting the p95-tr_3 apart from this regressor having the highest Breustch-Pagan value of 0.19.
Response: The feature combinations presented in Table 2 are the best found after the exhaustive exploration of all possible combinations of features defined in the appendix. The predictive accuracy measured by R2 is almost the same for them, thus we focus on filling the assumptions of the linear regression models, thus we choose Breustch-Pagan test as the most influential as pointed out by the reviewer. We have changed Line 402 to “For the optimal selection of variables we focus on the Breusch-Pagan test as heteroscedasticity is to be avoided in regression models, thus the selected model is {p95,tr_3}, i.e…..”
With technological advancements in (3D) Lidar data acquisition and the need by research community to improve on the accuracy of reference data for vegetation structure variables, biomass inclusive, the reader would want to know the rationale behind the current study opting for low density Lidar data. This motive is not clearly explained by the authors.
Response: Collecting LiDAR data is costly and cannot be afford easily by researchers, therefore we aim to the exploitation of already available LiDAR data collected for other purposes, such as cartography. In fact, LiDAR data collecting campaigns for cartography are carried out periodically by the institutions and their results are made public free of cost. Hence, they are an extraordinary rich resource for research. The aim of our work is to show that the lower density is not an obstacle for its exploitation for biomass estimation. We have rewritten the abstract (Line 104) in order to highlight this fundamental motivation of the paper. The end of the introduction specifically provides this explanation already.
A point at hand, the authors in Lines 526 – cites a study finding that “…R2 increased as the point density increased.” The difference between the modelled biomass in this study and the validation dataset from HAZI foundation can be attributed to several errors cited by authors in Lines 537 onwards. How has the current study compensated for these sources of error? Allometric errors are propagative. With additional modelling stage, the error is increased and transferred, unless a method or assumptions are laid down on how to handle these errors. This is important for carbon account purposes especially to ecologists and conservationists in climate related policy formulation. The question would be: Whether factoring in lower Lidar points (2 m) improved the current model.
Response: The first part of the comment refers to the studies that have reported the effect of point density in the estimation of biomass. This paragraph serves as justification of our use of low density LiDAR data produced for other purposes. We have rewritten the paragraph initial phrases in order to clarify this. The differences between the HAZI estimation and our own model are not to be considered due to some kind error, but due to variabilities in the computation process, explained in the paragraph. We have rewritten the initial phrase to highlight this nuance. Regarding the accumulation of allometric errors, our approach is direct, does not combine various computational processes, while the HAZI approach estimates the allometric equation input factors from the LiDAR data, thus risking to introduce more computational errors. We have rewritten the paragraph to highlight this. Regarding the question about lower limit on LiDAR points of 2m, we cannot make a direct comparison with HAZI because of the quite different nature of the models, ours is a direct regression model on LiDAR data features, HAZI uses the LiDAR-extracted tree characteristics as inputs for the allometric equation.
From the conclusion, the most relevant biomass predictor variables are found to be height-related. One obvious reason would be because canopy density is not only a variable in the allometric development in Equation 2, but is never a direct variable for most biomass allometries. Does leaving out tree volumetric variable in the current model affect the results?
Response: We have followed a systematic procedure to select the variables of the optimal model. Therefore, the exclusion of the tree volumetric variables is due to a lower accuracy and worse statistical properties regarding the statistical tests reported in Table 2.
Author Response File: 
Author Response.docx
Reviewer 2 Report
General comments:
The manuscript is addressing a highly discussed topic within forestry scientific communities with the implications for forest management practice, such as, the use of low-density LiDAR datasets for forest biomass estimations. Even that the topic is not new, it supports the knowledge improvements of using LiDAR data in the forestry sector.
The introduction section provides sufficient background and it is well documented. The study area is located in the northern part of Spain, in Basque Country, within pine forests of P. radiata. The field data, used as ground truth were collected in 2011, in the Basque Country, on the occasion of the 4th NFI. The LiDAR data were acquired in 2012 by a Lite Mapper 6800 ALS device. What I missed are some words about the changes of the forests between the inventory year and the LiDAR acquisition year. The methods and statistical analysis are used appropriately. I suggest you to add a figure with the entire workflow of the methodology, it will be much easier for the reader to follow the entire manuscript. The results are presented in an appropriate form. You are referring to R2 but you give also R2 adj. (see Table 2), I suggest replacing R2 with R2 adjusted in the statistical analysis.
In the discussion section, in comparison with their results, the authors present similarities with other studies. The conclusions section of the manuscript need improvements. I suggested move some parts to the discussion section (see the specific comments). No references should be in the conclusion sections.
Specific comments:
L 167-168: Please insert figures with the distributions (histograms) of the two variables (dbh, h)
L 231 : Figure 7 – Replace cleaning with filtering. Fusion metrics do not relate to anything? You are not using it in the further analysis?
L266 : use the adjusted coefficient of determination (R2 adj)
L279-280: “To minimize the bias introduced in the model, a correction factor was applied, which depended on the standard error of the estimate (SEE) (Eq. 11)” Please insert eq.11 properly.
L317: Figure 8 is similar with figure 1, please indicate a) and b) in the figure.
L363: In case of using R2adj the same value is 0.79, not 0.80
L367 I suggest using R2adj instead of R2
L397 Figure 11b use lnBiomass instead lnBiomasa
L471, 483, 487,491: you use RMSE for the comparation with other studies, please insert in methodology section a reference to RMSE
L559: Use significantly instead of drastically
L563-565, 569-570: please move to the discussion section, the conclusions should refer to your results and the comparison with other studies should be made in the discussion section.
Author Response
Prediction of Aboveground Biomass from Low Density LiDAR data: Validation Over P. Radiata Data From a Region North of Spain
Leyre Torre Tojal, Aitor Bastarrika, Brian Barret, Javier Maria Sanchez Espeso, Jose Manuel Lopez Guede and Manuel Graña
We would like to thank the editor for the careful and timely handling of our paper, and the anonymous reviewers for their valuable comments and suggestions. All of the comments and suggestions have been addressed, as a result of which we believe that our submission has improved considerably.
Reviewer 2
Comments and Suggestions for Authors
General comments:
The manuscript is addressing a highly discussed topic within forestry scientific communities with the implications for forest management practice, such as, the use of low-density LiDAR datasets for forest biomass estimations. Even that the topic is not new, it supports the knowledge improvements of using LiDAR data in the forestry sector.
The introduction section provides sufficient background and it is well documented. The study area is located in the northern part of Spain, in Basque Country, within pine forests of P. radiata. The field data, used as ground truth were collected in 2011, in the Basque Country, on the occasion of the 4th NFI. The LiDAR data were acquired in 2012 by a Lite Mapper 6800 ALS device. What I missed are some words about the changes of the forests between the inventory year and the LiDAR acquisition year.
Response: A paragraph commenting this fact is added at the end of the discussion section (Line 608).
The methods and statistical analysis are used appropriately. I suggest you to add a figure with the entire workflow of the methodology, it will be much easier for the reader to follow the entire manuscript.
Response: Figure 7a details the workflow of the LiDAR data. We include Figure 7b with an overall diagram of the research study. An explanation is added as last paragraph of 2.6.2
The results are presented in an appropriate form. You are referring to R2 but you give also R2 adj. (see Table 2), I suggest replacing R2 with R2 adjusted in the statistical analysis.
In the discussion section, in comparison with their results, the authors present similarities with other studies. The conclusions section of the manuscript need improvements. I suggested move some parts to the discussion section (see the specific comments). No references should be in the conclusion sections.
Specific comments:
L 167-168: Please insert figures with the distributions (histograms) of the two variables (dbh, h). Response: Distribution of the diameter and height has been displayed using box-plots in Figure 5. We believe that they provide a better summary than histograms.
L 231 : Figure 7 – Replace cleaning with filtering.
Response: Done in Figure 7a
Fusion metrics do not relate to anything? You are not using it in the further analysis?
Response: We calculate FUSION metrics on one hand, and density metrics, using PostGIS, on the other hand. With this two collection of variables we choose the ones whose combinations show greater R2adj values.
L266 : use the adjusted coefficient of determination (R2 adj).
Response: Done in Line 285.
L279-280: “To minimize the bias introduced in the model, a correction factor was applied, which depended on the standard error of the estimate (SEE) (Eq. 11)” Please insert eq.11 properly.
Response: In our version of the manuscript the reference is to Eq. 4.
L317: Figure 8 is similar with figure 1, please indicate a) and b) in the figure. The figure 1 has been replaced by a more general picture. Captions are changed to explain inset and picture in Figure 1 and Figure 8.
L363: In case of using R2adj the same value is 0.79, not 0.80. Done in Line 390.
L367 I suggest using R2adj instead of R2. This commentary has not been found in this line number, but all the R2 has been replaced to R2 adj in this paragraph.
L397 Figure 11b use lnBiomass instead lnBiomasa. Done in Figure 11b.
L471, 483, 487,491: you use RMSE for the comparation with other studies, please insert in methodology section a reference to RMSE.
Response: definition and reference are at the end of section 2.6.4 (Line 355).
L559: Use significantly instead of drastically. Done in Line 621.
L563-565, 569-570: please move to the discussion section, the conclusions should refer to your results and the comparison with other studies should be made in the discussion section.
Response: Conclusions ad discussion have been corrected along the suggested lines.
Author Response File: Author Response.docx
