Identifying Forest Structural Types along an Aridity Gradient in Peninsular Spain: Integrating Low-Density LiDAR, Forest Inventory, and Aridity Index

Round 1

Reviewer 1 Report

Dear editor,

The text is OK and I recommend for acceptance. I just ask the authors to take care of the graph figures the official paper is produced, because some of them might be too large (e.g., Fig 1 and plot in the supplementary). Maybe these figures could be reedited in this sense. Overall, it is OK.

L214. Please reshape this sentence in parenthesis. 

Author Response

Comments and Suggestions for Authors Reviewer #1

The text is OK and I recommend for acceptance. I just ask the authors to take care of the graph figures the official paper is produced, because some of them might be too large (e.g., Fig 1 and plot in the supplementary). Maybe these figures could be reedited in this sense. Overall, it is OK.

We appreciate the reviewer’s positive feedback and suggestions. We took into consideration the recommendations and we double-checked the figures and graphs in the manuscript. We reedited them according to the right dimensions.

L214. Please reshape this sentence in parenthesis. 

Done.

Author Response File: Author Response.docx

Reviewer 2 Report

The current study comprises the definition of 4 types of forest structure from clustering forest attribute data of the National Forest Inventory. Authors also perform a “random forest” classification to describe their forest structure types with LiDAR derived metrics to finally obtain a wall to wall cartography of forest structure. The general focus of methodology using different cluster and classification algorithms is also proposed in other works. Their methodology is therefore solid enough and they contribute with a novelty using different algorithms. They work with huge amount of data, i.e. forest inventory plots and lidar data, which is also remarkable.

Main objection to the manuscript comes from the meaning and interpretation of their forest structure types. Forest structures in order to provide useful significance and information should be labeled according to vertical strata and/or age and/or cover: single layered, multi-layered, young, mature, even-aged, uneven-aged, reversed J, open canopy cover, etc.. (this can be seen in some of manuscript references: [17], [26], [43] [84]). To have this, authors should select some information regarding distribution of trees according to DBH or height (something synthetizing histogram of DBH or height). The mentioned references [17], [26], [43] provide some examples: Gini Coefficient, % of the total number of trees×ha^-1 with diameter DBH ∈ (7.5 cm; 17.5 cm) (non-commercial wood), Lorey´s height or height strata might be interesting.

Therefore, the forest structure attributes selection to define forest structure types weighs on the LiDAR data results. Authors map or classify the 4 provinces according to some classes or types that have no easy interpretation and therefore these classes don´t have too much interest to forest or territorial managers.

Authors should improve their forest attributes selection to fed the k-medois algorithm. Alternatively they  could also coherently try to interpretate and label their current 4 types using DBH, N, etc.. values. Although this might not be easy to do in a proper way. Providing additional results usually also requires improving  discussion.

Some minor concers are:

• Authors should include in the introduction some lines as state of art of forest structure mapping.
• 1 caption is too much synthetic. Martone aridity index values?, range?
• Line 194. I suppose it should be “…discriminate forest structural types” instead of “…discriminate fuel structural types”.
• Figure 2 is the “wall-to-wall map of structure types”. I suggest to locate the figure in the results epigraph.
• Lines 386 – 387 Lidar variables: CC2M, MCHP, Dif 90_50, etc.. should be defined in the text (not only in supplementary material).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments to the Author

This paper focuses on the integration of lidar, field inventory and aridity index data to characterise the forest structure using a random forest model. There are several papers published on a similar topic but the idea of integrating low-density point data to classify many forest species is unique. I would recommend this paper for publication as it has already incorporated most of the comments from previous reviewers however, I would like to urge for some minor revisions to be incorporated as mentioned below.

Comments:

1. The title of paper should be changed by replacing “climatic data” with “aridity index” as the authors haven’t used any data than the aridity index.
2. First sentence of the abstract is a bit confusing so please split in two.
3. Details of lidar specification should be placed in the appendix.
4. Caption of Figure 5 should be written as “Figure 5” in page 11.
5. If possible, please mention the variance explained by the Random Forest model.
6. There are lots of most relevant, recent and impressive literature for identifying/classifying forest structural types using lidar data so please incorporate them in the introduction and discussion section. These references demonstrate the recent work on the vertical structure and distribution profile of the forests at the tree, stand and landscape scale for example, (Wilkes et al., 2015), (Woodgate et al., 2015), (Karna et al., 2019; Karna et al., 2020), (Kane, 2010; Kane et al., 2010a; Kane et al., 2010b; Kane et al., 2013; Kane et al., 2014; Kane et al., 2017) The bibliography of the citations is mentioned below.

Kane, V.R., Bakker, J.D., McGaughey, R.J., Lutz, J.A., Gersonde, R.F., Franklin, J.F., 2010a. Examining conifer canopy structural complexity across forest ages and elevations with LiDAR data. Canadian Journal of Forest Research-Revue Canadienne De Recherche Forestiere 40, 774-787.

Kane, V.R., McGaughey, R.J., Bakker, J.D., Gersonde, R.F., Lutz, J.A., Franklin, J.F., 2010b. Comparisons between field- and LiDAR-based measures of stand structural complexity. Can. J. For. Res. 40, 761-773.

Kane, V.R., North, M.P., Lutz, J.A., Churchill, D.J., Roberts, S.L., Smith, D.F., McGaughey, R.J., Kane, J.T., Brooks, M.L., 2014. Assessing fire effects on forest spatial structure using a fusion of Landsat and airborne LiDAR data in Yosemite National Park. Remote Sensing of Environment 151, 89-101.

Karna, Y.K., Penman, T.D., Aponte, C., Bennett, L.T., 2019. Assessing Legacy Effects of Wildfires on the Crown Structure of Fire-Tolerant Eucalypt Trees Using Airborne Lidar Data. Remote Sens. 11 (20), 2433.

Karna, Y.K., Penman, T.D., Aponte, C., Hinko-Najera, N., Bennett, L.T., 2020. Persistent changes in the horizontal and vertical canopy structure of fire-tolerant forests after severe fire as quantified using multi-temporal airborne lidar data. Forest Ecology and Management 472, 118255.

Wilkes, P., Jones, S.D., Suarez, L., Haywood, A., Mellor, A., Woodgate, W., Soto‐Berelov, M., Skidmore, A.K., 2015. Using discrete‐return airborne laser scanning to quantify number of canopy strata across diverse forest types. Methods in Ecology and Evolution 7, 700-712.

Woodgate, W., Jones, S.D., Suarez, L., Hill, M.J., Armston, J.D., Wilkes, P., Soto-Berelov, M., Haywood, A., Mellor, A., 2015. Understanding the variability in ground-based methods for retrieving canopy openness, gap fraction, and leaf area index in diverse forest systems. Agricultural and Forest Meteorology 205, 83-95.

Author Response

Comments and Suggestions for Authors Reviewer #3

This paper focuses on the integration of lidar, field inventory and aridity index data to characterise the forest structure using a random forest model. There are several papers published on a similar topic but the idea of integrating low-density point data to classify many forest species is unique. I would recommend this paper for publication as it has already incorporated most of the comments from previous reviewers however, I would like to urge for some minor revisions to be incorporated as mentioned below.

Comments:

We appreciate the comments. We added the following information in the manuscript:

1.The title of paper should be changed by replacing “climatic data” with “aridity index” as the authors haven’t used any data than the aridity index.

L2-4:

Identifying forest structural types along an aridity gradient in peninsular Spain: integrating low-density LiDAR, forest inventory, and aridity index.

2.First sentence of the abstract is a bit confusing so please split in two.

L11-12:

Done. The new sentence reads: “Forest structure is a key driver of forest functional processes. The characterization of forest structure across spatiotemporal scales is essential for forest monitoring and management.”

3.Details of lidar specification should be placed in the appendix.

We included a table in the appendix with som LiDAR specification:

4.Caption of Figure 5 should be written as “Figure 5” in page 11.

Done.

5.If possible, please mention the variance explained by the Random Forest model.

We added the variance explained by the Random Forest model in the manuscript:

L295-296:

The variance explained by the RF model reached 61.73%.

6.There are lots of most relevant, recent and impressive literature for identifying/classifying forest structural types using lidar data so please incorporate them in the introduction and discussion section. These references demonstrate the recent work on the vertical structure and distribution profile of the forests at the tree, stand and landscape scale for example, (Wilkes et al., 2015), (Woodgate et al., 2015), (Karna et al., 2019; Karna et al., 2020), (Kane, 2010; Kane et al., 2010a; Kane et al., 2010b; Kane et al., 2013; Kane et al., 2014; Kane et al., 2017) The bibliography of the citations is mentioned below.

We incorporated the references suggested by the reviewer#3 at the manuscript.

Literature cited:

Torresan, C., Corona, P., Scrinzi, G., & Marsal, J. V. (2016). Using classification trees to predict forest structure types from LiDAR data. Annals of Forest Research, 59(2), 281-298.

Adnan, S., Maltamo, M., Coomes, D. A., García-Abril, A., Malhi, Y., Manzanera, J. A., ... & Valbuena, R. (2019). A simple approach to forest structure classification using airborne laser scanning that can be adopted across bioregions. Forest Ecology and Management, 433, 111-121.

Lin, H. T., Lam, T. Y., von Gadow, K., & Kershaw Jr, J. A. (2020). Effects of nested plot designs on assessing stand attributes, species diversity, and spatial forest structures. Forest Ecology and Management, 457, 117658.

Barbati, A., Marchetti, M., Chirici, G., & Corona, P. (2014). European forest types and forest Europe SFM indicators: tools for monitoring progress on forest biodiversity conservation. Forest Ecology and Management, 321, 145-157.

Peng, C. (2000). Growth and yield models for uneven-aged stands: past, present and future. Forest ecology and management, 132(2-3), 259-279.

Moreno-Fernández, D., Álvarez-González, J. G., Rodríguez-Soalleiro, R., Pasalodos-Tato, M., Cañellas, I., Montes, F., ... & Pérez-Cruzado, C. (2018). National-scale assessment of forest site productivity in Spain. Forest Ecology and Management, 417, 197-207.

Reque, J. A., & Bravo, F. (2008). Identifying forest structure types using National Forest Inventory Data: the case of sessile oak forest in the Cantabrian range. Investig. Agrar. Sist. Recur. For, 17, 105-113.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors made an effort to provide a label for their forest types, which it improves their applicability.

In the previous review, I suggested that “Authors should include in the introduction some lines as state of art of forest structure mapping”. This suggestion involved implicitly to provide arguments to highlight the original contribution that any article requires to be published. 

I think introduction should provide the right context and highlight the novelty of their work. As suggestion, they can use the provided arguments they included in authors´ reply letter.

Some paragraphs extracted from their letter are:  
…” We agree with this comment and we understand the reviewer’s point, deeper information on forest structure could provide more accurate forest types. Nevertheless, the spatial scale considered in this study is a limiting factor since we are contemplating forest structure at regional scale”….. “Here, we used size (dbh and height) and density variables (basal area and number of trees per ha) to discriminate forest groups and we relate these groups to tree composition, which is an attribute of forest structure (Lin et al. 2020). In fact, studies classifying forest types at broad scales used forest composition as the main discriminant factor (Barbati et al. 2014)”…..

…” Our classification can be understood in four broad categories that are relevant to establish management and to implement mitigation and adaptation measures in Iberian forests”….” 
…”These four structural typologies suggest divergent functional performance in terms of exposure and vulnerability (sensitivity and adaptive capacity) in response climate change hazards (chiefly hot spells and intense droughts, windstorms and wildfires). Therefore, their definition can be assist to the definition of adaptation priority regions (i.e. targeting thinning in order to decrease drought vulnerability) (T3) or alternative plantations to foster recruitment (Type 2)”.

Main ideas that can be extracted from previous authors lines (that might be incorporated to introduction) are:
-    main-novel contribution of their work is mapping forest structure types (maybe just "forest types") at regional (large) scale. 
-    The relevance and utility of large scale forest types is to assist forest management decisions to implement climate change mitigation and adaptation measures.

State of art (references) should support these ideas in case they consider them. 
 

Author Response

We appreciate the reviewer’s suggestions and the positive feedback.

In order to implement the suggestions made by the reviewer about main-novel contributions to mapping forest types at a large scale and the relevance of a large-scale forest type definition to assist forest management decisions, we added some lines in the manuscript:

L46-52:

“It is remarkable that these large scales can be utilized in order to establish and improve management, and to implement mitigation and adaptation measures in Iberian forests [23]. It is also necessary to understand the divergent functional performance of these forests in terms of exposure and vulnerability (sensitivity and adaptive capacity) in response to climate change hazards (chiefly hot spells and intense droughts, windstorms, and wildfires) [24,25]. Therefore, forest structure definitions can assist in outlining adaptation priority regions (e.g., targeting thinning in order to decrease drought vulnerability or pinpoint alternative plantation sites to foster recruitment).”

L80-85:

“These previous studies were limited to local sites, yet the PNOA LiDAR coverage offers an excellent opportunity to provide accurate, detailed, wall-to-wall (covering the whole terrain) information on forest structure regionally (large scales), capturing its spatial heterogeneity in a way previously unachievable. It is important to highlight that improving the knowledge of this forest structural variation at large scales is essential for understanding the processes that drive forest dynamics, habitat variability, and biodiversity [72,73], along with nutrient and carbon dynamics [74,75] in forested landscapes.”

Author Response File: Author Response.docx

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

This study integrates LiDAR data into forest structure recognition, which is a very meaningful manuscript. This manuscript integrates survey data and LiDAR data, and systematically studies the application of LiDAR data to forest structures. At present, there are still some minor errors in the manuscript, such as the formula (1) of Line 88. I suggest that the author systematically correct these small errors to improve the readability of the manuscript.

Author Response

Dear Reviewer:

We are very pleased with the interest you have taken in our work and the comments you made which we agree with.

Comments and Suggestions for Authors Reviewer 1

This study integrates LiDAR data into forest structure recognition, which is a very meaningful manuscript. This manuscript integrates survey data and LiDAR data, and systematically studies the application of LiDAR data to forest structures. At present, there are still some minor errors in the manuscript, such as the formula (1) of Line 88. I suggest that the author systematically correct these small errors to improve the readability of the manuscript.

Thanks for this comment. We have revised the entire manuscript and corrected these errors as that you pointed out in L88

Author Response File: Author Response.docx

Reviewer 2 Report

This ms has a good amount of details and is easy to follow. However, I still have several concerns and suggestions which hopefully can help the authors improve the quality of the study.

First, lidar has been used in forest structure and tree type estimation for over two decades. To be honest, the methods in the ms are quite typical. But the authors mentioned the evaluation of low-density point clouds, which has the potential advance national forest inventory. Here, the authors simply tested three point densities in four provinces. The question is how point density is linked to model performance? Could you add some quantitative analyses to help the practitioners understand the relationship? And what is defined as low density? The three densities used in the study do not seem adequate. Why didn’t you test more densities? This could be done by resampling the point clouds. In addition, a review of the lidar point-density effect on forest structure estimation should be added to the Intro section.

Second, the accuracy of the results is low. Do they meet the SNFI needs? I think you need to carefully justify the usefulness of the results. You might want to give more emphasis to my first comment, and find the relationship between accuracy and point density. The shift of emphasis may help you better justify the usefulness of the findings.

Third, multiple things are missing or are placed in the wrong sections. For example, lidar metrics should be described in detail in the methods section. You should also give more details about random forest modeling, e.g., parameterization, and samples for calibration and validation.   

Figures 3&4: Please add a scale bar. You have two Type 3 structure types in Figure 3.

Author Response

RESPONSE TO THE SUGGESTIONS MADE FROM REVIEWER 2:

Dear Reviewer:

We are very pleased with the interest you have taken in our work and the comments you made which we agree with. Next, we are going to explain how your comments have been taken into account:

Comments and Suggestions for Authors Reviewer 2

This ms has a good amount of details and is easy to follow. However, I still have several concerns and suggestions which hopefully can help the authors improve the quality of the study.

First, lidar has been used in forest structure and tree type estimation for over two decades. To be honest, the methods in the ms are quite typical. But the authors mentioned the evaluation of low-density point clouds, which has the potential advance national forest inventory. Here, the authors simply tested three point densities in four provinces. The question is how point density is linked to model performance? Could you add some quantitative analyses to help the practitioners understand the relationship? And what is defined as low density? The three densities used in the study do not seem adequate. Why didn’t you test more densities? This could be done by resampling the point clouds. In addition, a review of the lidar point-density effect on forest structure estimation should be added to the Intro section.

Thanks for this comment. We would like to indicate that we did not aim at testing three point densities but using two open-source data (PNOA-LiDAR project) and NFI to identify forest structures. The point density of PNOA-LiDAR project varies among Spanish provinces (from 0.5 to 4 points m^-2). Additionally, the NFI is in progress and, therefore, is not completed. All of this restricts the number of provinces to be used as well as the point density. Regarding the adequacy of the point densities used, we did not have the chance of testing different point densities as we are using open-access data. In order to appropriately evaluate the impact of point density on the results, it is necessary to carry out specific surveys. Although a common approach to testing different point densities is to randomly remove points from the dataset to reach the lowest density, this approach is not exempt from errors since in an operative context, point density is changed by changing PRF, flight elevation, etcetera. This implies other parameters such as footprint size, energy, co-vary with point density, etc. (see García et al., 2017 Carbon Balance Manage (2017) 12:4 DOI 10.1186/s13021-017-0073-1

With respect to the impact of point density on model performance, we honestly think that evaluating the effect of point density is out of the scope of the study. There are multiple examples where different point densities LiDAR are tested to see how these different densities affect the model performance (e.g. Ruiz et al., 2014; Jakubowski et al., 2013; Garcia et al. 2017). However, our study aims at identifying forest structure changes along an aridity gradient and checking if it is possible to classify structural groups using low point density LiDAR data more than test how different point densities affect our model.

Concerning the definition of low point density, we used the classification of LiDAR density defined by Dr. Felix Rohrbach (https://felix.rohrba.ch/en/2015/point-density-and-point-spacing/), where point densities between 0.5-1 point m^-2 are defined as “sparse point densities”, densities between 1-2 point m^-2 are defined as “low point densities”, densities between 2-5 point m^-2 are defined as “medium point densities”, densities between 5-10 point m^-2 are defined as “high point densities, and densities over 10 point m^-2 are defined as “extremely point densities”.

In order to fulfill the reviewer’s suggestion, we have added the following lines in the introduction (L 58-64):

“However, although LiDAR has proven to be a capable tool for directly measuring structural variables (e.g., tree height), and modelling some others (e.g., aboveground biomass), its accuracy depends heavily on flight characteristics, especially on point density [46–48]. However, some authors have pointed out that plot size has a greater effect than LiDAR point density. In general, minimum plot areas of 500–600 m^-2 are needed for volume, biomass and basal area estimates, and of 300–400 m^-2 for canopy cover. Larger plot sizes do not significantly increase the accuracy of the models [49]. Also, it is remarkable that the accuracy of low point density LiDAR worsens when we shift from plot to landscape scales [50]”

Ruiz, L. A., Hermosilla, T., Mauro, F., & Godino, M. (2014). Analysis of the influence of plot size and LiDAR density on forest structure attribute estimates. Forests, 5(5), 936-951.

Jakubowski, M. K., Guo, Q., & Kelly, M. (2013). Tradeoffs between lidar pulse density and forest measurement accuracy. Remote Sensing of Environment, 130, 245-253.

Garcia, M., Saatchi, S., Ferraz, A., Silva, C. A., Ustin, S., Koltunov, A., & Balzter, H. (2017). Impact of data model and point density on aboveground forest biomass estimation from airborne LiDAR. Carbon balance and management, 12(1), 1-18.

Second, the accuracy of the results is low. Do they meet the SNFI needs? I think you need to carefully justify the usefulness of the results. You might want to give more emphasis to my first comment, and find the relationship between accuracy and point density. The shift of emphasis may help you better justify the usefulness of the findings.

Regarding the accuracy level of our study, our results are similar to other studies using similar datasets in Mediterranean environments. The main contribution is to provide spatially explicit information on forest structural types to forest managers. This variability at the landscape scale is not appropriately captured by NFI.

Moreover, we have run the Random Forest algorithm for each province. The following table shows the results for each province independently and it can be seen that overall accuracy did not change significantly among sites.

Third, multiple things are missing or are placed in the wrong sections. For example, lidar metrics should be described in detail in the methods section. You should also give more details about random forest modeling, e.g., parameterization, and samples for calibration and validation.   

We agree with the reviewer’s suggestion.

A LiDAR metrics reference has been pointed out in L (185-186):

“A summary of the LiDAR metrics are detailed in Table A3 (Supplementary material).”  

We have also added more information about the random forest modeling in L (193 – 196):

“A sample of 70% of the SNFI-4 plots (n = 1938 plots) were used for training the model and the rest 30% of the SNFI-4 plots (n = 832 plots) were used for the validation process. The RF model was set to grow to 500 trees while the parameter of mtry (number of variables randomly sampled as candidates at each split) was one third of the variables.”

Figures 3&4: Please add a scale bar. You have two Type 3 structure types in Figure 3.

We have modified both figures according to the comments of the reviewer.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear editor,

This manuscript aimed to classify forest areas according to the aridity index using LiDAR data and Random Forest model. A success of about 65% of accuracy was found for the classification considering four areas defied by them using clustering methods (k-medoids). The work is interesting, opportune, and straightforward, but major corrections are required.

I’m pointing below some main comments and other specific comments are listed later.

1. In the text you mentioned Figure A, Table A1, and A2 but it is not available in the text. I don’t know if they are available as supplementary, or you just forgot to add them in the text. By taking from the manuscript as it is, I could not understand the results correctly because of that. Since you have the maps of the area, you must also provide the classification maps according to your model.

2. Regarding what was said in L178: airborne LiDAR scanners are quite accurate nowadays, so the LiDAR-based estimates are more related to other factors. The point density is just one of these factors, while others might be regarding the modeling process (the ALS metrics, modeling approach, training data size), plot co-registration, forest composition, etc. A brief comment about these must be made in the introduction or discussion.

3. If I understood correctly, the Spanish NFI data used plots with considerably different sizes, varying from 80 to 2000 m² (L130). And then, you predicted values for pixels of 2500 m² (L183). Although your ABA models were used for classification and not for regression, you should comment why or how this difference in plot size would not affect your estimates. Please check this paper: Packalen et al 2019 Resolution dependence in an area-based approach to forest inventory with airborne laser scanning (https://0-doi-org.brum.beds.ac.uk/10.1016/j.rse.2019.01.022)

4. You should also use Kappa Index to measure classification accuracy because many works use it too. This would help to make comparisons.

5. Please double-check the format of all scientific names used in the text. Put them in the correct form (italic).

Specific comments:

L59. (e.g., tree height) ?

L94. Brackets are “[x]” symbols, here you are using parenthesis “(X)”.

L130. Were these plots concentric, i.e., had the same plot centers, or were they clustered? Please make it clear.

L146. Using auxiliary scripts is not a problem at all, but I wonder why did you use MATLAB to apply a simple threshold to separate trees from understory? In other words, why not just computing the metrics considering the 2-m-threshold in FUSION?

L154. Please consider moving figure 2 to the end of the methods. We cannot understand it correctly before reding all the processes.

L192. How did you split the dataset, randomly? Besides, why don’t you used iterative approaches such as a 5-fold-cross validation? It is very likely that if you had used such an approach your accuracy for this model (L263) would be increased.  

L221. Please drop this sentence “depicted the spatial distribution…. SNFI-4 data”.

L224. Was this figure produced with the predicted or observed plot class? Please clarify it and add this information in the figure caption too.

L248. This sentence is not clear “we can observe how T3 was the structural type most widely distributed across our study zones”. It looks like a discussion/conclusion sentence. Here you must describe this fact first so the reader can understand ‘how’ T3 class was the most widely distributed.

L225. Since aridity is the core of this work, the aridity index M should be key to order this table. For instance, it would be much more intuitive to call the aridest regions (here the type 4) as type 1 lest arid (here the type 3) as type 4. Please consider changing it and adapting the text accordingly.

L250. Confusing. What “a relative distribution” would be?

L256. Please, use intuitive names to refer to these metrics (e.g.: CC2m for canopy cover above 2m, and not ‘A2/FR’). Also, give more details about how their definition (because it is just six, you could even provide their formulas).

L262. “Producer’s accuracy”?? Shouldn’t be just “the accuracy”? See also “user’s accuracy” in (L264) and tables 4 and 5. What are user and producer accuracy? You did not mention it in the methods.

L264. Why are you calling Figure 3 here? Was it built using predicted classes? Please clarify this.

L274. Which “both cases” are you referring to?

L275. Where are Figure A1 and table A1 and A2?

L318. Seems that Gorgens is also using ref 87 and 88 for this statement, but it was part of his conclusions. Please modify this reference to something like: “(see also [87] and [88])”

417. Here you are referring to your own work, but then you call four other references, which seems to me that they were who made this choice for you. Please correct it.

Author Response

RESPONSE TO THE SUGGESTIONS MADE FROM REVIEWER 3:

Dear Reviewer:

We are very pleased with the interest you have taken in our work and the comments you made which we find very useful and will help improve the quality of the ms. Next, we are going to explain how your comments have been taken into account:

Comments and Suggestions for Authors Reviewer 3

Dear editor,

This manuscript aimed to classify forest areas according to the aridity index using LiDAR data and Random Forest model. A success of about 65% of accuracy was found for the classification considering four areas defied by them using clustering methods (k-medoids). The work is interesting, opportune, and straightforward, but major corrections are required.

I’m pointing below some main comments and other specific comments are listed later.

1. In the text you mentioned Figure A, Table A1, and A2 but it is not available in the text. I don’t know if they are available as supplementary, or you just forgot to add them in the text. By taking from the manuscript as it is, I could not understand the results correctly because of that. Since you have the maps of the area, you must also provide the classification maps according to your model.

As you indicate, Figure A1, Table A1, and A2 are in supplementary material. We have added this information in the ms. We have included the classification maps according to our model in Figure 3.

1. Regarding what was said in L178: airborne LiDAR scanners are quite accurate nowadays, so the LiDAR-based estimates are more related to other factors. The point density is just one of these factors, while others might be regarding the modeling process (the ALS metrics, modeling approach, training data size), plot co-registration, forest composition, etc. A brief comment about these must be made in the introduction or discussion.

We agree with the reviewer’s suggestion. We have added in the discussion the following lines (L 450 - 453):

“Regarding data accuracy, airborne LiDAR scanners are highly accurate nowadays, so the LiDAR-based estimates are connected to factors like point density, modeling process (ALS metrics, modeling approach, training data size), plot co-registration, forest composition, among other factors. In this direction, we had into account those requirements in order to ensure a correct design and treatment of LiDAR data.”

            and these in the Introduction (L 58 – 64):

“However, although LiDAR has proven to be a capable tool for directly measuring structural variables (e.g., tree height), and modelling some others (e.g., aboveground biomass), its accuracy depends heavily on flight characteristics, especially on point density [46–48]. However, some authors have pointed out that plot size has a greater effect than LiDAR point density. In general, minimum plot areas of 500–600 m^-2 are needed for volume, biomass and basal area estimates, and of 300–400 m^-2 for canopy cover. Larger plot sizes do not significantly increase the accuracy of the models [49]. Also, it is remarkable that the accuracy of low point density LiDAR worsens when we shift from plot to landscape scales [50]”

1. If I understood correctly, the Spanish NFI data used plots with considerably different sizes, varying from 80 to 2000 m² (L130). And then, you predicted values for pixels of 2500 m² (L183). Although your ABA models were used for classification and not for regression, you should comment why or how this difference in plot size would not affect your estimates. Please check this paper: Packalen et al 2019 Resolution dependence in an area-based approach to forest inventory with airborne laser scanning (https://0-doi-org.brum.beds.ac.uk/10.1016/j.rse.2019.01.022)

The Spanish NFI follows a nested design, i.e., concentric plots. This design aims at reducing the sampling effort (and cost) of small trees. Spanish NFI uses plots with 25m radius as maximum and, therefore, there are concentric plots with variable radius where it measures trees depending on their dbh. We mean, the concentric field plots of variable radii measure trees depending on the diameter at breast height (DBH): a radius of 25 m for trees with DBH ≥ 42.5 cm, a radius of 15 m for trees with DBH ≥ 22.5 cm, a radius of 10 m for trees with DBH ≥ 12.5 cm, and a radius of 5 m for trees with a DBH ≥ 7.5 cm. Trees with 2.5 ≤ DBH ≤ 7.5 cm are counted but not measured. Plot variables were calculated from tree measurements at plot level by using expansion factors. Therefore, plot size is equal for all the sampling points.

We have modified the text as follows (L 131 - 133):

“The SNFI project establishes plots at 1×1 km UTM grid with a concentric system plot model. The concentric field plots of variable radii measure trees depending on the diameter at breast height (DBH):”

1. You should also use Kappa Index to measure classification accuracy because many works use it too. This would help to make comparisons.

Although the kappa index is commonly used and presented with confusion matrices, there is some controversy about its suitability and it has been suggested to abandon its use (Pontius and Millones; 2011- Death to Kappa: Birth of quantity disagreement and allocation disagreement for accuracy assessment; https://0-doi-org.brum.beds.ac.uk/10.1080/01431161.2011.552923). These authors also stated that the inclusion of kappa never resulted in a change of the conclusion reached by using the overall accuracy.

Still, we calculated it and the results are shown in the table below. As can be seen, the Kappa index does not alters substantially what has already been stated.

1. Please double-check the format of all scientific names used in the text. Put them in the correct form (italic).

 We have gone throughout the manuscript and revised the scientific names.

Specific comments:

L59. (e.g., tree height) ?

Done.

L94. Brackets are “[x]” symbols, here you are using parenthesis “(X)”.

Done.

L130. Were these plots concentric, i.e., had the same plot centers, or were they clustered? Please make it clear.

Yes, Spanish NFI plots are concentric. Please, see the response to the main comment number 3.

L146. Using auxiliary scripts is not a problem at all, but I wonder why did you use MATLAB to apply a simple threshold to separate trees from understory? In other words, why not just computing the metrics considering the 2-m-threshold in FUSION?

We probably did not explain ourselves clearly in this regard. MATLAB was not simply used to apply a height threshold but to compute additional metrics to those derived from FUSION.

L148-L151: Added to the ms

“Canopy LiDAR metrics for each SNFI-4 plot were extracted using FUSION/LDV software [69], as well as several scripts developed in MATLAB. We applied a predefined height threshold of 2m to separate trees from understory vegetation with FUSION/LDV as well as with MATLAB.”

L154. Please consider moving figure 2 to the end of the methods. We cannot understand it correctly before reding all the processes.

Done

L192. How did you split the dataset, randomly? Besides, why don’t you used iterative approaches such as a 5-fold-cross validation? It is very likely that if you had used such an approach your accuracy for this model (L263) would be increased.  

We split the data by randomly selecting 70% of the sample for training and the remaining 30% for independent validation. The data set was randomly split. We pointed it out in L (192-193): “Data set was randomly split”.

By leaving an independent dataset for validation, we can provide a more robust validation of our models.

L221. Please drop this sentence “depicted the spatial distribution…. SNFI-4 data”.

Done

L224. Was this figure produced with the predicted or observed plot class? Please clarify it and add this information in the figure caption too.

This figure was produced with the predicted and balances classes. We clarified and added the information in the figure.

L225. Since aridity is the core of this work, the aridity index M should be key to order this table. For instance, it would be much more intuitive to call the aridest regions (here the type 4) as type 1 lest arid (here the type 3) as type 4. Please consider changing it and adapting the text accordingly.

We changed structural types numbers and placed them in descending order of aridity

L248. This sentence is not clear “we can observe how T3 was the structural type most widely distributed across our study zones”. It looks like a discussion/conclusion sentence. Here you must describe this fact first so the reader can understand ‘how’ T3 class was the most widely distributed.

L268-L269: Added to the ms (attending to the following commentary):

“Considering all provinces, T1 was the structural type most widely distributed across our study zones, i.e. the structural group that presented the highest abundance of all those evaluated in this work.”

L250. Confusing. What “a relative distribution” would be?

This sentence refers to the abundance of each group within each province. We have rephrased the sentence as follows L (270-271 ): “Considering all provinces, T1 was the structural type most widely distributed across our study zones, i.e. the structural group that presented the highest abundance of all those evaluated in this work.”

See supplementary material, Table A1, for more information.

L256. Please, use intuitive names to refer to these metrics (e.g.: CC2m for canopy cover above 2m, and not ‘A2/FR’). Also, give more details about how their definition (because it is just six, you could even provide their formulas).

We agree with the reviewer. Thus, we have added the following information to the manuscript (L199-203):

The six variables used in the second RF were CCM2, (All return above 2 meters/Total first returns) *100; MCHP represents the relative vertical distribution of canopy surface area (vertical vegetation profile), and accounts for occlusion of the laser energy by the canopy [80]; Area of the Curve, related with the forest stand biomass; P05 and P95 represent height where the 5% and 95%, respectively, are below of it; Dif 90_50, shows the difference between P90 and P50 and it is related with the biomass location.

L262. “Producer’s accuracy”?? Shouldn’t be just “the accuracy”? See also “user’s accuracy” in (L264) and tables 4 and 5. What are user and producer accuracy? You did not mention it in the methods.

To facilitate the understanding of the manuscript and clarify up the concepts of user’s and producer’s accuracy, we have added the following lines (L 196 – 198):

“Producer’s accuracy (how often are real features on the ground correctly shown on the classified map or the probability that a certain land cover of an area on the ground is classified as such) and user’s accuracy (how often the class on the map will be present on the ground, i.e., reliability)[79] were calculated using the confusion matrix.”

L264. Why are you calling Figure 3 here? Was it built using predicted classes? Please clarify this.

Rectified

L274. Which “both cases” are you referring to?

Cases refers to the RF with or without the downsampling process. We have added it to the ms (L301-302):

“The performance of the RF with or without the downsampling process (balanced and unbalanced data) was similar in both cases, i.e. quite similar results in analysis with balanced and unbalanced data.”

L275. Where are Figure A1 and table A1 and A2?

These tables and figures are in Supplementary material. We think that supplementary material was not properly attached.

L318. Seems that Gorgens is also using ref 87 and 88 for this statement, but it was part of his conclusions. Please modify this reference to something like: “(see also [87] and [88])”

Done and added to the ms (L346-L348):

“Görgens et al. [89] pointed out that metrics selected using automated processes might differ completely between studies and thus require the identification of stable metrics to be used as predictors to facilitate model generalization (see also [90] and [91]).”

417. Here you are referring to your own work, but then you call four other references, which seems to me that they were who made this choice for you. Please correct it.

Corrected in the ms.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors did not address my major concern of point density evaluation. They stated that "T(t)he main contribution is to provide spatially explicit information on forest structural types to forest managers." While SNFI does not have spatially explicit maps of forest structure, using lidar to map forests have been conducted for quite a while. I do not see novel 'scientific contribution' from this study by simply using lidar metrics and random forest.

Reviewer 3 Report

Dear editor,

I first thank the authors for their clarifying answers and for had improved the text. I recommended the acceptance. I just do not understand why the supplementary materials are not available for us reviewers in the journal system, for the second time. Although I shouldn’t, I will assume these materials are OK. I please ask the authors to double check them very carefully before submitting for publication.

There are also minor comments. I would like to see some discussion about how the prediction accuracy could be improved. Maybe they could be enhanced if satellite image metrics were used, for instance.

Other specific comments are given bellow.

L157. Could not be extracted from LiDAR or from FUSION/LDV ?

L211-215. Please reshape this sentence. Maybe braking it in two would let it more straightforward.

L219. Canopy curve or canopy surface?

L257. Figure 3 is totally confuse in the reviewing pdf. Please double check it.