Extending ALS-Based Mapping of Forest Attributes with Medium Resolution Satellite and Environmental Data

Round 1

Reviewer 1 Report

Extending ALS-based mapping of forest attributes with medium resolution satellite and environmental data

Overall this paper is well written and covers all the bases that I would expect from a piece of research in this scientific space. I have a few minor comments I would like to see addressed before publication.

L49. Add annotated citation followed by [14].

Paragraph from L72. Add brief review of other similar studies that extend to spaceborne LiDAR (i.e. GLAS)? IMO, a brief mention here is warranted for completeness. In some way these additional studies can be seen as a more valid approach than using optical data as GLAS is capable of capturing within canopy structural information.

L121. Explain 9 cm DBH cut-off here. It is explained below, but this is the first occurrence and should be explained up front.

Table 1 appears to be a result, not a method. Suggest moving to the appropriate section.

L150. LiDAR data was acquired ‘from 15 August’, until when? Was it acquired over multiple days? Please clarify.

L164. More info needed on ‘height, structural, density, and cover metrics’. Maybe a table of metrics, of just some info on height percentiles and how they were calculated, and similar info for structural, density and cover metric.

L165. ‘We calculated all height, structural, and density metrics using vegetated returns ≥2 m representing treed vegetation and ≤30 m to avoid erroneous points that exceeded that maximum tree height of the region.’ – Sentence needs a citation.

L174. Why remove these bands over the 'other' correlating bands? What was the qualifier for this?

L174. Explicitly state what 'r' is here (Pearson coefficient? Other?). Also why was 0.95 the cut-off? Seems arbitrary.

L178. Provide equivalent measure of 0.75 arcseconds in metres for clarity

L179. What are the ‘transformations’ in this context? Elaborate.

L186. Build à built (?)

L195. Need to justify the use of 500 trees. Seems overkill.

L200. What were the values of mtry tested? And which was best?

Figure 2. Need to explicitly define what the black and grey lines represent

Section 4.2. Discuss the option of utilising spaceborne LiDAR data for extension purposes. Not only GLAS (some citations from Nelson, 2009, Margolis 2015, Mahoney, 2018), but also considerations for future missions such as ISCESat-2.

Sections 4.2 & 4.3. Reads a lot like a summary rather than a discussion. Needs more implications of the work to be explored and articulated.

L464. What sort of areas would you focus your ‘further research’? There are a lot of papers that have employed both kinds of approaches (as you note in the introduction). We need a tangible direction to head in to keep this great research progressing. A little more information is appreciated.

Author Response

Response to Reviewer 1 Comments

Thank you very much for your attention to our manuscript.  We appreciate your very useful suggestions.

Point 1: L49. Add annotated citation followed by [14].

Response 1:  We modified the text as suggested.

Point 2: Paragraph from L72. Add brief review of other similar studies that extend to spaceborne LiDAR (i.e. GLAS)? IMO, a brief mention here is warranted for completeness. In some way these additional studies can be seen as a more valid approach than using optical data as GLAS is capable of capturing within canopy structural information.

Response 2: We extended this paragraph to include references to studies that use spaceborne LiDAR as follows:

“Others have combined field, airborne, and spaceborne LiDAR (i.e., Geoscience Laser Altimeter System (GLAS) data) to produce regional maps of stand attributes [44] and to estimate aboveground biomass and carbon of boreal forests using multilevel sampling strategies [45–50]. Although these studies have demonstrated improved mapping and estimation capabilities using indirect or multilevel approaches, further evaluation is warranted with different data sets, at different scales, and for different forest conditions.”

The following references were added to represent studies that have used spaceborne LiDAR in a multilevel approach: Boudreau et al RSE 2008; Nelson et al CJFR 2009; Margolis et al CJFR 205; Nelson et al RSE 2017; Holm et al RSE 2017; Mahoney et al RS 2018.

Point 3: L121. Explain 9 cm DBH cut-off here. It is explained below, but this is the first occurrence and should be explained up front.

Response 3: We modified the text as suggested.

Point 4: Table 1 appears to be a result, not a method. Suggest moving to the appropriate section.

Response 4: We prefer to leave the reference to Table 1 in the methods as the statistics of the ground plots describe the data employed in the study.  These statistics are often included in methods in this journal (e.g. Mauro et al. RS 2019; Saarella et al. RS 2018; Deo et al. RS 2017).  

Point 5: L150. LiDAR data was acquired ‘from 15 August’, until when? Was it acquired over multiple days? Please clarify.

Response 5: We modified the text for clarity as suggested. “Wall-to-wall ALS data were acquired between 15 August and 24 September 2016 ….”

Point 6: L164. More info needed on ‘height, structural, density, and cover metrics’. Maybe a table of metrics, of just some info on height percentiles and how they were calculated, and similar info for structural, density and cover metric.

Response 6: We provided more information on the metrics in the text as suggested as the table of metrics only included those metrics that we used for modeling. We also included citations for the various groups of metrics.

“Height metrics represented basic statistics of mean (MEAN) and max (MAX) height and the heights of various percentiles of first returns (i.e., P10…P90, P95) [6]. Structural metrics were statistical measures of skewness (SKEW), covariance (COVAR), vertical distribution ratio (VDR) [67] and vertical complexity index (VCI) [68]. To compute density metrics, we divided the range of LiDAR heights into 10 equal intervals and calculated the cumulative proportion of LiDAR returns found in the first nine intervals (i.e., D1 … D9) [6]. We calculated cover metrics at 2 m height intervals (i.e., CC2 … CC14) from the CHM as the number of 1 m × 1 m cells with a height value > 2 m divided by the number of nonvoid 1 m × 1 m cells [53]. Finally, we selected a reduced set of metrics for this study avoiding very highly correlated predictor variables within each group (Pearson correlation coefficient, r > 0.95) (Table 2).”

Point 7: L165. ‘We calculated all height, structural, and density metrics using vegetated returns ≥2 m representing treed vegetation and ≤30 m to avoid erroneous points that exceeded that maximum tree height of the region.’ – Sentence needs a citation.

Response 7: We added a citation as requested.

Point 8: L174. Why remove these bands over the 'other' correlating bands? What was the qualifier for this?

Response 8:  For bands that were highly correlated, we retained the bands that were least correlated with other bands in order to minimize redundancy.  We modified the text as follows
 “… bands 3, 7, 8A, and 12 were omitted to minimize redundancy with other bands …”

Point 9: L174. Explicitly state what 'r' is here (Pearson coefficient? Other?). Also why was 0.95 the cut-off? Seems arbitrary.

Response 9: We indicated that ‘r’ is the Pearson correlation coefficient as suggested.  We chose a very high cut-off to avoid redundancy in the predictor variables as is now indicated.

“… bands 3, 7, 8A, and 12 were omitted to minimize redundancy with other bands (Pearson correlation coefficient, r > 0.95).”

Point 10: L178. Provide equivalent measure of 0.75 arcseconds in metres for clarity

Response 10: We added the following information to clarify:

“… with a base resolution of 0.75 arc seconds (~23 meters) in the north–south direction …”

Point 11: L179. What are the ‘transformations’ in this context? Elaborate.

Response 11: We included the names of the transformations here along with the reference for their computation

“We generated solar radiation transformations including CosAspect, SCOSA, and SINA [72] by combining slope and aspect information derived from the CDEM.”

Point 12: L186. Build à built (?)

Response 12: We made this correction.

Point 13: L195. Need to justify the use of 500 trees. Seems overkill.

Response 13: The R package randomForest uses 500 trees by default. Our understanding is that the only limitation of having a large number of trees is computation time, which was not a problem in this context. We revised the text to indicate that 500 trees was the default.

Point 14: L200. What were the values of mtry tested? And which was best?

Response 14: We used tuneRF to determine the optimal number of predictor variables to retain. We added the following text to clarify our use of tuneRF:

“The tuneRF algorithm starts with a default value of mtry (i.e., the number of predictors divided by three) and searches for the optimal value according to out-of-bag error estimates. The optimal values of mtry varied depending on the attribute modeled.”

Point 15: Figure 2. Need to explicitly define what the black and grey lines represent

Response 15:  We added the following text to the figure caption:

“Black lines are 1:1 reference lines; gray lines are regression lines.”

Point 16: Section 4.2. Discuss the option of utilising spaceborne LiDAR data for extension purposes. Not only GLAS (some citations from Nelson, 2009, Margolis 2015, Mahoney, 2018), but also considerations for future missions such as ISCESat-2.

Response 16: Thank you for providing these citations. We added the following text to Section 4.4. Data and Technical Considerations:

“On the other hand, other types of remote sensing data may offer complementary information with the potential for improved results. For example, many researchers have shown that GLAS data are useful for extension purposes [44–50] and that metrics derived from Landsat time-series data have improved estimation of forest attributes beyond the use of single-date imagery [93,94]. Spaceborne LiDAR from the Advanced Topographic Laser Altimeter System (ATLAS) sensor on board ICESat-2 or future Global Ecosystem Dynamics Investigation (GEDI) mission could also facilitate estimation across large regions when used in combination with ground and airborne LiDAR data.”

Point 17: Sections 4.2 & 4.3. Reads a lot like a summary rather than a discussion. Needs more implications of the work to be explored and articulated.

Response 17: We added a new section entitled 4.5 Implications for Forest Inventory.

4.5. Implications for Forest Inventory

Forest inventories require information on a broad suite of forest attributes. In this study, we focused on key attributes commonly mapped with ALS that characterize the structure of vegetation. However, forest type and tree species information is a key information requirement of forest management [17]. Although substantive research has been conducted on species characterization with ALS data (reviewed by Vauhkonen et al .[95]), methods for mapping species have not yet reached the same level of maturity as those for mapping structural attributes such as height and volume. As a result, we did not address individual species. Rather, we stratified our plot database into coniferous, broadleaf, and mixed forest types as is common practice for area-based modeling, and we based our study on the coniferous forest strata. We did not produce models for broadleaf and mixed forest because of the limited ground sample data representing these forest types. Furthermore, we did not include the broadleaf and mixed forest plots in the models because preliminary analysis showed that doing so resulted in substantive underestimation of the high volume coniferous stands. Given that coniferous forest represents ~90% of the productive forest of this region and is the primary forest of industrial interest, we decided to sacrifice coverage for better estimation of attributes of the commercial forest. This limited application of the models to the area of coniferous forest and required a spatial layer representing forest type for mapping purposes, which we obtained from a conventional photo-based inventory of the area. Further work (and additional field sampling) is required to develop models for the broadleaf and mixed forest types.

The indirect approach used in this study does not preclude the requirement for ground plots. On the contrary, spatially precise and well-distributed ground plots are essential for building high quality ALS models. However, the indirect approach has the potential to optimize the efficiency of ground plot acquisition. For example, in this study, we used a priori ALS data to characterize the range of forest structural conditions across the study area prior to field sampling. There is general consensus that the use of a priori ALS data can maximize efficiency and reduce costs of ground plot acquisitions for area-based ALS modeling and mapping [58]. Recent research on the transferability of ALS-attribute models suggests potential for cost savings of some attributes by applying models to data with different point cloud characteristics or different areas [96]. Furthermore, in this study, we demonstrated that supplementing ground plots with ALS samples improved prediction over a direct modeling approach that uses the more limited set of ground plots. Other studies have also shown improved estimation of volume and biomass by combining field plots, ALS data and satellite data [41–43]. Supplementation of ground plots with ALS samples could significantly reduce inventory costs for remote and less accessible areas. Further research is needed in the design of ground and ALS sampling systems for multilevel mapping and estimation and to make inventories most cost-efficient (e.g.,[97]).

Finally, our study was carried out in the boreal forest conditions of western Newfoundland, Canada. The practical objective was to extend ALS-based mapping of key forest structural attributes from an area covered by wall-to-wall ALS data to an area of similar ecological conditions representing a full forest management district. Extending the mapping beyond these ecological conditions would result in predictions with unknown and likely increased errors. Additionally, the performance of the approach under different conditions (i.e., forest types, stand structures, difficult terrain) and with different datasets requires further research.”

Point 18: L464. What sort of areas would you focus your ‘further research’? There are a lot of papers that have employed both kinds of approaches (as you note in the introduction). We need a tangible direction to head in to keep this great research progressing. A little more information is appreciated.

Response 18: We incorporated areas for further research in Section 4.5 above.

Reviewer 2 Report

General: This article deals with the evaluation of a two-phased indirect approach for extending ALS-based maps forest maps for the prediction of forest attributes for the respective research area in research area in Western New Foundland.

The article is well written and well structured. Although the study talkes place in a relevant research area, the character of the manuscript is very much methodological and sometimes abstract. This is why detailed and structured explantions are of special importance. From a foresters’ perspective a bit more on forests and the meaning of the study in terms of practical implications could be discussed. I recommend minor revisions before considering the manuscript for publication.  

Introduction:

The introduction is well written and explains stepwise how recent research on ALS and satellite imagery for mapping moved forward. Some parts are written in the way of a literature review, with is suitable for making the reader familiar with the topic. Literature on the concept of combining ground plots, ALS and satellite imagery is mentioned.

Materials and Methods:

The research area seems to be suitable and sufficiently large for such a study.

Line 115- 128: The description of the ground plots may be described a bite more focused and structured. The authors jump there sometimes between ground plot and ALS and I think it would be better for reading to separately explain just the design and parameter of the ground plots. For example: line 122- 123 says that the respective attributes where also recorded for trees above 1.3 m in a separate circle with r= 4 m. So this circle was situated in the center or where? Potentially the authors may even add a small figure of the plots design: Plot design and attributes, distance between plots and also the representativity of each plot are of special importance since this was a calibration and validation basis.

Line: 142-143: … “four plots had structural characteristics exceeding the range of the calibration data…” What exactly do you mean? Maybe you can state this a bit more concrete?

Line 229-230: why samples with less than 75 % coniferous forest were removed?? Show would look the results if these parts were not removed? Is this a kind of limitation?  If yes, please state this!

Table 1: Maybe you can add in the captions or below the table the explanations for the used shortcuts (HGI, BA, GMV etc…).

Some parts of the methodology are very detailed and shortening may be considered.

Discussion:

Certain technical and methodological aspects are discussed. However, I think me and other readers may be interested to read a bit more on practical implications, concrete forest related / forest inventory related considerations. Maybe the authors can add a subchapter on “implications for forest inventory/management”? Are there any? To my knowledge ground plots still play big role in forest inventories and are basically indispensable. How do the authors feel about this? If ground plots would be only little necessary, why are they used worldwide and present the basis of any comprehensive inventory that includes forest structural attributes? Also: how such a method would perform under different conditions (stand structure, difficult terrain like high and steep mountains)? How do the authors see this or can compare from literature? Maybe I missed it , but how was the performance with regard to different tree species? Please touch these points. Even if this submission is for a remote sensing and not explicit forest journal: From a foresters’ point of view, it may also be interesting to have more practical comparison of real data / information on certain attributes with regard to the last forest inventory of the region and provide in some sentences some practical/visual examples on how useful the method of this study appears to be. In the foresters will be a main target group for the readership of this article, because it is about application in forest related context.

Finally, a few words about the literature: It appears to me that the literature overall covers the methodological aspects. However, the authors may keep in mind, that Remote Sensing is an international journal and so is the readership. Therefore, source beyond Canada and North America may be considered a bit more. How are experiences from other regions of the world (Asia, Europe etc.). I guess it would be interesting to add some comparing sources from other regions of the world (if available).

Author Response

Response to Reviewer 2 Comments

Thank you very much for your attention to our manuscript.  We appreciate your very useful suggestions.

Point 1: Line 115- 128: The description of the ground plots may be described a bite more focused and structured. The authors jump there sometimes between ground plot and ALS and I think it would be better for reading to separately explain just the design and parameter of the ground plots. For example: line 122- 123 says that the respective attributes where also recorded for trees above 1.3 m in a separate circle with r= 4 m. So this circle was situated in the center or where? Potentially the authors may even add a small figure of the plots design: Plot design and attributes, distance between plots and also the representativity of each plot are of special importance since this was a calibration and validation basis.

Response 1:  We made substantive changes to this section. We provided more details on the sampling design and clarified that that subplot was situated in the center of the larger plot. In lieu of a figure of the plot design, we provided a reference to Canada’s National Forest Inventory Ground Sampling Guidelines upon which the plot layout was based.  We hope this improves the structure and clarity of this. section The revised text for the beginning of this section is as follows:

“2.2. Ground Plots

We used two independent data sets of ground plots for calibrating and validating models of forest attributes. Both datasets were measured following ground sampling guidelines established for Canada’s National Forest Inventory [56]. The sample design aimed to capture the range of variability in forest structure of the study area. For calibration, we selected potential sample locations using ALS data (described in Section 2.3) as a basis for stratification following others who found improved accuracies using ALS-assisted plot selection for area-based inventory [57–59]. Instead of using specific ALS metrics (e.g., mean and standard deviation [58]), we performed a principal component analysis (PCA) [60] and represented forest structure by the first two components, which accounted for ~83% of the variance in ALS metrics. We used photo-interpreted forest inventory stand polygons to create stand-level dominant species masks and divided the range of values for each PCA component into 10 equal strata under each mask. We then selected a random sample location for ground-based assessment from each combination of species mask and PCA strata that existed within our study area. For validation, we selected potential sample locations using a stratified random sampling design using total volume predictions for conifer-dominated stands (result from Section 2.6) as the basis for stratification. We stratified the range of values of volume into 20 strata and randomly selected a minimum of two sample locations per strata.

 We established circular plots with 11.28 m radius and recorded species, DBH (measured at 1.3 m), height, and status (live or dead) for all merchantable trees (trees ≥ 9 cm DBH). We recorded these same attributes for all trees with a minimum height of 1.3 m for a subplot of radius 3.99 m located in the center of each plot. We recorded the center location of each plot with a Trimble GeoExplorer, 6000 series XH global positioning system (GPS) with floodlight technology and used base station data to differentially correct the roving receiver data to within submeter accuracy. In total, we measured 89 calibration plots in 2016 and 2017, and 43 validations plots in 2018.

From the standard measurements of species, height, and DBH, we derived a suite of structural attributes for all plots “

Point 2: Line: 142-143: … “four plots had structural characteristics exceeding the range of the calibration data…” What exactly do you mean? Maybe you can state this a bit more concrete?

Response 2:  We modified this text as follows

“Of the 43 validation plots, we removed four plots with GMV or TVOL that exceeded the range of the calibration data and were, therefore, not representative of the conditions modeled... “

Point 3: Line 229-230: why samples with less than 75 % coniferous forest were removed?? Show would look the results if these parts were not removed? Is this a kind of limitation?  If yes, please state this!

Response 3:  We added a rationale for modeling the coniferous forest in Section 2.2 and for removing samples with  ≤ 75% coniferous in Section 2.7. We also added a new section to the Discussion to explain the implications/limitation associated with modeling only the coniferous forest. 

“2.2. Ground Plots

… For the current study, we focused on the coniferous forest, which represents ~90% of the productive forest of our study area and is the forest type of commercial importance to the region.”

“2.7. Development of Extended Inventory (Phase 2)

“…we removed samples from areas with ≤75% coniferous forest according to a recent forest management inventory of the area as models only represented the coniferous forest. …”

“4.5. Implications for Forest Inventory

Forest inventories require information on a broad suite of forest attributes. In this study, we focused on key attributes commonly mapped with ALS that characterize the structure of vegetation. However, forest type and tree species information is a key information requirement of forest management [17]. Although substantive research has been conducted on species characterization with ALS data (reviewed by Vauhkonen et al. [95]), methods for mapping species have not yet reached the same level of maturity as those for mapping structural attributes such as height and volume. As a result, we did not address individual species. Rather, we stratified our plot database into coniferous, broadleaf, and mixed forest types as is common practice for area-based modeling, and we based our study on the coniferous forest strata. We did not produce models for broadleaf and mixed forest because of the limited ground sample data representing these forest types. Furthermore, we did not include the broadleaf and mixed forest plots in the models because preliminary analysis showed that doing so resulted in substantive underestimation of the high volume coniferous stands. Given that coniferous forest represents ~90% of the productive forest of this region and is the primary forest of industrial interest, we decided to sacrifice coverage for better estimation of attributes of the commercial forest. This limited application of the models to the area of coniferous forest and required a spatial layer representing forest type for mapping purposes, which we obtained from a conventional photo-based inventory of the area. Further work (and additional field sampling) is required to develop models for the broadleaf and mixed forest types.”

Point 4: Table 1: Maybe you can add in the captions or below the table the explanations for the used shortcuts (HGI, BA, GMV etc…).

Response 4:  Footnote added to Table 1 as requested.

Point 5: Some parts of the methodology are very detailed and shortening may be considered.

Response 5: We tried to present the methods as concise as possible. We considered reducing the text but found it difficult to do so without sacrificing important details. If there are specific areas which could be reduced, we are willing to reassess.

Point 6: Discussion: Certain technical and methodological aspects are discussed. However, I think me and other readers may be interested to read a bit more on practical implications, concrete forest related / forest inventory related considerations. Maybe the authors can add a subchapter on “implications for forest inventory/management”? Are there any? To my knowledge ground plots still play big role in forest inventories and are basically indispensable. How do the authors feel about this? If ground plots would be only little necessary, why are they used worldwide and present the basis of any comprehensive inventory that includes forest structural attributes? Also: how such a method would perform under different conditions (stand structure, difficult terrain like high and steep mountains)? How do the authors see this or can compare from literature? Maybe I missed it , but how was the performance with regard to different tree species? Please touch these points. Even if this submission is for a remote sensing and not explicit forest journal: From a foresters’ point of view, it may also be interesting to have more practical comparison of real data / information on certain attributes with regard to the last forest inventory of the region and provide in some sentences some practical/visual examples on how useful the method of this study appears to be. In the foresters will be a main target group for the readership of this article, because it is about application in forest related context.

Response 6:  As suggested, we added a section 4.5 Implications for Forest Inventory.  We addressed the topic of species, the requirement for ground plots, and the applicability to different conditions.  We did not compare our results with the last forest inventory of the region due to the temporal discrepancy between our data and the last photo-acquisition.

“4.5. Implications for Forest Inventory

Forest inventories require information on a broad suite of forest attributes. In this study, we focused on key attributes commonly mapped with ALS that characterize the structure of vegetation. However, forest type and tree species information is a key information requirement of forest management [17]. Although substantive research has been conducted on species characterization with ALS data (reviewed by Vauhkonen et al. [95]), methods for mapping species have not yet reached the same level of maturity as those for mapping structural attributes such as height and volume. As a result, we did not address individual species. Rather, we stratified our plot database into coniferous, broadleaf, and mixed forest types as is common practice for area-based modeling, and we based our study on the coniferous forest strata. We did not produce models for broadleaf and mixed forest because of the limited ground sample data representing these forest types. Furthermore, we did not include the broadleaf and mixed forest plots in the models because preliminary analysis showed that doing so resulted in substantive underestimation of the high volume coniferous stands. Given that coniferous forest represents ~90% of the productive forest of this region and is the primary forest of industrial interest, we decided to sacrifice coverage for better estimation of attributes of the commercial forest. This limited application of the models to the area of coniferous forest and required a spatial layer representing forest type for mapping purposes, which we obtained from a conventional photo-based inventory of the area. Further work (and additional field sampling) is required to develop models for the broadleaf and mixed forest types.

The indirect approach used in this study does not preclude the requirement for ground plots. On the contrary, spatially precise and well-distributed ground plots are essential for building high quality ALS models. However, the indirect approach has the potential to optimize the efficiency of ground plot acquisition. For example, in this study, we used a priori ALS data to characterize the range of forest structural conditions across the study area prior to field sampling. There is general consensus that the use of a priori ALS data can maximize efficiency and reduce costs of ground plot acquisitions for area-based ALS modeling and mapping [59]. Recent research on the transferability of ALS-attribute models suggests potential for cost savings of some attributes by applying models to data with different point cloud characteristics or different areas [96]. Furthermore, in this study, we demonstrated that supplementing ground plots with ALS samples improved prediction over a direct modeling approach that uses the more limited set of ground plots. Other studies have also shown improved estimation of volume and biomass by combining field plots, ALS data and satellite data [41–43]. Supplementation of ground plots with ALS samples could significantly reduce inventory costs for remote and less accessible areas. Further research is needed in the design of ground and ALS sampling systems for multilevel mapping and estimation and to make inventories most cost-efficient (e.g.,[97]).

Finally, our study was carried out in the boreal forest conditions of western Newfoundland, Canada. The practical objective was to extend ALS-based mapping of key forest structural attributes from an area covered by wall-to-wall ALS data to an area of similar ecological conditions representing a full forest management district. Extending the mapping beyond these ecological conditions would result in predictions with unknown and likely increased errors. Additionally, the performance of the approach under different conditions (i.e., forest types, stand structures, difficult terrain) and with different datasets requires further research.”

Point 7: Finally, a few words about the literature: It appears to me that the literature overall covers the methodological aspects. However, the authors may keep in mind, that Remote Sensing is an international journal and so is the readership. Therefore, source beyond Canada and North America may be considered a bit more. How are experiences from other regions of the world (Asia, Europe etc.). I guess it would be interesting to add some comparing sources from other regions of the world (if available).

Response 7:  We added the following references to studies that combining ground plots, LiDAR and satellite data for mapping or estimating forest attributes beyond Canada and North America. These include: Neigh et al RSE 2013; Saarela et al. RSE 2015, Holm et al. RSE 2017; Saarela et al RS 2018.

Author Response File: Author Response.docx

Reviewer 3 Report

I have my comments attached in the paper. Plz revise accordingly.

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 3 Comments

Thank you very much for your attention to our manuscript.  We appreciate your very useful suggestions.

Point 1:  Minor text edits throughout the manuscript.

Response 1:  We addressed the minor text edits according to the suggested changes.

Point 2: also check these useful references:
 (1) https://www.fs.fed.us/nrs/pubs/jrnl/2017/nrs_2017_deo_001.pdf
(2) https://www.fs.fed.us/nrs/pubs/jrnl/2017/nrs_2017_deo_002.pdf
LiDAR strip samples are used in a regional study in https://www.fs.fed.us/nrs/pubs/jrnl/2017/nrs_2017_deo_002.pdf

Response 2:  Thank you for these suggestions. We incorporated these citations in the paragraph that introduces the combined use of ALS data and satellite imagery and we added the following text:  

“Not surprisingly, models combining ALS with satellite variables generally provide better estimates than models using satellite imagery alone [32,33]…”

Point 3:  why first two principle components? add a citation

Response 3:  We added a citation for the principal component analysis and explain the selection of the first two principal components as follows: 

“… Instead of using specific ALS metrics (e.g., mean and standard deviation [58]), we performed a principal component analysis (PCA) [60] and represented forest structure by the first two components, which accounted for ~83% of the variance in ALS metrics.”

Point 4: dead standing trees also have volume and biomass and are captured in LiDAR point cloud. Give a reason why you ignored dead trees

Response 4: We added the following explanation for calculating attributes with live trees:

“We calculated attributes based on live trees (i.e., growing stock) for consistency with volume tables used by provincial and industrial agencies and because live-tree attributes are of interest for most operational and planning applications in the region.”

Point 5: If your focus is on large area mapping and estimation then why you did not consider pooled data (combined deciduous and conifer plots) for modeling? Even if you obtain good model with conifer plots, you can not generalize over the entire area, for example in mixed and broadleaved forests

Response 5:  We provide a rationale for modeling only the coniferous forest in Section 2.2. In Section 2.7, we added that “… models were developed only representing the coniferous forest”. We also added a new section to the Discussion to explain the implications/limitation associated with modeling only the coniferous forest. 

“2.2. Ground Plots

… For the current study, we focused on the coniferous forest, which represents ~90% of the productive forest of our study area and is the forest type of commercial importance to the region.”

“2.7. Development of Extended Inventory (Phase 2)

“…we removed samples from areas with ≤75% coniferous forest according to a recent forest management inventory of the area as models only represented the coniferous forest. …”

“4.5. Implications for Forest Inventory

Forest inventories require information on a broad suite of forest attributes. In this study, we focused on key attributes commonly mapped with ALS that characterize the structure of vegetation. However, forest type and tree species information is a key information requirement of forest management [17]. Although substantive research has been conducted on species characterization with ALS data (reviewed by Vauhkonen et al. [95]), methods for mapping species have not yet reached the same level of maturity as those for mapping structural attributes such as height and volume. As a result, we did not address individual species. Rather, we stratified our plot database into coniferous, broadleaf, and mixed forest types as is common practice for area-based modeling, and we based our study on the coniferous forest strata. We did not produce models for broadleaf and mixed forest because of the limited ground sample data representing these forest types. Furthermore, we did not include the broadleaf and mixed forest plots in the models because preliminary analysis showed that doing so resulted in substantive underestimation of the high volume coniferous stands. Given that coniferous forest represents ~90% of the productive forest of this region and is the primary forest of industrial interest, we decided to sacrifice coverage for better estimation of attributes of the commercial forest. This limited application of the models to the area of coniferous forest and required a spatial layer representing forest type for mapping purposes, which we obtained from a conventional photo-based inventory of the area. Further work (and additional field sampling) is required to develop models for the broadleaf and mixed forest types.”

Point 6: what did you do to remove spikes in DTM and CHM?

Response 6:  We did not observe any spikes in our DTM or CHM.

Point 7: did you do any atmospheric correction? were these surface reflectance? how did you deal with cloud and shadow?

Response 7:  We did not do any atmospheric correction. All four images were acquired on the same day under virtually clear skies.  There were a couple of very small pockets of cloud and cloud shadow, which we filtered as explained in section 2.7.

“We also removed potential samples located in areas of cloud or cloud shadow and areas harvested on the satellite imagery since the time of ALS acquisition. To do this, we established thresholds for band 2 (> 0.7 & < 0.09), 3 (> 0.6 & < 0.11) and 4 (> 0.03 & < 0.08) by visually assessing the imagery.”

Point 8: units for SKEW and COVAR?

Response 8:  These metrics are unitless.

Point 9: why did you consider only three density metrics?

Response 9:  We considered nine density metrics, but Table 2 only describes the reduced set of metrics following removal of those metrics that were considered to be redundant (r > 0.95).  We improved the explanation of the metrics in section 2.3 and included a citation as follows:

“To compute density metrics, we divided the range of LiDAR heights into 10 equal intervals and calculated the cumulative proportion of LiDAR returns found in the first nine intervals (i.e., D1 … D9) [6].”

Point 10: why did you consider only three CHM metrics? Not clear; are these raster grids?number of CHM cells within ???

Response 10:  We considered seven CHM metrics, but Table 2 only describes the reduced set of metrics following removal of those metrics that were considered to be redundant (r > 0.95).  We improved the explanation of the metrics in section 2.3 and included a citation as follows:

“We calculated cover metrics at 2 m height intervals (i.e., CC2 … CC14) from the CHM as the number of 1 m × 1 m cells with a height value > 2 m divided by the number of nonvoid 1 m × 1 m cells [53].”

Point 11: check units for Sentinel 2 variables

Response 11:  The Sentinel 2 pixel radiometric measurements are Level1C Top of Atmosphere (TOA) reflectances. Values range from 0 to 1, which we interpret as % expressed in decimal form.

Point 12: how that works? (i.e. tuneRF)

Response 12:  We added the following explanation for tuneRF

“The tuneRF algorithm starts with a default value of mtry (i.e., the number of predictors divided by three) and searches for the optimal value according to out-of-bag error estimates. The optimal values of mtry varied depending on the attribute modeled.”

Point 13: how were potential samples filtered

Response 13:  We added the following explanation of how we filtered potential samples.

“We converted the raster cells to a points shape file using ESRI ArcGIS [86] and we input the points shape file in R [76], where we filtered the potential samples to provide calibration data that were representative of the relevant forest conditions as follows …”