Integrating LiDAR, Multispectral and SAR Data to Estimate and Map Canopy Height in Tropical Forests

Round 1

Reviewer 1 Report

In general, I found the paper to raise an important research topic, showing the relevance of integrating satellite remote sensing (multispectral and Radar) and LiDAR data for a better understanding of complex tropical forests. The objective of the study is supported by the method used, and the results presented and discussed. The combination of different remote sensing data sources, the different statistical algorithms tested, and predictions made could provide an exemplary approach for other forest monitoring and conservation studies. Despite its merits, I have a few comments that could help improve the manuscript.

The manuscript can benefit from the inclusion of recent papers in the field, especially in the introduction and discussion section. Here are a few suggestions:e.g. https://www.nature.com/articles/s41467-019-12737-x#citeas https://www.mdpi.com/1999-4907/10/3/291 Since the manuscript starts with the concept of EBVs, it would be interesting to make a statement in the discussion on how the findings of this research relate to different EBV classes as well Even though the GEDI mission has been mentioned on line 97, there has been no explanation given about it or regarding its relevance in the mapping of tropical forest structure. A statement can be made in the LiDAR section of the introduction (L62-73)   Figure 1 could be presented clearer, the study areas are bounded as a polygon but are presented/appear as a point on the map, the size of the maps is also too big in section 2.2 information regarding the total area coverage of the airborne LIDAR scans is missing figure 2 (b) appears rather complicated, could be simplified L179: citation for GEE

Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment.

A nice table could be made to summarize the information on data, regression methods used, indices etc the details on the number of training vs. validation sites should be made clearer Wouldn't it be also important to use the Adjusted R2 instead of using the R2 to compare the 5 regression models, as the predictors used across the models vary in number? Table 1/3: what does the bold and the underlined values represent? not mentioned in the caption  more discussion could be made on how these findings would relate to future missions, especially GEDI

Author Response

Reviewer #1

 

In general, I found the paper to raise an important research topic, showing the relevance of integrating satellite remote sensing (multispectral and Radar) and LiDAR data for a better understanding of complex tropical forests. The objective of the study is supported by the method used, and the results presented and discussed. The combination of different remote sensing data sources, the different statistical algorithms tested, and predictions made could provide an exemplary approach for other forest monitoring and conservation studies. Despite its merits, I have a few comments that could help improve the manuscript.

1) Citation suggested by reviewer: Bae et al. (2019).

R=The paper of Bae et al. (2019 ) on the use Sentinel-1 and airborne laser scanning (ALS) to study the forest structure was included in the citations of the introduction at the next lines were added 94-95: However, high correlations between the C-band backscatter and forest structural variables were found lately in temperate forests[48].

 

2) Since the manuscript starts with the concept of EBVs, it would be interesting to make a statement in the discussion on how the findings of this research relate to different EBV classes as well Even though the GEDI mission has been mentioned on line 97, there has been no explanation given about it or regarding its relevance in the mapping of tropical forest structure

R= We have divided the first paragraph of the discussion in two paragraphs to include some lines about the use of our methods and results in the spatial estimation of EBvs. Thus, the next discussion was added at the lines 419-424: Because we used free and open resources (imagery and software), this improvement of the CH mapping can be adopted to facilitate the accurate long-term global monitoring of this EBV. Annual metrics of multispectral and SAR predictors also could be examined to model spatially other EBVs (such as the other ecosystem structure variables) using discrete or waveform LiDAR sensors as ground reference (including GEDI that is available from 2019) because both LiDAR technologies generate accurate measures of vegetation structure.

 

3) Figure 1 could be presented clearer, the study areas are bounded as a polygon but are presented/appear as a point on the map, the size of the maps is also too big in section 2.2 information regarding the total area coverage of the airborne LIDAR scans is missing figure 2 (b) appears rather complicated, could be simplified L179: citation for GEE.

R= In relation to the figure 1. We have modified the figure 1 to improve the understanding of our methods. Thus, we have included the areas mapped (MAT, TAP, and CHOCO) in a detailed spatial resolution. We also included the LiDAR coverage in each study area.

In relation to the figure 2.  Although this figure was done with real data and shows a detailed area of TAP, the figure 2 is only an illustration to facilitate the understanding of our methods. We believe that the new information provided in the figure 1 allows the better understanding of the figure 2. Therefore, we did not modify the figure 2.

 

4) Citation suggested by reviewer: Gorelick et al. (2017).

R= We have included this citation on the use of Google Earth Engine in the line 180

 

5)  A nice table could be made to summarize the information on data, regression methods used, indices etc the details on the number of training vs. validation sites should be made clearer.

R= We have included a new table, the Table 1. Extent of study areas, LiDAR coverage, and training sites for the regression models.

 

6) Wouldn't it be also important to use the Adjusted R2 instead of using the R2 to compare the 5 regression models, as the predictors used across the models vary in number?

R= This is a valuable correction of the reviewer. We have estimated the R2 adjusted for each model and changed the results agree with the estimation of the adjusted R2. We also made a new graphic using the adjusted R2 (figure 4B). The results using the Adjusted R2 are very close the previous results using R2; thus, the paper was not affected by these changes.

7) Table 1/3: what does the bold and the underlined values represent? not mentioned in the caption

R= We used bold and underlined to highlight significant values in the Table 2. As well, we used bold and underlined to highlight significant values with r >0.2 or k>0.2 in the Tables 3 and 4. We have added this information to the caption of each table.

 

Reviewer 2 Report

First of all, congratulations for the work. The introduction is very consistent and didactic, the methodology is robust and the achieved results are highly relevant. This manuscript is a great and concrete contribution for the mapping of the canopy height in tropical forests.

Here is my suggestions:

1) Manuscript Presentation:
1.1. The figure 1 could be more informative, showing the actual LiDAR area sampled in MAT, TAP and CHOCO (see Fig. S1. of SI Materials
and Methods - Asner et al., 2010);
1.2. I believe these lines have typos: "...four study areas... (line 234)", "...in in..." (line 385);
1.3. The figure 4 could improve. I really can't distinguish the groups of variable selection in the charts.
1.4. Better not use the term "map accuracy", because in fact what was generated is the accuracy of the model, not the accuracy of the map (Line 359, 386). To generate a map accuracy, you will need representative samples for the mapped area (i.e. 100 times the coverage of the LiDAR surveys), preferably, completely isolated from the training samples.
1.5. Do you think this method could be used to estimate forest degradation ? What are the challenges to generalize this methodology to, for example, a WRS landsat scene (e.i. ~31.000 km2) ? Maybe the conclusions section could have a few words about it.

2) Technical Approaches:
2.1. What are the area and size of each LiDAR coverage for MAT, TAP and CHOCO ?
2.2. Why you used bilinear interpolation to resample 1m to 30m ? (line 156)

Congratulations again.
Great work !

Author Response

Reviewer #2

1) Manuscript Presentation:

1.1. The figure 1 could be more informative, showing the actual LiDAR area sampled in MAT, TAP and CHOCO (see Fig. S1. of SI Materials and Methods - Asner et al., 2010);

R= We have improved the information of the figure 1; we added three maps with the details of the study area (imagery of high spatial resolution in real color) and the LiDAR coverage. We also added a new table, the table 1, with the extent of study areas, the LiDAR coverage, and the training sites for the regression models.  

 

1.2. I believe these lines have typos: "...four study areas... (line 234)", "...in in..." (line 385);

R= We have corrected these two typos.

1.3. The figure 4 could improve. I really can't distinguish the groups of variable selection in the charts.

R= We have improved the visual resolution of the graphic and added the levels (V.1, V.2, and V.3) in the axis X to distingue the three groups of variable selection.

1.4. Better not use the term "map accuracy", because in fact what was generated is the accuracy of the model, not the accuracy of the map (Line 359, 386). To generate a map accuracy, you will need representative samples for the mapped area (i.e. 100 times the coverage of the LiDAR surveys), preferably, completely isolated from the training samples.

R= The reviewer provided a valuable correction. We have changed accuracy of the map to accuracy of the model at Line 359, 386 (now lines 383, 405).

1.5. Do you think this method could be used to estimate forest degradation? What are the challenges to generalize this methodology to, for example, a WRS landsat scene (e.i. ~31.000 km2) ? Maybe the conclusions section could have a few words about it.

R= The reviewer is right, and his comment is so valuable to improve our conclusions. Annual metrics of multispectral and SAR can help to study forest degradation and many other ecological phenomena in tropical forest. Thus, we added the lines 545-550 to the las part of the conclusions: Additionally, the study of some other forest traits could be improved by integrating annual metrics of multispectral and SAR bands, such as forest degradation, diversity, and phenology.

2) Technical Approaches:

2.1. What are the area and size of each LiDAR coverage for MAT, TAP and CHOCO?

R= We have included a table (Table 1) with information about the sizes of the study areas, the LiDAR coverage, and the quantity of the training and validation points.

2.2. Why you used bilinear interpolation to resample 1m to 30m? (line 156).

We needed to resample the CH to 30m in order to match this response variable to the spatial resolution of the predictors (multispectral and SAR). As we mentioned in the methodology, response and predictor variables were resampled to Landsat pixels. This multispectral sensor had the lowest spatial resolution and produced the highest number of predictors; thus, the resampling of the other sensors to the Landsat pixels was high appropriate in methodological terms. We selected bilinear interpolation among the most three usual methodologies (bilinear interpolation, Nearest neighbor, and Cubic convolution) because our variables are continuous and bilinear interpolation is recommended for continuous variables.

 

Reviewer 3 Report

This work presents a method for the CH (Canopy Height) calculation in three study cases in South America forests. The methods sounds interesting and its value as a tool for CH calculation of large areas is hight. 

After reading the article, I'm not sure if I understood the presented method. Authors have said in different parts of the article that the method integrates LiDAR data with SAR and multispectral data (Title, abstract, introduction...) but I think that this data is not integrated. In my opinion LiDAR data is used as the Ground Truth, and with the CH calculated with this information they calculate different models for the CH calculation just using SAR and multispectral information. If this is the case, I suggest that the authors should correct the article to make this information clear (change the title, abstract, intro, etc). 

Also, in the conclusions authors should discuss how the use of satellite data with pixel of 25-30m affects the results and how they could be improve using information with a higher spatial resolution. 

For future works, I suggest the authors to use UAV (Unmanned Aerial Vehicles) to obtain the multiespectral informaiton. A fixed wing platform can cover large areas in a single flight and can carry a multiespectral and thermal camera. UAV are not expensive and can be a usefull tool to obtain multiespectral data with higher resolution. 

Author Response

Reviewer #3

 

1) After reading the article, I'm not sure if I understood the presented method. Authors have said in different parts of the article that the method integrates LiDAR data with SAR and multispectral data (Title, abstract, introduction...) but I think that this data is not integrated. In my opinion LiDAR data is used as the Ground Truth, and with the CH calculated with this information they calculate different models for the CH calculation just using SAR and multispectral information. If this is the case, I suggest that the authors should correct the article to make this information clear (change the title, abstract, intro, etc).

 

The next is the recommendation of the assistant editor for this specific comment of reviewer 3.

Ms. Muriel Zhang: Reviewer 3 suggests your analysis is not really fusion, as there is no deep integration between LiDAR and the other multispectral and SAR datasets. While I understand where the reviewer is coming from, I believe there is enough integration in your method, with all SAR and ms datasets reliant on your LiDAR Canopy Height models. So I do not believe you need to change the title, nor the structure and emphasis of your paper. However, I would ask that you clarify some of the text in the Abstract and Methods to make it clearly what you do and don't do regarding 'integration', and potentially add a paragraph to the Introduction or Discussion looking into what you and others regard as 'data fusion' or 'data integration', and where your study fits in withthis narrative. I believe such edits would help address the concerns of Reviewer 3, and prevent any such confusion from other reviewers.

 

 

R= The reviewer is right: we used LiDAR-CH estimation as ground truth (as other authors have done previously) and then calculated different models for the CH calculation using SAR and multispectral information (annual metrics). Thus, we clarified the definition that we used for “integration” in the abstract and methods.

 

Lines 22-25 In the abstract: We modeled and mapped CH estimated from aircraft LiDAR surveys as ground reference, using annual metrics derived from multispectral and SAR satellite imagery in a dry forest, a moist forest, and a rainforest of tropical South America.

 

Lines 201-204 in the methodology: Our method for integrating discrete return LiDAR, Lansat-8, Sentinel-1, and PALSAR-2/PALSAR data was not a strict remote sensing fusion; we used the CHs estimated from LiDAR surveys as ground reference [see 21,46,47] to model CHs to areas with not LiDAR coverage using annual metrics of the multispectral and SAR as predictors.

 

2) Also, in the conclusions authors should discuss how the use of satellite data with pixel of 25-30m affects the results and how they could be improve using information with a higher spatial resolution.

 

R= The reviewer suggests a valuable improve to the conclusion; we have added the lines 550-555 to the last part of the conclusion. The measure of other forest traits could be also improved by integrating annual metrics of multispectral and SAR bands, such as forest degradation, diversity, and phenology. We have interpolated the respond and predictor variables to the spatial resolution of 30m, the lowest spatial resolution among these variables. Thus, the estimation of canopy height and other forest traits could be even improved more by integrating annual metrics of remote sensors with higher spatial, spectral, and temporal resolution, such as the multispectral sensor Sentinel-2.

 

 

Round 2

Reviewer 3 Report

In my opinion, after the revision made by the authors the paper is ready for publication. The misunderstanding about 'integration' is solved and the conclusion section has been improved according to my comments. 

 

Nice work, congratulations!.