Combining Hyperspectral, LiDAR, and Forestry Data to Characterize Riparian Forests along Age and Hydrological Gradients

Round 1

Reviewer 1 Report

Suggestion on line 13 (Abstract):

Abstract: Riparian forests are complex ecosystems shaped by their connectivity in a river system,.....

Author Response

Thank you for taking the time to review our paper.

 

Point 1. 

Suggestion on line 13 (Abstract): Abstract: Riparian forests are complex ecosystems shaped by their connectivity in a river system, … 

 

Reply 1. 

We wish to keep the sentence as “connectivity to a river system” rather than modify it because this grammatical construction is the one that fits our intended meaning the best, even though riparian forests are a part of that river system.

Reviewer 2 Report

Dear authors,

Thanks for your submission. The comments are as follows:

1. The introduction is too long, and what's the scientific question in this manuscript?

2.  The thematic map of fig. 1 is not conform to cartography specifications. Please double check.

3.  What's the innovation in the methods? 

4. How to validate the extraction accuracy? Ground observation or public dataset?

5. From line 1023 to end, please double check all mistake in references.

Author Response

Thank you for taking the time to review our paper, and for your suggestions.

 

Point 1.

The introduction is too long, and what’s the scientific question in this manuscript?

Reply 1. 

Following your comment, we shortened the introduction by condensing the text and removing some parts that were less important for understanding the manuscript. This resulted in a reduction of about 15%. As we believe this manuscript could appeal to people with differing levels of familiarity with remote sensing or riparian forests, we tried to make sure we still covered both of those elements sufficiently. 

To answer your comment about the scientific question of the manuscript, we made changes throughout the text. First, we reworked the end of the introduction in order to highlight our main objectives (lines 132 to 138). We then referred back to those objectives by their numbers in the discussion section of the manuscript (lines 841 and 842). 

 

Point 2.

The thematic map of Fig. 1 is not conform to cartography specifications. Please double check.

Reply 2. 

We made slight adjustments to Fig. 1 in response to this comment. We made sure to use more intense and contrasting colors to highlight the coverage of our remote sensing data. We moved the part of the legend referring to the map of France just below it in order to make sure to distinguish it from the one for the study site. We added both rivers to the legend of the map, in addition to keeping their names in italics next to each river, as is traditional for location maps. We also tried to highlight the fact this is supposed to be a reference map rather than a thematic map by adding a satellite layer as background. If further adjustments need to be made, please specify because we are unsure where the issue was for a reference map, even after double checking.

 

Point 3.

What’s the innovation in the methods?

Reply 3 - Answer

As highlighted in the introduction, the combined use of hyperspectral, LiDAR, and field data to study riparian forest is novel because most studies in the field did not exploit this kind of dataset, tending instead towards data with more limited spectral resolution.

In addition, the application presented in this paper is very innovative in the field of geomorphology because we can study and demonstrate the simultaneous impacts of forest aging and geomorphic changes on the health of riparian forests, by integrating those two gradients in our study and in the methodology we used to explore our dataset.

 

Point 4.

How to validate the extraction accuracy? Ground observation or public dataset?

Reply 4 - Answer.

By “extraction accuracy” are you referring to extracting remote sensing indexes as described in 4.1?

This was done using classical GIS techniques to extract them for the location of our ground observations. Those ground observations (forest plots) had their planar accuracy assessed in the field with a DGPS in the case of the EVS survey. In the case of the ONF, the methodology was similar but using a GPS survey.

For the remote sensing data, their planar accuracy was checked using publicly available data produced by IGN which includes the series of aerial photographs that we used to determine plot age. In the case of the LiDAR data, vertical accuracy was checked by comparing the mean height in the EVS plots with ground observations of height classes of all the trees in the plot. 

Therefore, we mostly used ground observations. All the indexes derived from LiDAR and hyperspectral data were extracted for those ground observations. To make this clearer to the reader, we added a few words at the beginning of 4.1 (line 374).

 

Point 5.

From line 1023 to end, please double check all mistake in references.

Reply 5.

We double-checked all of the references after editing the manuscript and incorporating the changes to the introduction.

Reviewer 3 Report

This is a well-written very comprehensive manuscript, with clear objectives, and well described dataset.

One item that needs work is the definition of a mature forest.  In places this seems to be determined by age, in others “…most mature stage of growth (developed understory and high tree diameter)” –it is not clear what “high tree diameter” means, a large tree diameter?  And in others mature seemed focused on height of canopy, and in one place the text indicates “…the most mature stages of the poplar forest where it transitions towards post-pioneer hardwood forest” which seems to indicate that mature forests depend on successional phase.  This is important because recent policy requests in other nations are focused on definitions of “mature” for policy purposes.  Please clarify.

The second item that needs further consideration is the conclusion about L892-896, …the possibility of using reflectance data alone for mapping the connectively of riparian forests…” and included in the abstract.  This study relied heavily on analyses of forestry data which was then used with Lidar, and some hyperspectral.  There was a short discussion of what could be perhaps be done if one used on the hyperspectral, but there was not a direct comparison of what could be done with just hyperspectral versus all that was learned and discussed with the remote sensing, lidar, and forestry data.

It should not take long to deal with the above items.  Modifying text should be enough to deal with these

Minor items:

See line 205, fill in for X’s in XXth century

 In terms of possible relevant citations of work involving Lidar and hyperspectral, there is a large study/inventory in interior Alaska that may be of interest.  One citation is:

 

Shoot, Caileigh; Andersen, Hans-Erik and others. 2021. Classifying forest type in the National Forest Inventory context with airborne hyperspectral and lidar data. Remote Sensing. 13(10): 1863. https://doi.org/10.3390/rs13101863.

Author Response

Thank you for taking the time to review our paper, and for your suggestions.

 

Point 1.

One item that needs work is the definition of a mature forest. In places this seems to be determined by age, in others “... most mature stage of growth (developed understory and high tree diameter)” -it is not clear what high tree diameter means, a large tree diameter? And in others mature seems focused on height of canopy, and in one place the text indicates “... the most mature stages of the poplar forest where it transitions towards post-pioneer hardwood forest” which seems to indicate that mature forests depend on successional phases. This is important because recent policy requests in other nations are focused on definitions of “mature” for policy purposes. Please clarify.

Reply 1.

We made slight adjustments to the text in the Study Site, Materials, and Methods sections in order to clarify how we used “mature forest” in the rest of the manuscript. For the study site section, we adjusted the legend of Figure 2. For the Materials section, we added some information on our ground observations (lines 317 to 320). For the Methods section, we added information when presenting the classification targets that highlights how this was determined for practical purposes (lines 437 to 442). In addition, in the quote you highlighted we changed “high” to “large” since this was the intended meaning (line 754).

Our definition of mature forest is based on the successional stages identified in the field by the French Forestry Office, and on the interests of local stakeholders. The forest plots we consider as “growing” (as opposed to mature in the manuscript) are the ones that are of key interest for them because they are the first successional stages and tend to feature mostly pioneer species. For practical purposes with our dataset, we therefore used age, tree height, and the proportion of pioneer species in the plot to see at which age the shift between those two environments was occurring, and we used the corresponding age when training the classifier (Methods section).  

 

Point 2.

The second item that needs further consideration is the conclusion about L892-896, “...the possibility of using reflectance data alone for mapping the connectivity of riparian forests” and included in the abstract. There was a short discussion of what could be done if one used only the hyperspectral, but there was not a direct comparison of what could be done with just hyperspectral versus all that was learned and discussed with the remote sensing, lidar, and forestry data.

Reply 2.

As you pointed out, there was a short discussion of what could be done with only hyperspectral data, and Table 6. presents a direct comparison of the classification errors obtained by using hyperspectral data alone, versus LiDAR alone and LiDAR + hyperspectral, but there was no such comparison for the first part of our results in 5.1, which was critical to characterizing the riparian forest and understanding the processes occurring. Therefore, we have changed the text accordingly to make this point clearer (lines 925 to 929). 

 

Point 3.

See line 205, fill in for X’s in the Xxth century.

Reply 3.

This sentence used roman numerals for the 20th century. Therefore, we applied the proposed change by using the decimal numeral system in order to clarify the text (line 195). We also applied this change in other portions of the manuscript where roman numerals were used (line 166).

 

Point 4.

In terms of possible relevant citations of work involving Lidar and hyperspectral, there is a large study/inventory in interior Alaska that may be of interest. One citation is

Shoot, Caileigh; Andersen Hans-Erik and others. 2021 Classifying forest type in the National Forest Inventory context with airborne hyperspectral and lidar data. Remote Sensing. 13(10): 1863. https://doi.org/10.3390/rs13101863 

Reply 4.

Thank you for the reference. We added it to the introduction (line 107), but due to demands towards shortening the introduction we did not give much details.

Reviewer 4 Report

Dear authors

The manuscript with the title “Combining hyperspectral, LiDAR, and forestry data to characterize riparian forests along age and hydrological gradients” aimed to assess the ability of remote sensing to characterize and monitor riparian forests. Riparian forests are a strategic resource, because they provide us water quality and regulation, habitat quality, among others. In this sense, this type of studies are needed to monitor the influence of anthropogenic pressure.

This manuscript is well organized. There are a few comments and suggestions about this article that should be addressed:

Introduction: The authors explained the current state of knowledge. The knowledge gap was identified. Outline objectives were addressed. However, I would suggest summarizing this section, because is quite long.

Material and Methods

-        Study area was well described

-        3.1. Remote Sensing Information: I would suggest a nice table with remote sensing information used in each sub section.

-        Why do not used a multispectral imagery such as Pleiades or World-View or Sentinel-2?

-        What type of mechanism do you used to prevent overfitting of Random forest classifier?

Results

Why do not calculate error matrix? I would suggest presenting the error matrix. Commission and omission errors are very import descriptors of your classifications.

 

Discussions: The authors discuss the results obtained in this section, but the results are not in perspective with other similar papers. The authors explained the significance of results and how they contribute to advance knowledge.

Conclusions: The conclusions are consistent with the evidence and arguments presented.

 

Author Response

Thank you for taking the time to review our paper, and for your suggestions.

 

Point 1.

Introduction: [...]. However, I would suggest summarizing this section because it is quite long.

Reply 1.

This suggestion is similar to one given by another reviewer, and we shortened the introduction as a result (by around 15%), while trying to keep it accessible for readers with differing degrees of familiarity with either remote sensing or riparian forests. 

 

Point 2.

3.1 Remote Sensing Information: I would suggest a nice table with remote sensing information used in each sub section.

Reply 2.

We added such a table and a few lines of text right before the description of each individual dataset (lines 218 to 222). 

 

Point 3.

Why do not used a multispectral imagery such as Pleiades or World-View or Sentinel-2?

Reply 3.

The main objective of this paper was to focus on a one-shot study to explore how coupling LiDAR and hyperspectral data could help characterize riparian forests. 

From this study, we currently hypothesize that the physical disconnection of the riparian forest in the incised reach could lead to tree water stress. Therefore, we want to first explore this hypothesis in order to cross-validate our findings and to better understand the potential temporal dynamics before attempting to map and monitor forest connectivity from space by using a time-series of multispectral data. Therefore, a time-series analysis of multispectral data is outside the scope of this work for now, and using a single image would have required a timed acquisition with the hyperspectral data in order to account for external parameters such as differences in temperatures and precipitations. 

 

Point 4.

What type of mechanism do you used to prevent overfitting of Random Forest classifier?

Reply 4.

The performance of the classifier with the validation dataset (see the reply to Point 5) is comparable to those of the calibration dataset, which suggests the classifier does not overfit. In order to ensure the classifier was able to be generalized, we selected a balanced training data (50 in each of the two classes) otherwise there would have been a strong bias towards the most abundant class.  

 

Point 5.

Why do not calculate error matrix? I would suggest presenting the error matrix. Commission and omission errors are very import descriptors of your classifications.

Reply 5.

In the original manuscript, there was a confusion matrix for the training data (Table 7). For each classifier, the training data is a balanced set of 50 vegetation plots for each of the two classes we attempt to predict.

Following your comment, we added a new Table (Table 8) and a short description of this table to the discussion section (lines 682 to 694). In this table, we present the confusion matrix obtained when applying the classifier to our whole dataset as validation (circa 400 plots total). 

As explained in our reply to Point 4, the performance of the classifier is comparable for both the training and validation data. 

Round 2

Reviewer 2 Report

Dear authors,

The comments are as follow:

1, All picture should be updated using a high resolution.

2, The class error HS only were more than 20% in table 7, and the mean error were higher than 11%. What's the innovation in this manuscript? 

Author Response

Thank you for taking the time to review our paper. 

 

Point 1

All pictures should be uploaded using a high resolution.

Answer 1

Instructions for authors specify that all figures should have a minimum of 1000 pixels width/height or a resolution of 300 dpi or higher. All figures fit the 1000 pixels width requirement, and following this comment we made sure every figure that did not meet 300 dpi was upgraded in resolution to 300 dpi.

 

Point 2

The class error HS only were more than 20% in table 7, and the mean error were higher than 11%. What’s the innovation in this manuscript?

Answer 2

Errors between 10% and 20% are in the range observed in the literature for classifications using similar datasets, with some variability depending on the type of classification targets. 

In this paper, the innovation is not related to improving classification accuracy. Similar datasets are still novel for fluvial studies, and exploring the characteristics of both age and connectivity gradients simultaneously on a riparian forest is an innovation. For classification-related innovations, they are more related to the classification targets (and their relationship to forest - river connectivity) rather than the classification process itself.

Therefore, we fully cover the scope of this journal by featuring innovative applications based on combining hyperspectral, lidar, and field data which creates a cutting-edge dataset for our field of study. Please, refer to our previous answers for further information, or to the part of the introduction that explains which combining such data is innovative for riparian forests.