Classifying Forest Type in the National Forest Inventory Context with Airborne Hyperspectral and Lidar Data

Round 1

Reviewer 1 Report

Mapping forests at species level remains a challenging task despite the great advancements in both remote sensing data and methods. In this study the authors identify forest types according to the specifications of the FIA program using the vegetation spectral and structural characteristics, both derived by remote sensing data. Five different classifiers are employed and different combinations of data reaching an overall final accuracy of 78%.

In my view the novelty of the study is not high enough for a journal such as Remote Sensing. There are numerous studies over the last 20 years dealing with the performance of the various classifiers and the suitability of data for forest and ecosystem mapping. In my view there is no classifier that can be considered superior to the others under any circumstances. The performance of each classifier is strongly dependent, among others, on the data employed, the training set, and the structure and composition if the landscape being mapped. Perhaps the manuscript would be more suitable for a journal like FORESTS since it presents an applied piece of research on forest mapping for management and assessment purposes. However, this decision belongs to the editor and not to the reviewer.

Regarding the manuscript it is well written and well presented and it reads nicely throughout. However, I would recommend the authors to change the units to the metric system (meters instead of feet) because currently at some points it is the English and in other the Metric system which is used.

The overall performance of the various classifiers achieved in the manuscript is not particularly high even for the best performing classifier. Table 14 for instance shows classification accuracies as low as 0% for some particular forest types and the high overall accuracy is simply the result of the high accuracy of black spruce which also has the highest participation in the ground truth data. The ground truth dataset is extremely unbalanced and I am sure this affects the classifier and of course the accuracy assessment. The authors could also provide a rough assessment based on previously published works on the relevant participation of each forest type in the landscape.  I would like to see the performance of the classifiers with a more balanced training set, and I wouldn’t be surprised if a different classifier performs better. Perhaps the authors could omit some plots of Black spruce and Paper Birch (using a maximum of 100 plots) and assess the performance of the classifiers and the data. The way the training and the testing was made using the entire set of available plots produce in my view misleading results.

The manuscript has also some minor spelling mistakes which I am sure they will be spotted and corrected in the revised version of the manuscript.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

See attached Word file for comments.

Comments for author File: Comments.docx

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

This paper proposes a fusion of airborne hyperspectral and lidar data for classifying forest type with Forest Inventory and Analysis (FIA) plot data. Five classification algorithms were tested with six different combinations of raw hyperspectral data, hyperspectral vegetation indices, and lidar-derived canopy and topography metrics. Some experiments demonstrate that combining both structural (LiDAR) and spectral (imagery) data yields the best results for forest classification. The main comments are

1. The motivation and innovation of this paper is unclear in abstract. Please rewrite the abstract.
2. Please give some descriptions about hyperspectral and lidar data referring to 10.1109/TGRS.2020.2963848 and 10.1109/JSTARS.2014.2332337 in introduction.
3. A process steps should be given to explain the flowchart of the proposed method.
4. How to fuse the hyperspectral and lidar data should give a detailed description.
5. In experiments, please enhance the analysis about why the fusion hyperspectral and lidar data can achieve better results than the other data.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

See attached comments.

Comments for author File: Comments.docx

Author Response

Dear Reviewer #2,

Thank you again for taking the time to further improve this manuscript. Your feedback is truly appreciated and has been taken into consideration in this most recent round of edits. In your first comment you suggested that we acknowledge how forest class is defined and discuss some of the limitations at the start of the manuscript. In response to this we have added some language to both the introduction (section 1) and the FIA field data (section 2.2) to clarify how forest type is classified and the limitations associated with this. We left the language in the discussion to ensure we discuss this point throughout the manuscript.  I was unable to find literature that discusses this issue, as FIA forest type is not often used.

In your second comment you expressed concern over the use of Kappa in this manuscript. We agree that this may be a situation where we must “agree to disagree”. We firmly believe in including kappa in this manuscript because it greatly enhances our ability to compare to similar studies such as Dalponte, 2008 (reference 69). You mentioned having issue with our use of table 10 as it is a very old table from a 1977 paper. We have removed the table and simply integrated the kappa values seen in table 10 into the manuscript.

In your third comment you had two primary points: first you suggested that the language in L260 suggested that our techniques eliminate the possibility of overfitting. We agree that this is not true and have modified the language of L260 (or Line 284 in the most recent version) to make it clear that overfitting can still occur, but cross-fold validation helps reduce the likelihood. In your second point you expressed concern over the uncertainty associated with the metrics used in this manuscript. We agree that the metrics we used in this manuscript have uncertainty associated with them. An alternative approach for future would be to carry out a hypothesis test to determine if the differences in accuracy between the models are statistically significant. This is beyond the scope of this work, but this should be addressed in future research directions. I have included this language in the discussion to make this clear.

In your fourth comment you suggested changing the accuracy measures to precision and recall. We agree that this is something that future work should consider including, but we were unable to do at this time. To accommodate readers who wish to calculate their own accuracy measures, I have added in the original counts to Table 14. This study has limitations, as the main objective of the study was to provide an exploratory comparison of these techniques using well-known accuracy statistics (that are somewhat robust to unbalanced dataset).

Finally, in your minor comment you advised we remove the appendix lines as they are not applicable to this manuscript. We agree and have removed these lines.

Thank you again for taking the time to review this!

Sincerely, 

Caileigh Shoot 

Remote Sensing Research Specialist 
Strategy & Technology 
Advanced Forest Systems 
Weyerhaeuser 
Email: [email protected]

Reviewer 3 Report

All my concerns have been addressed.

Author Response

We are so glad your concerns have been addressed. Thank you again for taking the time to review this!