Improving Forest Baseline Maps in Tropical Wetlands Using GEDI-Based Forest Height Information and Sentinel-1

Round 1

Reviewer 1 Report

It is a very interesting and timely article, as defining the forest boundary with respect to climate change and global forest loss is one of the key challenges. 

Although the introduction is somewhat brief and could be expanded, from my point of view it contains all the necessary references and is sufficient. The objectives and hypothesis of the study are also clearly defined.

I have comments on the selection of surfaces where manual selection by the user has been used. Personally, I would have preferred a regular grid of points. In addition, it is not clear from the text whether the selection of training points was limited to sites near watercourses. The distribution of training and validation points is not uniform and the validation results may be affected by the different selection method. 

Although the results of this work show a refinement of the FNF mask when using GEDI/Sentinel 1 data, the refinement is relatively small and the accuracy  of the GEDI data will be sufficient for many users. In addition, working with radar data requires significantly higher processing requirements.

Despite the above criticisms, I consider the results of the study to be beneficial and recommend the paper for publication, as the developed approaches and validation of the GEDI data may be beneficial for further studies.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The reviewed manuscript uses Sentinel-1 and the forest height product that is generated by GEDI and Landsat to produce the Forest/Non-Forest (FNF) masks in tropical wetland areas. The experiments have shown that the accuracy of their FNF is greatly improved compared to the existing global FNF masks. 
The reviewed manuscript is well organized and the language is sufficiently good. The methods presented in this paper are straightforward and written in details. The results clearly show that the GEDI forest height, as well as the texture metrics derived from Sentinel-1 could provide valuable information for the classification of forest and non-forest. 

There are some parts of this paper that need to be improved before the manuscript can be accepted:
1. For the introduction, are there any other forest height products? I suggest to add more literatures regarding the forest height product and why the authors choose the one generated from GEDI and Landsat over other products?
2. It’s not very clear for me how the PT (probability threshold) is used to determine the class of a pixel. Is the PT a threshold that is applied on the probability of forest class (or non-forest class), only the pixels with the probability of forest that is higher than PT, are classified as forest. Or is the PT a confidence measurement of the classification? The implementation of the PT is not elaborated in the paper, so I would recommend the authors add more details regarding to the PT and how the class is determined by the PT.
3. In line 319-320, the authors made the statement that the GEDI-FH/S1 mask is more conservative in allocating the forest class compared with the GEDI-FH mask, which can be verified based on the results of Area 1 and 2 in Fig.6. While in the result of Area 3, GEDI-FH mask tends to give less forest and more non-forest. Therefore, in my opinion, it would be better to add some quantitative evidences to support this statement, like the accuracy (UA and PA) of forest or the confusion matrix of the validation data.
4. In the conclusion line 384, the authors have written that “our model yields the highest forest UA and non-forest PA of all masks compared”, however, I could not find any description of the forest UA in the paper. I would suggest that the authors either add the description of the forest UA or modify this statement.
6. For the FNF mask produced by Global Forest Change, why different canopy cover thresholds are used in the validation data (60% in Tab. 4) and the test area 1, 2 and 3 (80% in Fig. 6).
7. As the authors mentioned in line 151, the used global FNF mask is based on different inputs that have different forest definitions. You should discuss this issue and it would be good if you could estimate the influence of the different products to the assessed accuracies within you study.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This study combined a GEDI Forest Height product and Sentinel-1 radar data to improve FNF masks in wetland areas in Gabon using a Random Forest model. Overall, this study is contributive. While the writhing is insufficient, especially in INTRODCUTION and DISCUSSION parts.The specific suggestions are follows：

1. Abstract: 1) Authors said they improved the accuracy by 20%. There are many FNF products, which one is authors compared to? Please be specific. 2) The sentence, “In radar-based applications, moisture dynamics in unmasked non-forest areas can lead to false detections”, is unclear. Please directly report the mechanism and advantages on combining GEDI height and Sentinel-1 to remove the moisture dynamics in unmasked non-forest areas.
2. Introduction: Page 2 Line 53, …where short vegetation…-> …where short vegetations….; Why authors used Sentinel-1 radar data? Author just reported that “Combing a forest height product with radar data could improve the forest mapping accuracy and remove Landsat artifacts from the resulting map” without any reference. And why combing a forest height product with radar data can improve the forest mapping accuracy? What the role of radar data? Because radar data are sensitive to the water content and surface roughness? Why this sensitivity can improve the accuracy and remove Landsat artifacts? Please be specific.
3. Materials and Methods: 1) The vegetation and topography should be introduced in the 2.1 Study Area; 2) Pagle 3 Line 97, the shortage of radar FNF product (ALOS FNF) should also be reported; 3) The spatial resolution is different among the FNF products, how authors solved the uncertainty resulting from different resolutions? 4) Page 4 Line 115, did author use Landsat and GEDI data to acquire wall-to-wall forest height products? Or just acquire the foot-level forest height products? Please be specific. I suggest authors draw the figure to display the acquired forest height; 5) Wy authors use RF model for classification? Please add the reason as well as the reference.
4. Discussion: The roles of GEDI and Sentinel-1 data should be discussed. The uncertainty of classification should be discussed. Cautions about the application this methodology in other wetland areas should be discussed.
5. Conclusions: This part should be concise. Do not repeat the content that reported in above parts. Pleas highlight the contribution of this study, and cautions about the application this methodology in other wetland areas.
6. Reference: Please revise the format according to requirement of this journal.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Dear authors,

my comments are considered in an appropriate form and thus the manuscript can be accepted from my point of view.

 

Reviewer 3 Report

Actually， this paper aims to improve the non-forested area classification accuracy. Hence, I suggest the author further clearly illustrate or explain the aims in the section of Abstract, Intruduction and conclusions.