Next Article in Journal
Genetic Programming for High-Level Feature Learning in Crop Classification
Next Article in Special Issue
Forest Height Inversion Based on Time–Frequency RVoG Model Using Single-Baseline L-Band Sublook-InSAR Data
Previous Article in Journal
Spatiotemporal Variation in Compound Dry and Hot Events and Its Effects on NDVI in Inner Mongolia, China
Previous Article in Special Issue
A Selection of Experiments for Understanding the Strengths of Time Series SAR Data Analysis for Finding the Drivers Causing Phenological Changes in Paphos Forest, Cyprus
 
 
Article
Peer-Review Record

Deep Learning for Mapping Tropical Forests with TanDEM-X Bistatic InSAR Data

Remote Sens. 2022, 14(16), 3981; https://doi.org/10.3390/rs14163981
by Jose-Luis Bueso-Bello *, Daniel Carcereri, Michele Martone, Carolina González, Philipp Posovszky and Paola Rizzoli
Reviewer 1: Anonymous
Reviewer 2:
Reviewer 3:
Remote Sens. 2022, 14(16), 3981; https://doi.org/10.3390/rs14163981
Submission received: 13 July 2022 / Revised: 5 August 2022 / Accepted: 9 August 2022 / Published: 16 August 2022
(This article belongs to the Special Issue SAR for Forest Mapping II)

Round 1

Reviewer 1 Report

The paper aimed to establish an RS image classification technique using the TanDEM-X data, which may benefit forest resources assessment in the Tropical region often covered with clouds for most of the year.

I appreciate the authors' effort and contribution. I suggest some revisions for publication.

1. Abstract is long. The journal guidelines suggest about 300 words.

2. The description of 2.2 is confusing. I would suggest decomposing and including them in the method section, such as Landsat Tree Cover Map, into 4.2. 

3. Likewise, 4.4 should elaborate more on the accuracy assessment procedure incorporating some parts of 2.2.

4. Elaboration on how to downsize the images to the resolution of 50m is required, which has been done for several archives. 

5. L258 Please elaborate on how the weight was determined.

6. L260-265 The point of the description is not clear. Please elaborate.

7. 5.3 The use of NDVI and NDMI should be explained in the Method section. Likewise, elaboration on threshold selection, 0.7 and 0.25, is required.

8. L606-610 I would suggest deleting this part. Not directly relevant to the main topic.

8. Please place a scale or coordinates for images.

Author Response

Dear reviewer, please find attached our responses. Thank you.

Author Response File: Author Response.pdf

Reviewer 2 Report

A fairly informative paper devoted to an topical and important problem of a forest map automatic extraction from radar interferometry data.

Unfortunately, the authors, justifying by the lack of computational resources, discard of full resolution (12 m) radar data processing, and limited it by quicklooks (50 m) only. It would be useful for readers to add an example of full resolution radar data analysis over at least a small area of the land surface.

The authors employ the “supervised clustering algorithm” term (for example, in line 6). In fact, the “clustering” and “supervised" concepts contradict each other. The authors should explain in detail how the “supervised” is applied to the well-known data-driven fuzzy clustering algorithm. Or, alternatively, leave either the “clustering algorithm” or the “supervised classification algorithm”.

The derivation of equations (1) – (4) for obtaining the main factor, i. e. volume decorrelation γvol almost completely rewords the already published material [Martone et al., 2018] by the same authors and thus bear the mark of partial self-plagiarism.

It seems insufficient for a convolutional neural network to engage a training/control set of 455 (line 336) and 376 (line 345) images only.

As declared by the authors the overall accuracy (OA) achieved is 85-90 % (Table 3), while in the mentioned work [Martone et al., 2018] a 90 % average OA is claimed using a fuzzy clustering algorithm. The authors should more convincingly justify the need to apply a convolutional neural network instead of the previous algorithm.

It would be good to describe the possible future research in conclusion.


Reference

 

Martone M., Rizzoli P., Wecklich C., Gonzalez C., Bueso-Bello J.-L., Valdo P., Schulze D., Zink M., Krieger G., Moreira A. The global forest/non-forest map from TanDEM-X interferometric SAR data. Remote Sensing of Environment, 2018, vol. 205, pp. 352-373. DOI: 10.1016/j.rse.2017.12.002

Author Response

Dear reviewer, please find attached our responses. Thank you.

Author Response File: Author Response.pdf

Reviewer 3 Report

Forests are one of the key players for the environment of our planet. In this paper, based on the use of deep convolutional neural networks, the authors develop a specific supervised training strategy for large-scale forest mapping over tropical areas by utilizing TanDEM-X bistatic InSAR acquisitions. The product is highly accurate, time-tagged and gap-free. This is a good work to map the tropical forest using the SAR data, therefore, I suggest to accept the manuscript after minor revision.

The detailed comments are as follows:

1. The coherence is a reliable indicator of the presence of water on the ground, since water surfaces typically appear as the completely decorrelated areas. That is, the water should be easily classified as the non-forest. However, in this study, the water is thought as the third class, why? Moreover, from this aspect, the seasonal change of the forest is apparent in the temperate region. On such condition, the coherence of the TanDEM-X bistatic data is also bad. Then, what is the difference between the coherences of the water and the forest? Please give a brief introduction in theory.

2. The authors feed the network using the following set of input feature maps: the SAR backscatter, the interferometric coherence, the volume decorrelation factor, the local incidence angle, and the height of ambiguity. Please give the reason to select these features.

3.Line 113: change “asses” to “assess”.

4.Line 247: This is a wrong sentence in expression.

5.Lines 255-257: three different sets of cluster centers were derived for the generation of the global product, including “tropical, temperate and boreal forests”. I suggest to give a brief introduction of the three sets.

Author Response

Dear reviewer, please find attached our responses. Thank you.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

All comments are well addressed. Therefore, I have no objections to publication.

Back to TopTop