Mapping Deforestation in Cerrado Based on Hybrid Deep Learning Architecture and Medium Spatial Resolution Satellite Time Series

Round 1

Reviewer 1 Report

This manuscript provides a study about deforestation mapping in Cerrado using hybrid deep learning and time series remote sensing data. The study is interesting and the manuscript is well written. It is in the scope of the journal of Remote Sensing. I have some comments as follows:

1. Line 7, ‘LSTM’ should be given its full name when first appear.
2. Lines 208-212, the parameters of LSTM is important for the deforestation mapping model. How to determine these parameters should be given more information.
3. The input of the LSTM should be clearly presented, the time series NDVI, EVI or the spectral reflectance?
4. The combination of LSTM and U-Net is based on two steps in this study. There are also studies directly combine the LSTM and U-Net model in one step. Which method can give more accurate deforestation mapping accuracy?
5. The overall accuracy of more than 99% is really high, can you give an map to show the validation samples distribution?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript is well designed and structured, I just have a question about the reason of using three approaches of training samples. For approach 1, is the adjacent area landcover the same as the study area? 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Reviewer’s Report on the manuscript entitled:

Mapping Deforestation in Cerrado based on Hybrid Deep Learning Architecture and Medium Spatial Resolution Satellite Time Series

The authors propose a new method of detecting deforestation in Cerrado via  Long Short-Term Memory (LSTM) and U-Net applied to Landsat and Sentinel image time series. The paper is overall well-written and interesting but not organized well. Particularly, the Introduction must be improved. The limitation of the study should also be highlighted. Below, I have listed my comments to help the authors to improve their manuscript.

 

The Introduction must be improved. Currently, the objective of the paper is mentioned in the last paragraph only in one sentence.

The objective of the paper is to detect deforestation via Artificial Intelligence (AI) techniques. However, the introduction talks more about the study region (lines 19-31 which part of it may go to Section 2.1), politics, and management (lines 32-65) rather than the AI techniques and challenges. The authors may break the introduction into subsections and talk more about the AI techniques (time series analysis, change detection, machine learning, deep learning, etc.) for creating change maps. For example, the authors need to add the recently developed methods for creating change maps based on satellite image time series, such as

Jumps Upon Spectrum and Trend (JUST):  

https://doi.org/10.3390/rs12234001

whose software is freely available online at

https://geodesy.noaa.gov/gps-toolbox/JUST.htm

 

Breaks For Additive Season and Trend (BFAST) Lite:

https://doi.org/10.3390/rs13163308

whose software is freely available online at

https://github.com/bfast2/bfast

 

Change Detection via Convolutional Neural Networks:

https://doi.org/10.3390/rs12060901

 

Please also write a couple of lines in the Introduction to describe how the paper is organized.

Please define the acronym LSTM in the abstract and Introduction. All the acronyms must be defined the first time they appear in the paper. That also includes NDVI and EVI. For their definitions, the authors can also refer them to the first reference above (i.e., JUST).

Please add an acronym table at the end of the manuscript showing the list of all the abbreviations used in the paper.

Line 81. Please write it as “Taquary [31] proposed ….”. Please note that a sentence should never start with a reference number. In this case, the author’s names should come first. That also applies to Lines 119, 368, etc.

 

Line 84. Replace “He” with “They”. Please refer to other authors as “They”.

Line 86. “obtained in” not “obtained by”. Also, Line 264. “recommended in” not “recommended by”.

Line 395. Through further inspection? What inspection? Did the authors consider the day or month of the year (e.g., climate)?

Limitations of this study should be highlighted in the discussion section. There are currently two classes discussed: Deforestation, Natural Vegetation. So, what happens if the natural vegetations burn due to wildfire and then recover a few years after? Are they being classified as deforestation? What about changes caused by insect attacks, urbanization?   If these classes (i.e., changes caused by fire, urbanization, insect attack, drought, flood, phenological change, etc.) come to the proposed deep learning method, then most likely the f1-score and accuracy will decrease with a significant increase of false-positive and false-negative rates. In fact, I suspect that basic threshold techniques applied to the vegetation indices, such as NDVI or EVI can perform well for separating Deforestation from Natural Vegetation!

 

I leave the following comment optional:

The authors take three approaches for selecting training data sets, but what about showing a basic threshold method for comparison purposes? In Table 5, The overall accuracy is greater than 99%. Based on what I mentioned above, this study is only focused on separating Deforestation from Natural Vegetation that is a relatively easy task that can also be achieved by basic threshold methods. For example, if the authors plot the histograms of deforestation and natural vegetation using NDVI for the training regions and find the value where the two curves (histograms) intersect, then they can obtain the threshold for separating deforestation from natural vegetation and subsequently use that threshold for the study region to separate deforestation from natural vegetation. Would the result be better, worse, or almost the same with respect to the proposed deep learning method?

Thank you for your contribution

Regards,

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

the paper has been improved

Author Response

Dear Reviewer

We would like to thank you for reviewing our manuscript and also for your contributions.

Best regards,
B. M. Matosak.

Reviewer 3 Report

I would like to thank the authors for addressing my comments in a satisfactory manner. I have a few more minor comments to be considered:

Line 1. in the abstract please instead of “km^2” mention “square kilometers”. I also suggest doing the same in the manuscript too wherever you have km^2.

Line 18. At the beginning of the Introduction, please briefly describe how you organized the Introduction (i.e., Subsections 1.1, 1.2, 1.3, 1.4) so that the Introduction suddenly does not start with a subsection.

Line 230. Please mention “The rest of the paper is organized as follows.”

Please carefully proofread the article before publication

Thank you for your contribution

Author Response

Please see the attachment.

Author Response File: Author Response.pdf