Skip to Content
MDPIMDPI

Advanced

by
  • Rohit Mukherjee 1,*,† and
  • Desheng Liu 2,†

Reviewer 1: Anonymous Reviewer 2: Monica Pepe

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors


Comments for author File: Comments.pdf

Comments on the Quality of English Language

English is fine only minor corrections required. 

Author Response

Thank you for your comments and suggestions. These recommendations will help improve this paper.

 

Our responses to your comments:

  • Comment 1: Are this image translation technique works well over various geographical scales or limited to this specific area? Are the seasonality affects the translations accuracies?
    Spatial generalizability is difficult to achieve, however, this method can be applied to any location. A CGAN for each region trained on the samples from that specific region will be the most optimal. Additionally, seasonality will not affect the method if the training dataset contains training pairs from that season. Otherwise, predicting a S2-like image from winter when only L8 images from summer were used, will not yield accurate results. Since the cost of the training dataset is low, i.e., built through publicly available images, this CGAN can be trained for winter seasons too.
  • It is unclear what kind of data authors used in this translation’s technique? For example, raw DN values or Level 2 atmospherically corrected data? 

We have added more details to the data source of Landsat 8 OLI and Sentinel-2 MSI to the paper (Page 3, Lines 131-133). Each scene was Level 2 atmospherically corrected surface reflectance values.

  • Comment 3: Does the angular geometries of both sensors do affects translation accuracy?

The differences in angular geometries and other instrumental differences are taken into account and CGANs are expected to learn these differences from each training sample.

  • Comment 4: I this translation techniques appliable also for thermal infrared channels of both sensors?

Predicting S2-like thermal bands (that are not available in S2) from L8 thermal bands will be more complicated since there are no S2 bands that have a strong correlation with other S2 bands. Unlike, the red edge bands which had a strong correlation with other S2 bands like NIR. However, GANs have been used to translate between optical and radar which is an even more complicated task. Therefore, such an experiment can be attempted.

  • Comment 5: In this study authors used CGAN technique, however in many sections of the article authors indicate GAN instead CGAN this need to be correct. 

We have corrected all occurrences.

 

We have fixed all suggested technical corrections and fixed some grammatical errors. Thank you for reviewing our work thoroughly.

Reviewer 2 Report

Comments and Suggestions for Authors

The study deals with the topic of time series improvements by Landsat 8 OLI imagery translation to Sentinel 2 MSI. The manuscript is well organised and clear. The approach presented is scientifically sounding, and the results interesting.

I have no major concrns, just some minor comments, most of them could be found in the revised PDF.

I only add some general comments here: 1) the text in some parts is not as accurate as the rest of the manuscript (references or cross-references are missing, concepts not explored in depth), be careful to review the manuscript. with an eye on this aspect. 2) Authors are translating imagery acquired by a satellite sensor to another satellite sensor, not a mission to another mission, so please do not forget to add OLI and MSI throughout the text.

 

Comments for author File: Comments.pdf

Author Response

Thank you for helping us improve this paper. The suggestions are really useful.

These are the changes we made based on your recommendations:

  • We removed the section descriptions at the end of the introduction, We agree it adds nothing to the paper.
  • We fixed the confusing Figure 1 caption.
  • Fixed minor typographic suggestions:
    • incorrect/missing references (multiple locations)
    • unnecessary quotations (Page 6/7, Lines 197, 198)
    • Added space before paragraph (Page 7, Line 219)
    • Improved sentence construction (Page 14, Lines 403-404)
  • We changed the ground truth to reference (Page 6, Line 203)
  • Fixed the equation and the explanation (Page 7, 230-238)
  • Removed the sentences explaining the differences between CGAN and GAN
  • Explained acronyms for PSNR and RMSE. Added a reference for PSNR. (Page 8, 256-257)
  • The images from Landsat 8 are pixelated but the images after CGAN translations are not. It might not be very obvious visually, however, it is supported by the spatial correlation metrics. All the Landsat 8 imagery is less pixelated after applying CGAN.
  • The multistep approach included some spatial features that were highlighted due to the stretch we applied for the figures. That was a neat observation!
  • We have improved the explanation of why CGAN will perform better during testing as opposed to CNN (Page 13, Lines 381-387)
  • Added OLI and MSI to Landsat and Sentinel-2 respectively, wherever applicable.
  • We agree with the suggestion for the future work. Analyzing the performance of each biome and seasonal changes will be interesting.  Applying this method to capture phenological parameters, such as leaf area index, will demonstrate the usability of the method.