Gaussian Process and Deep Learning Atmospheric Correction

Round 1

Reviewer 1 Report

* it would be helpful to the reader if more detail on the MODTRAN run parameters were shared in either the text or figure captions for Figures 1 and 2, such as atmosphere model and aerosol model used

* it would be helpful if the results of the study were compared to direct runs of QUAC or FLAASH. there is a mention of comparing the results  to "the well known QUAC method", however there is only the author's  own "QUAC style" universal mean regression prediction, not the native QUAC code (such as is available in ENVI). or can the author's present a verification that their QUAC style implementation of atmospheric compensation does indeed match results from running the actual QUAC compensation?

* it would be helpful to the reader if the author stated more detail on where the reflectance spectra came from. are they from an atmospherically compensated Hyperspectral image? or from a reference library of some sort? most real hyperspectral image products don't have an even mix of spectra from a reference library, but have a large amount of background spectra (i.e. vegetation, soil, sand etc) mixed with a smaller amount of man-made material spectra

* there is some slight inconsistency in mentioning the acronym QUAC and other times QUACC. same with FLAASH and FLASH

Author Response

We included a word document addressing the comments from all reviewers.  Because there is overlap between reviewer comments, we are including the point-by-point response for all reviewers together.

Than you.

Author Response File: Author Response.docx

Reviewer 2 Report

Atmospheric correction is a critical step in the processing of airborne hyperspectral data, which accounts for path radiance, aerosol effects and gas absorption to produce an accurate surface reflectance. Two deep learning models for atmospheric correction are presented in this study. The effectiveness of the proposed modules is assessed using objective metrics. The reasoning in the Introduction and Results of this manuscript, however, is somewhat shaky and challenging to understand. Besides, this manuscript requires additional clarifications and modifications on several issues. Some specifics are as follows:

 

1.      The Introduction should be written using background, significance, related works, issues, and study content as a framework. As for the section of Results, authors should divide it into several parts, including the evaluation metrics, results for the two models, and results compared to other models.

2.      On line 16, instead of stating the content of this study, the significance of studying atmospheric correction should be outlined.

3.      On page 3, each meaning of acronym such as DD, should be indicated in Figure 2 (left).

4.      Why did authors propose two models for the atmospheric correction? What are the characteristics and benefits of each model? What issues can each approach address? Moreover, what are the flaws in these three assumptions? Authors ought to elaborate in more details about the principles and merits of your two models.

5.      In the Data section, what are the details of the images used in this study? Are they accessible to the general public or self-measured? Specifics should be provided.

In Section 2.3 Denoising autoencoder…, what is the structure of the decoder? What is the definition of your objective function, and what is the optimization process of your model? Furthermore, authors should demonstrate the specifics of your experiments and parameter settings.

Author Response

We included a word document addressing the comments from all reviewers.  Because there is overlap between reviewer comments, we are including the point-by-point response for all reviewers together.

Than you.

Author Response File: Author Response.docx

Reviewer 3 Report

Review Basener

Radiance measured at a spectral sensor is converted to the reflectance of materials in a multispectral or hyperspectral image so that materials present in the pixel spectra can be identified. Here, two machine learning models for atmospheric correction are trained and tested with 100,000 batches of 40 reflectance spectra for conversion to radiance using the radian transfer calculation software MODTRAN. A theoretically interpretable Bayesian Gaussian process model and a deep learning autoencoder treating the atmosphere as noise are created. Both methods were compared for estimating the gain in the correction model to the QUick Atmospheric Correction Code (QUACC) method for assuming a constant mean endmember reflectance. It is concluded that the prediction of reflectance using the Gaussian process model is better than the other methods in terms of accuracy and reliability.

General comments

The structure of the paper is unusual – the Introduction includes methods and the Conclusions include results. It would be helpful to read in Methods, why the materials and methods were selected. Are any applications of the described methods available?

The abbreviation in the Abstract should be explained.

The paper addresses relevant scientific questions within the scope of the journal.

The paper presents novel concepts, ideas, and tools.

The scientific methods and assumptions are valid and outlined mainly so that substantial conclusions are reached.

The results are sufficient to support the interpretations.

The description of experiments and analyses is complete and precise to allow their reproduction by fellow scientists.

The quality and information of the figures and tables are fine. The captions should be more extensive so that one can understand it without reading the manuscript.

Title and abstract are too short to reflect the whole content of the paper. It should be adapted to the objectives and Conclusions

The overall presentation is well structured and clear.

The mathematical symbols, abbreviations, and units are generally correctly defined and used.

Specific Comments

Line 8: QUACC instead of QUAC?

Lines 87, 93 and 117: Explain the abbreviations.

The references include only some peer-reviewed papers – any more should be included.

Technical corrections

There should be a free space between a number and a unit.

Lines 326 and 336: page number is missing.

Author Response

We included a word document addressing the comments from all reviewers.  Because there is overlap between reviewer comments, we are including the point-by-point response for all reviewers together.

Than you.

Author Response File: Author Response.docx

Reviewer 4 Report

Atmospheric correction is the process of converting radiance measured at a spectral sensor to the reflectance of the materials in a multispectral or hyperspectral image. The authors presented two machine learning models for atmospheric correction trained and tested on 100,000 batches of 40 reflectance spectra converted to radiance using MODTRAN. Their machine learning model learns the radiative transfer physics from MODTRAN. The authors performed a theoretically interpretable Bayesian Gaussian process model and a deep learning autoencoder treating the atmosphere as noise. They compared both methods for estimating gain in the correction model to the well-know QUAC method. Prediction of reflectance using the Gaussian process model appears to outperform the other methods in terms of criteria accuracy and reliability.

Comments:

    1) In opinion of this reviewer, the authors should provide explicitly justified principal contributions of their method. They only wrote “In this paper, we provide a Gaussian process atmospheric correction process that does not make assumption 3, and a denoising autoencoder atmospheric correction process that can be used avoiding all 3 assumptions.” This justification made only for QUAAC method, but the authors did not explain how they can resolve all other drawbacks mentioned in related work section.

    2) The authors put discussing in the conclusion sect. This reviewer thinks that authors should present additional discussion sect where they can explain drawbacks of the second proposed denoising autoencoder method that did not perform well in comparison to two other used in comparison. In this section, they can explain possible modification of this framework.

Author Response

We included a word document addressing the comments from all reviewers.  Because there is overlap between reviewer comments, we are including the point-by-point response for all reviewers together.

Than you.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The authors did not answer all questions and recommendations of reviewer 3. This should be finished.

Additional detailed comments：

<General comments not answered: "It would be helpful to read in Methods, why the materials and methods were selected. Are any applications of the described methods available?", "The captions should be more extensive so that one can understand it without reading the manuscript.", "Title and abstract are too short to reflect the whole content of the paper. It should be adapted to the objectives and Conclusions.">
<Specific comments not answered: "The references include only some peer-reviewed papers – any more should be included.">

Author Response

Please see the attachment.

Author Response File: Author Response.docx