The SSR Brightness Temperature Increment Model Based on a Deep Neural Network

Round 1

Reviewer 1 Report

Authors propose "  The SSR Brightness Temperature Increment Model Based on Deep Neural Network". The topic is very interesting, and the paper is, in general, well-written. I, however, have the following comments: 

1.  Authors were missing and didn’t explain some important research works conducted in the other parts of the world which discussed about the SSR (Sea Surface Roughness) brightness temperature increment model.
2. Authors should show study area location in a map, model calibration and validation locations and add an additional figure

3          3. Authors should show modeled vs observed time-series comparison.

      

      4.  How did the authors deal with unbalanced dataset.

       5. Authors also need to check the sentence structure. In some lines, sentence structure is not correct. 

Authors also need to check few typos,  sentence structure. In some lines, sentence structure is not correct. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

General

Wen et al. submitted the article The SSR Brightness Temperature Increment Model Based on

Deep Neural Network to mdpi remote sensing for review. The paper investigate the use of deep neural network to estimate SSS Brighness Temperature using DNN. The study is conducted offshore on an unknown location but at a single location. The DNN model variables are θ, U 10 , SD, ΔT, Bp_SSR variables with 35000 records. Results show that DNN provide better estimation of SSS than forward conventional model. The use of single location made the model only valid at that location even it can be understand that it have global potential validity. The lack of other location record is the main limitation of the approach. The fact that a DNN outperform a conventional model can easy be understood. A general introduction of ML method for this type of application is missing or at least on ocean variable estimation. It may better to illustrate your Table 6 with a graph that we could spot quickly the lowest accuracy. Otherwise, approach is logical and follow standard.

 

Questions

Q1

Can the author indicates what are usually the main issue of conventional forward model? What specific part limit the model?

 

Q2

Is there a way that DNN could be used to improve the forward model in estimating some constant used in equations?

 

Q3

What the performance of the DNN on extreme SSS values like high and low?

 

Q4

5 variables were used to constrain the DNN model. How to be sure that enough? Why not using V10 in addition to U10?

 

 

Specific lines

Line 9 opposite look better.

The distribution and change of SSS (Sea Surface salinity) have a significant influence on the sea dynamic environment, marine ecological environment, global water cycle, and global climate change.

-->

the sea dynamic environment, marine ecological environment, global water cycle, and global climate change have a strong influence on the distribution and variabity of SSS (Sea Surface salinity).

 

Line 35

Essential Climate Variables

-->

essential climate variables

 

Line 69

data Processing

-->

data processing

 

Line 75 ? what article

This article conducted an offshore platform

-->

A experimental study was conducted on an ....

 

Line 217 add a reference to universal approximation theorem

 

 

   

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Overall the manuscript is well-written with no major grammatical errors. I noticed some minor spelling mistakes, which I have marked below:

 

This paper investigates a critical issue in satellite remote sensing of sea surface salinity (SSS) - establishing an accurate sea surface brightness temperature forward model. It points out that the sea surface brightness temperature consists of the flat sea surface brightness temperature and the sea surface roughness (SSR) increment. The flat sea surface brightness temperature can be calculated using relative permittivity models combined with Fresnel formulas. However, different relative permittivity models yield different flat sea surface brightness temperatures, leading to divergent results from forward models based on different SSR models. This has exceeded the accuracy requirements of SSS inversion. To address this problem, the paper proposes a universal neural network architecture and training scheme. Using offshore experimental data, it provides 18 different SSR brightness temperature increment models (including 9 H-pol and 9 V-pol models) based on deep neural networks for 9 widely used L-band relative permittivity models. These models were compared to prevailing models. The results demonstrate that the neural network models significantly outperform existing models, with accuracy approaching that of radiometers (0.1K). Therefore, this study effectively resolves the issue of SSR brightness temperature correction under different relative permittivity models, providing theoretical support for high-precision SSS inversion research. The innovation lies in proposing a universal approach to generate matched SSR increment models for any given relative permittivity model, thereby eliminating errors caused by model discrepancies and improving SSS inversion accuracy.

 

In summary, this paper addresses a key challenge in satellite SSS inversion using an innovative approach, achieving positive results that will help improve the accuracy of future SSS satellite products. Overall, the language flows smoothly and there are no spelling or grammar mistakes

 

Minor revisions:

I recommend adding a figure, apart from the accuracy tables of different models, to display the inverted data from different models and the actual observed data for comparison.

 

Line 100 and Line 104: paceborne should be spaceborne

Overall the manuscript is well-written with no major grammatical errors. I noticed some minor spelling mistakes, which I have marked below:

 

This paper investigates a critical issue in satellite remote sensing of sea surface salinity (SSS) - establishing an accurate sea surface brightness temperature forward model. It points out that the sea surface brightness temperature consists of the flat sea surface brightness temperature and the sea surface roughness (SSR) increment. The flat sea surface brightness temperature can be calculated using relative permittivity models combined with Fresnel formulas. However, different relative permittivity models yield different flat sea surface brightness temperatures, leading to divergent results from forward models based on different SSR models. This has exceeded the accuracy requirements of SSS inversion. To address this problem, the paper proposes a universal neural network architecture and training scheme. Using offshore experimental data, it provides 18 different SSR brightness temperature increment models (including 9 H-pol and 9 V-pol models) based on deep neural networks for 9 widely used L-band relative permittivity models. These models were compared to prevailing models. The results demonstrate that the neural network models significantly outperform existing models, with accuracy approaching that of radiometers (0.1K). Therefore, this study effectively resolves the issue of SSR brightness temperature correction under different relative permittivity models, providing theoretical support for high-precision SSS inversion research. The innovation lies in proposing a universal approach to generate matched SSR increment models for any given relative permittivity model, thereby eliminating errors caused by model discrepancies and improving SSS inversion accuracy.

 

In summary, this paper addresses a key challenge in satellite SSS inversion using an innovative approach, achieving positive results that will help improve the accuracy of future SSS satellite products. Overall, the language flows smoothly and there are no spelling or grammar mistakes

 

Minor revisions:

I recommend adding a figure, apart from the accuracy tables of different models, to display the inverted data from different models and the actual observed data for comparison.

 

Line 100 and Line 104: paceborne should be spaceborne

Author Response

Please see the attachment.

Author Response File: Author Response.docx