Retrieval of Atmospheric Temperature Profile from Historical Data and Ground-Based Observations by Using a Machine Learning Algorithm

Round 1

Reviewer 1 Report

This study proposed a machine learning algorithm to retrieve the atmospheric temperature profile from historical data and ground-based observations. And the results show that this method seems to be feasible. However, some issues need to be resolved.

(1)  The horizontal and vertical axes of Figure 4.a are both represented by Temperature, and the authors should distinguish them.

(2)  For tables used to compare error values in the text (such as Table 3,4,5), the authors should bold the optimal value.

(3)  In Table 4, the convolution kernel with a size of 4*4 performs better than the convolution kernel with a size of 3*3. Why did the authors choose a convolution kernel with a size of 3*3? The training time of the two is not much different, and the priority of accuracy should be higher.

(4)  There are errors in lines 305 and 306, which do not correspond to the formulas, such as N_s, y_i, etc.

(5)  Please specify the specific composition of the dataset, including sample features, training labels, etc.

(6)  The size of the INPUT in Figure 6 is 10*28, please explain the meanings of 10 and 28.

(7)  The network model in this study is a combination of CNN and MLP, which are both the most basic network models. Please explain the reason the authors chose this architecture. Compared with other network models such as VGG, RNN, etc., what are the advantages of model in the study? Please explain the reason for your choice in terms of model effect and calculation cost.

(8)  It is mentioned that the CNN module can reduce the time complexity of the model, but whether this will reduce the effect of the model, if possible, please provide an ablation experiment to verify it.

(9)  As mentioned in section 2.1, all the temperature data and meteorological data in the area are averaged. Will this cause information loss?

(10)             Table 3, Table 4, and Table 5 respectively verify the influence of different parameters on the model, while the authors should control the variables to verify the impact of a change in a certain parameter on the results. For example, compare the influence of different channels in the first layer of convolution under the condition that keeping the number of neurons in the hidden layer and the size of the convolution kernel unchanged.

Author Response

Thank you for all your valuable comments. On this basis, we have carefully revised the manuscript and a detailed response can be found in the attached document.

Author Response File: Author Response.pdf

Reviewer 2 Report

Lines 32-35: Please specify the atmospheric regions and/or altitude range of the temperature profile focused on in this study

Line 49-50: “climatological temperature profiles” instead of “typical temperature profiles”

Lines 60-61: experimentally, radiosondes do not “require large amounts of resources”, they are fairly easy and inexpensive to launch. Better to emphasize points made afterward that they lack spatiotemporal resolution, global coverage, and continuous measurements in time.

Line 71: Temperature profiling lidars, which are typically rotational Raman lidar systems, can measure throughout the troposphere and even into the stratosphere. They are not limited to the boundary layer as this text implies.

Line 77-81: these belong in the previous paragraph

Near line 114: It would be good to point out limitations of remote sensing that ML studies could improve on, including: cost of building and running instruments; measurement limitations (e.g. cloud cover); limited global/spatial coverage

Lines 115-126: Please describe what kind of measurements, if any, go into ERA (e.g. does it include radiosonde measurements?)

Lines 115-126: This study could be strengthened if radiosonde or remote sensing temperature profiles were used for training data, rather than a model. This could be in addition to study using ERA data. Actually, it would be nice to see ML results and comparison using both sets of training data: actual measurements and ERA model data

Lines 164-166: I think this wording needs to be removed

Section 2.2: I see that at least the ERA was compared with measurements

Section 2.2: is the GIIRS instrument an example of the infrared hyperspectral remote sensing technique mentioned in Section 1? If so, it should be mentioned as an example in Section 1. If not, a description of this type of sounder should be added to Section 1.

Seminal work on rotational raman lidar should be included:

·        J. Cooney, “Measurement of atmospheric temperature profiles by Raman backscatter,” J. Appl. Meteorol. 11, 108–112 (1972).

·        Andreas Behrendt at the University of Hohenheim

Author Response

Thank you for all your valuable comments. On this basis, we have carefully revised the manuscript and a detailed response can be found in the attached document.

Author Response File: Author Response.pdf

Reviewer 3 Report

This manuscript proposed a machine learning method for retrieving atmospheric temperature profiles and compared the method with traditional methods. The study is a welcome addition to the literature for the field of machine learning (deep learning) and atmospheric research, especially in an era where available satellite data and re-analysis data are increasing dramatically. However, the manuscript has significant problems in terms of paper logic and language presentation. To be specific:

 (1) Introduction

L41-42: What is the logic between the first paragraph of the introduction, which already emphasizes the importance of the atmospheric temperature profile at the beginning, and the atmospheric temperature at the end? Is the study objective of the manuscript the atmospheric temperature or the atmospheric temperature profile?

L41-42: Is remote sensing becoming an important issue because of the need for spatial and temporal observations of atmospheric features?

L122-126: why does the data description appear in the last paragraph?

 (2) Data presentation.

The description of the study area overview is missing, and the data introduction section is mixed with a large amount of content on the analysis of results. The content of this chapter only needs to introduce the overview of the study area and the basic information of the data used in the study. The comparison of different data and the time series analysis of the data should not appear in this chapter.

 (3) Methods

The method is mixed with a large number of descriptions that are not relevant to the method of this study. The proposed neural network model is missing many key elements, such as training data generation.

 (4) Results and Discussion

The first paragraph seems to be part of the methodology section.

Therefore, the manuscript needs improved writing logic and additional experimental details.

Author Response

Thank you for all your valuable comments. On this basis, we have carefully revised the manuscript and a detailed response can be found in the attached document.

Author Response File: Author Response.pdf

Reviewer 4 Report

The study evaluates deep learning networks merging real-time ground observation and atmospheric profiles to retrieve atmospheric soundings. The authors consider three sources of data (ERA-5, FY-4A , and radiosounding) and data applicability for their methods. After a description of their model, they analyze its performance. The paper is generally unclear and challenging to evaluate due to poor descriptions and low-quality language. The introduction is incomplete and misses a deeper discussion of limits associated with several sounding retrieval methods. Moreover,  the state-of-art of retrieval methods using machine learning techniques is completely missing. It is unclear what is considered the "Reference", and the deep learning networks' performances are evaluated by basic metrics (cross-correlation, mean bias, and RMSE): accuracy, loss and confusion matrix are recommended. Finally, the paper is missing of a deeper discussion of the results, with the interpretation of agreements (or disagreements) in accordance with atmospheric physics.

Author Response

Thank you for all your valuable comments. On this basis, we have carefully revised the manuscript and a detailed response can be found in the attached document.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments for author File: Comments.pdf

Author Response

Thank you for your valuable comments. We have checked the 
manuscript and revised it according to the comments.

Author Response File: Author Response.pdf