Modeling and Forecasting Ionospheric foF2 Variation Based on CNN-BiLSTM-TPA during Low- and High-Solar Activity Years

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The results of the paper are of interest for the specialists in geophysics, the solar physics, and applied mathematics. The proposed method of predictions seems perspective, well tested, and competitive with another ones.

 

Some explanations should be added in the text, to make the paper more readable.

1. The Ionospheric F2 layer Critical Frequency depends on the measurement time (day or night). Is it possible to give any comments about the time of measurements during a day?

2. In the basic Eqs. (3)-(6) not all parameters are defined, for instance W_xf. Please add the description of all the parameters there.

3. Some comments to Eq. (8) are needed. Is any relation to statistical methods there?

4. Again, some comments about the parameter W_h are needed for Eqs. (14)-(15).

5. The results of Section 3 are very interesting. But it seems better to specify the significant digits in Tables 7-10.

6. Line 270. ...From the figures, the proposed model has the smallest cumulative distribution error in the samples, indicating that it achieves higher accuracy in predicting foF2 compared to other models...

Is it possible to give more comments about this important result? From Figs. (6)-(8) it seems that in the proposed model the percentage is bigger.

7. English may be improved, minor editing.

 

The paper may be published after the minor revision.

Comments on the Quality of English Language

A minor editing is assumed.

Author Response

We have responded seriously to your valuable comments. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper tried to propose a new model to forecasting the foF2 variation, however, I do not think the paper can be published in the present version. At least, a major revision is needed. Dataset is the crucial issue. Please find my specific comments below:

1.       Figure2: Ht on the left side might not be useful.

2.       Equation 17: The output should be the series consequence of the three layers, but not in parallel.

3.       Equation 16: UTs and UTc (DAYs and DAYc), there are two many input variables here, would it lead to bad results or the iteration speed?

4.       The experiments are divided into two groups (solar high and low). Are they trained and tested by the same model? Or are they tested by two different trained models?

5.       The data of the experiment is not enough, only data in one year, with 8 months of training and 4 months of testing. Conclusions are not reliable based on only one-year data. I suggest to use more data to rerun the experiment.

6.       The training and testing data sets are selected by the authors, the data sets do not have the continuity to ensure the reliable of the results. The testing data might also have some relationships with the training data, I do not think the data set used in this study have the scientific meaning.

7.       Discussion of the paper should provide the limitation of the algorithm. And, what should be done to improve the model in the future?

8.       TPA model does not have much contributions on the results. The novel of the TPA is not obvious.

Author Response

We have responded seriously to your valuable comments. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript presents a CNN-BiLSTM-TPA model and uses it to describe and predict the variation of the daily ionospheric foF2 indices from three southern hemispheric stations in terms of sunspot numbers, solar F10.7 flux, geomagnetic Ap and Dst indices, UT time, and season, in the solar minimum year 2009 and solar maximum year 2014. The ms is interesting enough and may eventually be published. However, I have several questions on model reconstruction, obtained results and on the overall presentation, which require a rather extensive revision before publication can be considered.

 

Firstly, all stations whose foF2 indices are here analyzed come from Australia, i.e., from the southern hemisphere. Why this very limited choice? Any reasonable model to predict the foF2 indices must be made to apply for both northern and southern hemisphere. Otherwise, the comparison, e.g., to IRI model is not fair since it is formed to apply to a larger spatial area, thus explaining the larger global variability of conditions. Even if limited to only low latitudes and a limited longitude range, the model should be made to include the two hemispheres. Moreover, the ms does not sufficiently clearly mention the limited latitude-longitude range. Currently, it hardly mentions the limitation that it only deals with the southern hemisphere.

 

It is not clear to me how the modeling is done in detail, since presentation leaves many gaps. 

I guess you prepare one single model for all stations, seasons etc? Please explain in more detail in text and maybe mention in Fig. 3. 

You use 3/4 of data for training and the rest for prediction. Judging from presentation, I conclude that the training time is successive, not randomly sampled. However, if the total time is one year, how can you make predictions for early and middle year? Please explain. 

Model configuration of Table 1 is obscure as there is no intelligible connection between the values given here with the earlier mentioned model parameters like w, k, delta or others. How long into the future the model is made?  Be more informative on all details. 

 

Figures 4 and 5 depict only one day in each month, and one month in each season. This is not clearly mentioned when introducing the figures. Do the statistics in Tables 3 onward use only data for these days or all data? They must be for forecast times, not training times, right? What are they (see above).

Model does not seem to match better than BiLSTM in all of these days. Please discuss this openly and try to find reasons why not. I also pay attention that BiLSTM seems to do a better job than the model especially in the equinox months, contrary to the authors' later specific mention. What causes this mismatch?

Figure 9 depicts rather questionable results. It seems to me that BiLSTM distributions are normal but model values not always. Especially for Darwin, the model distribution is clearly skewed. Normality should be tested. Also, text mentions that "..the proposed method is superior..". This is definitely not the case for Fig. 9, nor do authors give any statistics for this claim. If they intend to leave Fig. 9 in the ms, they should do a proper analysis and present, e.g., the first three moments of these distributions, analyze differences between model and normal curve etc. 

 

Figure 8 should be repeated for solar maximum results. This could even replace current Fig.9.

 

 

 

Year 2009 is the solar minimum year between cycles 23 and 24 (does not belong to cycle 24) and 2014 is the solar maximum year of cycle 24. These are unique years, not just times of active and inactive Sun. Please always use the terms of solar minimum/maximum years. 

 

Data availability. Where are the results of this model to be obtained? Please provide the model output data in a freely accessible data service for the other scientists to be able to use this model. I hope that Chinese researchers and authorities will join the FAIR policy of scientific data. 

 

Finally, reading the author contributions, I wonder why W.H. is not the first author? He has clearly been more important for this manuscript than all others, including B.X.

 

Other notes:

- Line 68: Better title for Sec 2.2: "Season, Daily Cycle, Solar and Geomagnetic Activity"

- Line 70: Correct reference is Williscroft and Poole [15].

- Line 73 and onward. The regular term is "day of year", abbreviated DOY (not DAY). Use this throughout. 

- Line 84. Reference should also include basic reviews of F10.7 index, like Tapping, Space Weather, 2013.

- Lines 84-95. It has recently been shown that the F10.7 index and sunspots drift with respect to each other (Mursula et al., Astron. Astrophysics, 2024), which lessens their mutual dependence and correlation. This increases the motivation to take both indices into account.

- Line 98. IMF Bz without dash is better. Use in all occasions, including figures.

- Line 101. Campbell is not a proper reference to the Dst index. Use always original sources as references, here Sugiura, Annals of the International Geophysical Year, 35, 9, 1964; 
Sugiura & Kamei, Equatorial Dst index 1957–1986. In A. Berthelier & M. Menvielle (Eds.), IAGA Bulletin 40. Saint-Maur-des-Fossés, France. ISGI Publ., 1991

- Fig. 1 This figure should include all explaining parameters. So, add sunspot numbers and IMF Bz. Are all parameters given at daily resolution? Please mention the data resolution in caption as well as in text. Caption should start "Time series of..." (not Changes in...).

- Fig. 1 and elsewhere. Use Ap, not ap, consistently in the paper.

- Lines 117-119. You write "CNN has the capability to capture and convey the characteristics of ionospheric variations due to the strengths of convolutional operation." What is this capability? This must be explained in more detail.

- Lines 219-220. What are "sunspot-R12" and "F-peak" parameters of IRI and what are the other settings that are left at default value? A more detailed presentation of IRI is needed.

- Line 237. Lower case "the".

- Line 242. Repeat the months, instead of "in each month".

- Line 244. What average is this exactly? Be more informative.

- Fig. 8 and text. Which errors are these, MAEs?

- Lines 265-271. Why is higher value better than lower in Fig.8, although "..smaller values in each part indicate a closer approximation to the actual observed values"? I understand this but presentation is poor on this point. Be more complete.

- Line 276. Fig. 9 does not present results "Similar to the previous section,..".

- References. Many references includes a strange [J], [C] or [M]. What are these? Please remove unless essential.

Author lists used different styles. Please unify.

Author Response

We have responded seriously to your valuable comments. Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

In the manuscript a Modeling and Forecasting of Ionospheric foF2 based on neural network technique are presented and compared with other methods. The results seem satisfactory, but using UT instead LT is a basic flaw of the work. The stations that are used are separated in longitude by only about 23 deg (about 1.5 hours). If stations were distributed  more globally then the modeling would work less well. Therefore I recommend that a major revision is done. Preferably the  model should be adjusted to LT instead of UT.

Lines 78-79 and equation (2): foF2 and other parameters at low latitudes depend much more on local time (LT) than on UT, which is the same time irrespective of the location. I'm surprised that UT was used.

Figure 4: Is the x axis (Time) UT or LT. The maximum foF2 at Brisbane is roughly in the time interval  2-10, but it should occur in the afternoon hours. f0F2 increases as the photo ionization from the sun enhances the ionosphere. The effect has some delay, and so occurs hours after local noon. Thus it seems that really UT is plotted in Figure 4. The LT of Brisbane at 130 deg longitude is UT+130/15=UT+10.2 hours, for Darwin it UT+8.6 hours. The daily variations of foF2 at Brisbane should be shifted by 1.5 hours compared to Darwin. Using LT instead of UT would take care of this. If the modeling should be applied to stations more widespread in longitude than across Australia, it would be successful only when using LT. Because here stations with a LT difference of only about 1.5 hours are used, the modeling with UT still works, but is not optimal. 

The coordinates (latitude, longitude) of the stations Brisbane, Darwin and Townsvile are indicated at the beginning of the paper and in the abstract, no need to repeat the coordinates in section "Conclusions"

Author Response

We have responded seriously to your valuable comments. Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thanks the authors for their responses. I suggest to accept the paper this time after reading the revised paper, although I still think the four year data sets are not enough to do the experiment and selected months are not the best way based on the physical variations of the solar activities. 

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript has been essentially improved as to its scientific significance and transparency, as well as presentation style, consistency and clarity. 

Therefore, although some small issues related to research methodology and presentation still reman, I am ready to  accept the manuscript as it now is.

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have thoroughly improved the manuscript and took into account the comments in the review report. The plots in Figures 4 and 5 make now more sense from a physics viewpoint with high f0F2 values in the local afternoon hours. Also the modeling seems to work better with LT instead of UT.

I have no further comments and recommend to publish the paper,