U-Net Based Forecasting of Storm-Time Total Electron Content over North Africa Using Assimilation of GNSS Observation into Global Ionospheric Maps

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript titled “U-Net–based forecasting of storm-time TEC over North Africa using assimilation of GNSS observations into the Global Ionospheric Map” presents the development of a nowcasting model for ionospheric total electron content (TEC) over the African region using a U-Net–based machine learning approach. The methodology is sound, and the results demonstrate improved performance compared to existing models through statistical evaluation. There are no major concerns but several minor points listed below.

1. Please define total electron content (TEC) in the title
2. TEC should be defined at the first appearance in the abstract. (e.g Line 28 : “This study presents a U-Net deep learning model of total electron content (TEC) …”)
3. The study considers 33 storm events, with 25 used for training the U-Net model. Please clarify whether this number of events is sufficient for training a deep learning model. It would be helpful to state the total number of training samples (e.g., number of 4-hour data segments per day) used for model training to better address concerns regarding sample size.
4. Specify the forecast lead time of the predicted models shown in Figure 7. E.g. indicate that the results correspond to a one-hour lead time in the figure caption.
5. Since recurrent neural network (RNN) architectures such as LSTM are often well suited for time-series prediction, please include a brief discussion comparing CNN-based U-Net models with RNN-based approaches and outline plans for exploring alternative ML techniques in future work.

Author Response

 

Comments and Suggestions for Authors

The manuscript titled “U-Net–based forecasting of storm-time TEC over North Africa using assimilation of GNSS observations into the Global Ionospheric Map” presents the development of a nowcasting model for ionospheric total electron content (TEC) over the African region using a U-Net–based machine learning approach. The methodology is sound, and the results demonstrate improved performance compared to existing models through statistical evaluation. There are no major concerns but several minor points listed below.

Re: Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions highlighted in track changes in the re-submitted manuscript attched. In addition, further modifications were undertaken to accommodate the valuable comments and requirements specified by other reviewers.

1- Please define total electron content (TEC) in the title

Re: Thank you very much for your suggestion. A clear full name instead of the TEC abbreviation has been added in the title.

2- TEC should be defined at the first appearance in the abstract. (e.g Line 28 : “This study presents a U-Net deep learning model of total electron content (TEC) …”)

Re: Thank you so much for your invaluable suggestion. A full definition of the TEC is added to line 28, besides we have revised the whole manuscript for any other modifications regarding it. Like caption of Figure 12, and also in the discussion section…etc

3- The study considers 33 storm events, with 25 used for training the U-Net model. Please clarify whether this number of events is sufficient for training a deep learning model. It would be helpful to state the total number of training samples (e.g., number of 4-hour data segments per day) used for model training to better address concerns regarding sample size.

Re: Thank you very much for your invaluable comment on this point. Providing further clarification on the number of samples used for training is indeed important for the reader. Worth noting that, for each geomagnetic storm, we selected a 7-day interval consisting of three days before and three days after the time of maximum depression in the Dst index, in addition to the day of maximum depression itself. The GIM-TEC data have a temporal resolution of 1 hour. Consequently, the number of samples per storm is 24 × 7 = 168 TEC maps. Based on this selection, the total number of training samples (25 storms) amount to 4200 TEC maps.

Also we have clarified the reason behind demonstrating the figure as 4-hour however the 1-hour time resolution of GIM data in section 3.2 as follows “ interval as follows. Worth noting the ground truth data is 1-hour time resolution, however the 4-hour time interval shown in Figures 7-9 is presented solely for visualization purposes, to illustrate the temporal evolution of TEC over the North African region. This reduced interval was chosen to limit the number of displayed maps while preserving the key temporal characteristics.

4- Specify the forecast lead time of the predicted models shown in Figure 7. E.g. indicate that the results correspond to a one-hour lead time in the figure caption.

Re: We are very grateful to the reviewer for prompting us to revise this point. It is indeed important to clearly emphasize in the figure that the U-Net model performs a 1-hour lead prediction. Accordingly, we have added a clarification in section 3.2 confirming that the U-Net output corresponds to a 1-hour lead TEC map prediction.

5- Since recurrent neural network (RNN) architectures such as LSTM are often well suited for time-series prediction, please include a brief discussion comparing CNN-based U-Net models with RNN-based approaches and outline plans for exploring alternative ML techniques in future work.

Response: We thank the reviewer for this valuable suggestion. We agree that while LSTMs have traditionally been the standard for temporal modeling, CNN-based architectures like the U-Net offer distinct advantages in terms of computational efficiency and multi-scale feature extraction. In the introduction section we have added a brief discussion comparing RNN and CNN-based U-Net which compares these two paradigms and and in the conclusion section we have outlined our future research directions regarding hybrid and Transformer-based architectures:

The section that has been added in the introduction section as follows: Worth noting that implementation of the CNN-based U-Net architectures have emerged as powerful alternatives for time-series tasks [2]. Its Net architecture utilizes convolutional filters to capture local and global patterns simultaneously across the en-tire input window. This is particularly effective for "TEC image-like" time-series data where multi-scale features (e.g., short-term spikes vs. long-term trends) are critical [3]. In contrast, RNN processes data sequentially, which can lead to vanishing gradient is-sues over very long sequences despite their internal memory cells. Because U-Nets are fully convolutional, they allow for massive parallelization during training, often resulting in significantly faster convergence compared to the sequential nature of LSTMs [4]. Finally, the hierarchical structure of the U-Net’s encoder-decoder paths allows it to establish a large receptive field, enabling the model to "see" a broad temporal context without the recursive overhead of an RNN.

In the conclusion we have added this sentence for future promising studies“ Besides, it is recommended to investigate Transformer-based models, which utilize self-attention mechanisms to weigh the importance of different time steps globally, in handling complex, non-linear dependencies in high-frequency data.”

 

 

 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

I have reviewed the revised manuscript titled "Geomagnetic Storm-Time Total Electron Content Modeling over North Africa: U-Net Architecture Validated Against AfriTEC". The authors have presented a well-designed hybrid framework combining data assimilation and U-Net-based deep learning for improved TEC forecasting over North Africa during geomagnetic storms. The comparative analysis with both IRI-2020 and AfriTEC models is comprehensive, and the results are convincing. The manuscript is suitable for publication in its current form after addressing the following minor revisions to enhance clarity, depth, and impact.

Comments and Suggestions for the Authors:

1. Clarification of Data Assimilation Methodology: In Section 2.2, the two-step assimilation method is described, involving a "fixed contribution factor (50%)" for the second adjustment. Please provide a brief justification for selecting the 50% value (e.g., based on preliminary tests, empirical balance between bias correction and overfitting prevention). This will enhance the transparency and reproducibility of the assimilation scheme.

2. Consistent and Careful Use of "Ground Truth": The assimilated GIM product is consistently referred to as "ground truth." While it serves as a valid reference for model evaluation, it is technically a fusion product. It is recommended to add a clarifying note upon its first use (e.g., "the assimilated GIM data, hereafter referred to as 'ground truth' for model evaluation") and maintain this terminology consistently to avoid confusion with raw GNSS observations.

3. Deepening the Discussion of Research Significance: The introduction and conclusion mention the challenge of sparse GNSS coverage in Africa. Please expand on the practical implications of your proposed framework (assimilation + deep learning) for operational space weather forecasting. Specifically, discuss its potential for near-real-time application, its contribution to mitigating GNSS positioning errors, and its significance for building space weather service capabilities in Africa and other data-sparse regions.

4. More In-Depth Discussion on Error Sources: Section 3.1 attributes residual errors after assimilation to "localized ionospheric structures and limited station density." Please elaborate further: given your linear interpolation-based assimilation scheme, which specific ionospheric phenomena (e.g., plasma bubbles, traveling ionospheric disturbances) are most challenging to capture accurately? This will help better define the current method's limitations.

5. Improved Readability of Figures: Some multi-panel figures (e.g., Figures 5, 6) contain small labels and may be difficult to interpret in grayscale. Please ensure all figures are legible when printed in black and white. Also, consider refining the captions for key comparative figures (e.g., Figures 7, 8, 9) to more directly highlight the most critical differences between the U-Net predictions and the IRI/AfriTEC outputs.

6. Extended Discussion on AfriTEC's Performance Degradation: Section 4 correctly points out that AfriTEC's training data (2000-2018) does not represent the strong activity of the current Solar Cycle 25. This is a crucial observation. Please briefly discuss the broader implication: What challenge does this pose for the general applicability of regionally-tuned empirical models based on historical data? What does this suggest for developing future ionospheric models with longer-term adaptability?

7. More Specific "Future Work" Section: The direction of "incorporating additional space-based observational datasets" mentioned in the conclusion is valuable. Please provide one or two concrete examples of the most relevant and accessible data sources (e.g., in-situ electron density from SWARM satellites, or radio occultation profiles from COSMIC-2) and briefly indicate how they could enhance model performance (e.g., vertical structure, global coverage).

8. Brief Note on Real-Time Applicability Potential: Given the forecasting nature of this study, it would be beneficial to add a sentence or two in the Discussion or Conclusion regarding the practical challenges for real-time or near-real-time operation. A comment on the computational demand, data latency requirements, or operational feasibility of the proposed U-Net model would increase its relevance for potential operational users.

9. Minor Language and Formatting Edits:

(a) In the abstract, "GPS-sparsing" should be corrected to "GPS-sparse" or "in regions with sparse GPS coverage."

(b) Ensure consistent formatting of "U-Net" throughout the manuscript (currently "U-Net," "U- Net," and "U-net" are used interchangeably). The standard "U-Net" is recommended.

(c) Please perform a final check to ensure all in-text citations and the reference list conform strictly to the journal's style guide.

Author Response

Comments and Suggestions for Authors

I have reviewed the revised manuscript titled "Geomagnetic Storm-Time Total Electron Content Modeling over North Africa: U-Net Architecture Validated Against AfriTEC". The authors have presented a well-designed hybrid framework combining data assimilation and U-Net-based deep learning for improved TEC forecasting over North Africa during geomagnetic storms. The comparative analysis with both IRI-2020 and AfriTEC models is comprehensive, and the results are convincing. The manuscript is suitable for publication in its current form after addressing the following minor revisions to enhance clarity, depth, and impact.

Re: Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions highlighted in track changes of attcahed manuscript. In addition, further modifications were undertaken to accommodate the valuable comments and requirements specified by other reviewers.

Comments and Suggestions for the Authors:

1. Clarification of Data Assimilation Methodology: In Section 2.2, the two-step assimilation method is described, involving a "fixed contribution factor (50%)" for the second adjustment. Please provide a brief justification for selecting the 50% value (e.g., based on preliminary tests, empirical balance between bias correction and overfitting prevention). This will enhance the transparency and reproducibility of the assimilation scheme.

Re: Thank you for this comment regarding the choice of the 50% contribution factor in the second adjustment of the two-step assimilation method. A brief justification of this selection has been incorporated into the revised manuscript throughout the second paragraph of section 2.2.     “ ….an error correction map. This interpolated error map was scaled by a fixed contribution factor of 50%, a value chosen based on preliminary sensitivity tests that balanced the correction of spatially residual errors against the risk of overfitting to the uneven GNSS station distribution. Specifically, lower values failed to adequately correct regional biases, whereas higher values occasionally accumulate errors in sparsely observed regions. Therefore, a 50% contribution factor was identified as the most robust and balanced choice. The scaled error map is then added to ……”.

 

2. Consistent and Careful Use of "Ground Truth": The assimilated GIM product is consistently referred to as "ground truth." While it serves as a valid reference for model evaluation, it is technically a fusion product. It is recommended to add a clarifying note upon its first use (e.g., "the assimilated GIM data, hereafter referred to as 'ground truth' for model evaluation") and maintain this terminology consistently to avoid confusion with raw GNSS observations.

Re: Thank you so much for stressing this point for more clarification. By the end of section 3.1 we have mentioned that from this point onwards we will refer to the assimilated TEC map as the ground truth. Subsequently we have replaced all assimilated TEC maps with the ground truth.

3. Deepening the Discussion of Research Significance: The introduction and conclusion mention the challenge of sparse GNSS coverage in Africa. Please expand on the practical implications of your proposed framework (assimilation + deep learning) for operational space weather forecasting. Specifically, discuss its potential for near-real-time application, its contribution to mitigating GNSS positioning errors, and its significance for building space weather service capabilities in Africa and other data-sparse regions.

Re: Thank you very much for this comment.

Several studies have incorporated additional datasets, such as COSMIC radio occultation measurements, to improve GIM-derived TEC estimates (e.g., Okoh et al., 2019, 2020). More recently, Okoh et al. (2025) demonstrated the strong potential of calibrated COSMIC data as a complementary dataset to ground-based GNSS observations for operational space-weather monitoring. Therefore, both calibrated COSMIC data/ground-based GNSS observations are required to produce corrected (assimilated) GIM products. These corrected GIMs are particularly important for deep learning models, such as our U-Net model, to reduce spatial errors arising from interpolation. So we have given additional details for the efforts achieved for finding more data sources besides the GNSS TEC data over the African as demonstrated in the 5th paragraph of the Introduction section.

Additionally, by the end of the discussion section we have expanded the significance of the assimilation+deep U-Net the for developing a space weather service capability in Africa as follows: These preliminary results indicate that the proposed assimilation-deep learning framework can better capture spatio-temporal variations that are poorly captured by either the AfriTEC or the IRI models. Worth noting that the proposed framework demonstrates favorable computational characteristics on a standard GPU hardware (16 GB, Tesla-P100) with  an inference time 0.003517 seconds per sample (TEC map in our study). This indicates that the model is capable of fast predictions, making it suitable for practical deployment in real-world applications. However, its update frequency is mainly governed by the latency of the input drivers. Consequently, the primary operational challenge is likely to be data latency rather than computational demand, as limited GNSS coverage across Africa can delay data availability.

 

4. More In-Depth Discussion on Error Sources: Section 3.1 attributes residual errors after assimilation to "localized ionospheric structures and limited station density." Please elaborate further: given your linear interpolation-based assimilation scheme, which specific ionospheric phenomena (e.g., plasma bubbles, traveling ionospheric disturbances) are most challenging to capture accurately? This will help better define the current method's limitations.

Re: Thank you for this important comment. Following the reviewer's suggestion, we have revised Section 3.1 to explicitly clarify the physical sources of residual errors. The added text in the second paragraph of section 3.1 specifies that small-scale to medium-scale ionospheric phenomena, including equatorial plasma bubbles, traveling ionospheric disturbances, and sharp storm-time TEC gradients, are difficult to resolve using a generic interpolation-based assimilation scheme under sparse GNSS station coverage.

In particular, small-scale to medium-scale ionospheric phenomena such as equatorial plasma bubbles, traveling ionospheric disturbances, and sharp storm-time TEC gradients are difficult to capture accurately using a generic interpolation-based assimilation scheme. These structures exhibit rapid temporal evolution and strong spatial gradients that are not well resolved by the sparse GNSS station distribution over Africa and are therefore partially smoothed during interpolation. As a result, localized TEC variations may persist as residual errors even after assimilation.

 

5. Improved Readability of Figures: Some multi-panel figures (e.g., Figures 5, 6) contain small labels and may be difficult to interpret in grayscale. Please ensure all figures are legible when printed in black and white. Also, consider refining the captions for key comparative figures (e.g., Figures 7, 8, 9) to more directly highlight the most critical differences between the U-Net predictions and the IRI/AfriTEC outputs.

Re: Thank you so much for stressing enhancing the figure. We have improved the quality Figures 5 and 6 to be more readable, also  we have re-organised figure 7 as the TECU of the color bar was missing. Finally, we have re-adjusted the captions of the figures to be concise and convey the meaning and highlight the most critical differences.

 

6. Extended Discussion on AfriTEC's Performance Degradation: Section 4 correctly points out that AfriTEC's training data (2000-2018) does not represent the strong activity of the current Solar Cycle 25. This is a crucial observation. Please briefly discuss the broader implication: What challenge does this pose for the general applicability of regionally-tuned empirical models based on historical data? What does this suggest for developing future ionospheric models with longer-term adaptability?

Re: Thank you so much for your comment. The current model is designed to capture the spatio-temporal variations of ionospheric TEC under intense storm conditions. This objective differs from that of AfriTEC and IRI, which are trained using both quiet and storm-time TEC data. Mixing quiet and storm conditions during training may introduce bias in storm-time predictions. Both AfriTEC and IRI continue to show the presence of double crests, even when these features are degraded in the ground-truth data during storms as shown in Figure 7 and 9. The U-Net model, however, successfully captures this storm-time degradation in TEC. We therefore suggest that the relatively low performance of AfriTEC may be due to its training period (2000–2018), which did not include extreme events such as the May 2024 storm. Alternatively, the improved performance of the proposed U-Net model may result from its focus on storm-time TEC only, allowing it to avoid bias introduced by quiet-time conditions. Subsequently few lines have been added to the discussion section stressing this biassing problem as follows: “ Besides, our U-Net model, which is trained only on storm-time TEC data, demonstrates non-biased quiet time behavior compared with the AfriTEC and IRI models. This quiet-time biasing behavior has been mentioned earlier in the study of  (doi.org/10.1029/2021SW002741). Therefore, the AfriTEC  and IRI modelled data shown in Figures 7-9 still show the two ionospheric crests around the equator even during severely disturbed conditions, although the crests are completely degraded in the observed data.”

 

7. More Specific "Future Work" Section: The direction of "incorporating additional space-based observational datasets" mentioned in the conclusion is valuable. Please provide one or two concrete examples of the most relevant and accessible data sources (e.g., in-situ electron density from SWARM satellites, or radio occultation profiles from COSMIC-2) and briefly indicate how they could enhance model performance (e.g., vertical structure, global coverage).

Re:  Thank you so much for raising such a comment. By the end of the conclusion we have added these few lines “..... For instance, incorporating space-based data such as COSMIC-2 would improve such study due to its global coverage and enhanced vertical resolution, making it highly suitable for studying ionospheric structure and variability over North Africa, where ground-based observatory coverage is sparse.”

 

8. Brief Note on Real-Time Applicability Potential: Given the forecasting nature of this study, it would be beneficial to add a sentence or two in the Discussion or Conclusion regarding the practical challenges for real-time or near-real-time operation. A comment on the computational demand, data latency requirements, or operational feasibility of the proposed U-Net model would increase its relevance for potential operational users.

Re: We sincerely thank the reviewer for highlighting the importance of clarifying the real-time and near-real-time applicability of the proposed forecasting framework. related to computational demand, data latency, and operational feasibility. While the U-Net model is trained offline, its deployment phase is computationally efficient and involves a single forward pass without iterative optimization. This makes the framework well suited for near-real-time operation once trained. The required inputs-hourly GIM TEC products and routinely available space-weather indices-are already provided with low latency by existing operational services, supporting practical implementation in monitoring and forecasting environments. This clarification enhances the relevance of the proposed approach for potential operational users. Also, the inference time of the model has been reported in an experiment and found to be 0.0003517 seconds. That was earlier explained through comment 3.

In response, we have briefly referred to that issue  in the Conclusion and future work section as follows: Moreover, given the moderate spatial resolution and one-hour-ahead prediction horizon, highlighting its potential applicability for operational ionospheric monitoring and space-weather forecasting systems.

 

9. Minor Language and Formatting Edits:

(a) In the abstract, "GPS-sparsing" should be corrected to "GPS-sparse" or "in regions with sparse GPS coverage."

Re: Thank you so much for your constructive comment. We have modified this sentence in the frame of your comment.

(b) Ensure consistent formatting of "U-Net" throughout the manuscript (currently "U-Net," "U- Net," and "U-net" are used interchangeably). The standard "U-Net" is recommended.

Re: Thank you so much for your constructive comment. We have modified this sentence in the frame of your comment to be U-Net. We have made sure that it is consistent throughout the manuscript in the format U-Net.

(c) Please perform a final check to ensure all in-text citations and the reference list conform strictly to the journal's style guide.

Thank you for this helpful suggestion. We have rechecked the references and made sure it adapts the MDPI 's reference style.

 

Author Response File: Author Response.docx