Automated High-Precision Recognition of Solar Filaments Based on an Improved U2-Net
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsComments on the Quality of English Language
minor editing work will be helpful.
Author Response
Thank you for your pointing out. We agree with this comment. We have finished minor editing work
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper presents an advanced approach for the high-precision recognition of solar filaments, a critical solar activity phenomenon, using an improved U2-Net deep learning model. The study emphasizes the importance of accurately identifying solar filaments in full-disk H-alpha solar images due to their association with significant solar events like solar flares and coronal mass ejections. The researchers processed images from the Chinese H-alpha Solar Explorer, captured in 2023, to generate filament labels. The preprocessing pipeline included several steps such as limb-darkening removal, grayscale transformation, K-means clustering, particle erosion, multiple closing operations, and hole filling. An attention mechanism was integrated into the U2-Net model to enhance its performance on this specialized dataset. The improved Attention U2-Net demonstrated superior performance with an average Accuracy of 0.9987, Precision of 0.8221, Recall of 0.8469, IoU of 0.7139, and F1-score of 0.8323, outperforming other U-net variants.
The paper addresses a crucial need in solar physics, where the automated and precise detection of solar filaments can significantly enhance our understanding and prediction of solar activity, which has broad implications for space weather forecasting. The integration of an attention mechanism into the U2-Net model is a notable innovation, which improves the model's ability to focus on important regions in the solar images, leading to higher accuracy and better performance in filament detection. The authors have meticulously detailed their preprocessing steps, ensuring the dataset is well-prepared for deep learning applications. This attention to detail is critical for achieving high precision in complex image recognition tasks.
Here are some suggestions for possible improvement:
- The paper would benefit from a discussion on the availability of the dataset used in this study. Providing access to the dataset or describing how it can be accessed would be valuable for other researchers looking to replicate or build upon this work.
- While the improved Attention U2-Net shows superior performance, the paper should include a more detailed comparison with baseline models, such as the original U2-Net and other popular semantic segmentation models, under identical conditions. This would highlight the specific advantages brought by the attention mechanism.
- Introduce more recent work on the application of improved U-net and its variants for segmentation tasks, such as: https://doi.org/10.1080/17452759.2024.2325572
- An ablation study could be included to systematically evaluate the impact of each preprocessing step and the attention mechanism on the model's performance. This would provide insights into which components are most critical for the observed improvements.
- Including visual comparisons of the filament segmentation results from the Attention U2-Net and other models could provide a clearer demonstration of the improvements in detection and segmentation quality. This would also help in understanding any remaining challenges in filament detection.
- While the paper focuses on accuracy, a discussion on the potential for real-time or near-real-time application of this model in operational settings would be valuable. This could include considerations of model deployment, inference speed, and integration with existing solar monitoring systems.
Overall, this manuscript presents a significant advancement in the automated detection of solar filaments using deep learning. The integration of an attention mechanism into the U2-Net model yields substantial improvements in performance, as evidenced by the strong results across multiple metrics. The research is timely, relevant, and has the potential for real-world application in solar physics and space weather forecasting.
Comments on the Quality of English LanguageEnglish language used throughout the manuscript is generally good.
Author Response
Comment 1: The paper would benefit from a discussion on the availability of the dataset used in this study. Providing access to the dataset or describing how it can be accessed would be valuable for other researchers looking to replicate or build upon this work.
Respose 1: Thank you for your pointing out. We agree with this comment. We have made our solar image dataset public in manuscript, which can be found in the Data Availability Statement.
Comment 2: While the improved Attention U2-Net shows superior performance, the paper should include a more detailed comparison with baseline models, such as the original U2-Net and other popular semantic segmentation models, under identical conditions. This would highlight the specific advantages brought by the attention mechanism.
Respose 2: Thank you for your pointing out. We agree with this comment. We have added the detailed introduction of Attention U2Net in 3.1.3. as well as the improvements and advantages of Attention U2Net compared to other baseline neural networks, such as U2Net.
Comment 3: Introduce more recent work on the application of improved U-net and its variants for segmentation tasks, such as: https://doi.org/10.1080/17452759.2024.2325572
Respose 3: Thank you for your pointing out. We agree with this comment. As a result, We have added AM-SegNet, the ecent work on the application of improved U-net and its variants for segmentation tasks.
Comment 4: An ablation study could be included to systematically evaluate the impact of each preprocessing step and the attention mechanism on the model's performance. This would provide insights into which components are most critical for the observed improvements.
Respose 4: Thank you for your pointing out. We agree with this comment. We have shown the results of the ablation experiment in Table 1 and Table 3 of the 3.4 results. Table 1 shows four kinds of U-Net neural networks controlling different parameters, and Table 3 shows the performance of four kinds of neural networks in the evaluation function.
Comment 5: Including visual comparisons of the filament segmentation results from the Attention U2-Net and other models could provide a clearer demonstration of the improvements in detection and segmentation quality. This would also help in understanding any remaining challenges in filament detection.
Respose 5: Thank you for your pointing out. We agree with this comment. We have shown the results in Figure 10 of the 3.4 Result.
Comment 6: While the paper focuses on accuracy, a discussion on the potential for real-time or near-real-time application of this model in operational settings would be valuable. This could include considerations of model deployment, inference speed, and integration with existing solar monitoring systems.
Respose 6: Thank you for your pointing out. The focus of this paper is to discuss the accuracy of solar filament detection, and we hope to further discuss the real-time and integrated detection system in future articles.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authorssee attache file
Comments for author File: Comments.pdf
Author Response
Comment 1：The following equations are numbered but are not mentioned in the text (not cited)
Respose 1: Thank you for your pointing out. We agree with this comment. We have added the reference to the equation number in the text on pages 8-9.
Comment 2: In References, Line 289, bibitem [13], substitute / 2017, . / 2017. /
Respose 2: Thank you for your pointing out. We agree with this comment. We have deleted the comma.
Comment 3: In References, Line 308, bibitem [22].Recommendation: To split the line.
Respose 3: Thank you for your pointing out. We agree with this comment. We have splited the line.
Comment 4: Lines: 155-156 Recommendation: substitute / Figure 3 / Fig. 3./
Comment 6: Lines: 115, 116, 163, 164, 165, 185, 215, 218, 219 Recommendation: substitute / Figure X/ Fig. X/
Respose 4: Thank you for your pointing out. We agree with these comments. We have substituted all Figure X to Fig. X in the text.
Comment 5: Page 14, Line 224 two times Recommendation: change: Figure, Indicate the relevant figures.
Respose 5: Thank you for your pointing out. We agree with this comment. We have deleted the two figures in 4.Disccussion in Page 14.
Comment 7: Pages 3 and 4 Recommendation: Re-scale Figures.
Respose 7 : Thank you for your pointing out. We agree with this comment. We have Re-scaled the two Figures in Page 3 and 4.
Comment 8: About Fig. 12 Recommendation: To change Oy in Logarithmic scale.
Respose 8: Thank you for your pointing out. We agree with this comment. We have changed one figure to two figures in Fig.11 and Fig.12. Now it is visible well in a linear scale.
Comment 9: And a few questions at the end.
- What is T N in Table 2 ?
- What is "Validation Loss" or "Loss" in Fig. 12 ?
Respose 9: Thank you for your pointing out. We agree with this comment. We have added the explanation of TN in subsection 3.4 and we have added the explanation of Validation Loss in subsection 3.5.
Author Response File: Author Response.pdf
