A Deep Learning-Based Echo Extrapolation Method by Fusing Radar Mosaic and RMAPS-NOW Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsAuthors trend to present a deep learning-based echo extrapolation method in this manuscript. There exists 5 major issues. Comments are listed as follows.
Major issue 1: Many references are too old for citing. 6 references are published before 1990. Reference 8 has error.
Major issue 2: Generator network is not presented. Why?
Major issue 3: Why not presented relevant comparison methods? For example, methods in references [32],[33],[34],[36].
Major issue 4: Authors should summarize motivation and contribution of this manuscript in the part of introduction.
Major issue 5 : In present form, the novelty is not enough for Remote Sensing. Authors should dig and present more novel points for reviewers.
Author Response
Major Issue 1: Many references are too outdated to be cited. Six references were published before 1990. Reference 8 contains errors.
Response: Thank you for your valuable suggestions regarding the reference section. In response to the issues you pointed out, we have made the following systematic revisions for the problems of "too outdated references" and "errors in Reference [8]": References [4], [5], [6], [7], [8], and [9] have been replaced with high-impact articles in recent years on cross-correlation methods, centroid tracking methods, and optical flow methods. Through updating outdated references and correcting citation errors, the academic timeliness and accuracy of the references have been significantly improved. Thank you again for your meticulous review. We believe these adjustments will enhance the credibility and academic rigor of the paper.
Major Issue 2: The generator network is not shown. Why?
Response: Thank you for your suggestion! Regarding the issue that the generator network was not shown, the generator in this paper adopts the mature architecture from the author's previous literature "Research on Radar Echo Extrapolation Based on ConvLSTM Integrated with RMAPS-NOW Data", so it is cited as a convention (see Reference [37] on page 7 of the main text). A diagram of the generator architecture has been added to page 7 of the revised version, and the details of this architecture have been described. Thank you again for your meticulous review. The description of the generator network in the revised version has formed a complete technical chain with other modules in the method section (such as the discriminator and loss function design), which helps readers comprehensively understand the model design logic. Please feel free to let us know if further adjustments are needed.
Major Issue 3: Why are no relevant comparison methods proposed? For example, the methods in References [32], [33], [34], and [36].
Response: Thank you for your suggestion! This paper focuses on the research on the improvement of radar echo extrapolation by fusing RMAPS-NOW weather background data with the DCGAN framework, and the effectiveness of the innovative method is verified through baseline experiments with the baseline of no weather background data and the traditional GAN architecture. The methods in References [32, 33, 34, 36] have fundamental differences in technical approaches and data usage from this study, and some models are difficult to reproduce under the same experimental conditions, so they are not included in the comparison. The revised version has supplemented the explanation in the discussion section and added comparative experiments with classic generative models to enhance the persuasiveness of the method.
Major Issue 4: The authors should summarize the motivation and contributions of this manuscript in the introduction section.
Response: Thank you for your suggestion! We have added a new paragraph at the end of the introduction to systematically elaborate on the research motivation (solving the problems of traditional convolution blur, data detail loss, and lack of physical constraints) and three major contributions (designing a fused encoder-decoder framework, introducing RMAPS-NOW weather background data, and proposing a time-series and strong echo adaptive weighting strategy). The structure of the introduction has been optimized through the "problem-motivation-contribution" logical loop to ensure the clear presentation of the research value.
Major Issue 5: In its current form, the novelty is insufficient for remote sensing. The authors should explore and provide more novel perspectives for the reviewers.
Response: Thank you for the expert feedback. The innovations of this paper include:
1.Introducing weather background information in radar echo extrapolation, which significantly improves extrapolation accuracy compared to models without such information.
2.Training different extrapolation models for varying extrapolation times to avoid blurring issues caused by single models as extrapolation time increases.
3.Designing a custom loss function。
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper proposes a novel Multi-channel Radar echo extrapolation Deep Convolutional Generative Adversarial Network (MR-DCGAN) that integrates radar mosaic data with RMAPS-NOW physical variables, addressing the limitations of traditional convolutional networks in radar echo extrapolation. The use of GANs to mitigate blurring effects is a notable contribution. The idea of this paper is easy to be followed. My comments are as follows:
1) The map in Figure 1 should be redrawn. Please add the north arrow and scale bar. In addition, give the names of the important cities in Hebei Province, for example, using Zhangjiakou instead of the "Zhangbei" Count, using Tianjin instead of the "Tanggu" Bay.
2) Correct some errors, for example, "Figure 101".
3) Provide figures (Figure 9/11/13) with much higher quality.
4) The paper mentions computational constraints but does not quantify the computational cost of training 20 independent models for 120-minute extrapolation. This raises concerns about scalability and real-time applicability. The trade-off between model complexity and computational resources should be discussed in greater detail, especially for operational deployment.
5) While the study compares MR-DCGAN with ConvLSTM and U-Net, it omits comparisons with other state-of-the-art GAN-based or transformer-based models. This limits the reader's understanding of how the proposed method stands against broader alternatives.
6) The absence of ablation studies to isolate the contributions of individual components (e.g., the role of the custom loss function vs. the GAN architecture) weakens the methodological justification.
7) The reliance on CSI, POD, and FAR may not fully capture the spatial and temporal coherence of predicted echoes. Additional metrics like Structural Similarity Index (SSIM) or Mean Squared Error (MSE) could provide a more comprehensive assessment of image quality.
8) Some key implementation details (e.g., hyperparameters for the discriminator/generator, architecture specifics of ConvLSTM layers) are insufficiently described.
Author Response
Major Issue 1: The map in Figure 1 should be redrawn. Please add a compass and a scale. In addition, give the names of important cities in Hebei Province, for example, use "Zhangjiakou" instead of "Zhangbei" County, and use "Tianjin" instead of "Tanggu" Bay.
Response: Thank you for your suggestion! We have redrawn Figure 1, added a black-and-white compass and scale in the upper right corner, and modified the city names to ensure the accuracy and compliance of geographical information.
Major Issue 2: Correct some errors, such as "Figure 101".
Response: Thank you for your correction! Numbering errors such as "Figure 101" in the text have all been corrected, and all chart references strictly correspond to the actual content. The revised version has been rechecked to ensure there are no errors.
Major Issue 3: Provide higher-quality charts (Figures 9/11/13).
Response: Thank you for your correction! Higher-quality figures have been provided for the indicated figures in this paper as required.
Major Issue 4: The paper mentions computational constraints but does not quantify the computational cost of training 20 independent models for 120-minute extrapolation. This raises concerns about scalability and real-time applicability. The trade-off between model complexity and computational resources should be discussed in more detail, especially for deployment.
Response: For deep learning, model training is time-consuming, but once trained, the deployed model requires minimal computing power as all parameters are pre-fitted. The 20 independent models consider the temporal evolution of weather systems and avoid echo blurring issues in single-model extrapolation.
Major Issue 5: Although this study compares MR-DCGAN with ConvLSTM and U-Net, it ignores the comparison with other state-of-the-art GAN-based or transformer-based models. This limits readers' understanding of how the proposed method compares with a wider range of alternatives.
Response: Thank you for the expert suggestion. Given the rapid evolution of deep learning algorithms, this study primarily aims to: 1) demonstrate that multiple models outperform single models in accuracy; 2) prove that introducing weather background information significantly improves echo extrapolation compared to models without it. It is foreseeable that any other model (GAN/Transformer-based) incorporating weather background information would similarly enhance extrapolation performance.
Major Issue 6: The lack of ablation studies to isolate the contributions of individual components (such as the role of the custom loss function versus the GAN architecture) weakens the methodological rationality.
Response: Thank you for the comments raised by the reviewer! Regarding the issue raised by the reviewer that the lack of ablation studies isolates the contributions of individual components (such as the role of the custom loss function and the GAN architecture), it should be noted that this paper focuses on the engineering practicality verification of the MR-DCGAN architecture and the multi-source data fusion overall scheme, and the custom loss function and the GAN architecture are the organic whole of collaborative design. Split verification may cut their internal relevance, which is in line with the common paradigm of exploratory research in the field. Deepening work such as modular ablation experiments and cross-dataset verification has been included in the follow-up research plan.
Major Issue 7: The reliance on CSI, POD, and FAR may not fully capture the spatial and temporal coherence of the predicted echo. Other metrics such as the Structural Similarity Index (SSIM) or Mean Squared Error (MSE) can provide a more comprehensive image quality assessment.
Response: Thank you for your suggestion! We have recognized the insufficiency of CSI, POD, and FAR in evaluating the spatiotemporal coherence of echoes. Therefore, the Structural Similarity Index (SSIM) has been added as a supplementary metric in the revised version. By quantifying the brightness, contrast, and structural similarity between the predicted image and the real echo, the ability of the model to maintain the spatial details of the echo is comprehensively evaluated, and the SSIM change curve of 20 time steps is given in the experiment section. The results show that the model in this study significantly outperforms the ConvLSTM and Unet models in echo structure retention, forming a dual evaluation system of "classification accuracy-structural fidelity".
Major Issue 8: Some key implementation details (such as the hyperparameters of the discriminator/generator and the architectural details of the ConvLSTM layer) are not fully described.
Response: Thank you for your suggestion! We have supplemented the key implementation details in the revised version, including annotating the parameters of the ConvLSTM layer in the generator MR-ConvLSTM model structure diagram, giving the hyperparameters in Section 3.4 Model Training, and adding the mathematical expression of the ConvLSTM gating mechanism in Section 3.1 to ensure the completeness and reprodu
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsAfter revision, this manuscript is improved. There still exists several issues.
Major issue 1: Authors should cite more references which are published within 5 years.
Major issue 2: In response, authors state that ‘A diagram of the generator architecture has been added to page 7 of the revised version’. But, page 7 of the revised version does not contain the generator architecture.
Major issue 3: Authors should compare the proposed method with relevant method which are published within 5 years. Comparing with classic models are not enough.
Author Response
Major Issue 1: Authors should cite more references which are published within 5 years.
Response:
We appreciate the reviewer's suggestion regarding the timeliness of references. We have systematically updated the literature, adding highly relevant studies published between 2020 and 2025 in the introduction section. These additions cover the latest applications of deep learning in radar echo extrapolation and cutting-edge advances in data assimilation technologies. All newly added references were screened based on criteria of direct relevance to the research topic, publication in authoritative field journals, and inclusion of key conclusions. The specific revisions have been marked in the reference section.
Major Issue 2: In response, authors state that ‘A diagram of the generator architecture has been added to page 7 of the revised version’. But, page 7 of the revised version does not contain the generator architecture.
Response:
We appreciate the reviewer for pointing out the oversight in our response! The "page 7" mentioned previously was a typo. The generator architecture diagram has actually been added to Figure 4 on page 10 of the revised version. We deeply apologize for the error in our response and will strictly verify the revision locations to avoid similar issues in the future. For reviewing the specific diagram or textual explanations, please refer to the corresponding page of the revised manuscript.
Major Issue 3: Authors should compare the proposed method with relevant method which are published within 5 years. Comparing with classic models are not enough.
Response:
We appreciate the reviewer's suggestion regarding the adequacy of method comparisons. In the revised experimental comparison section, we have conducted a same-scenario comparison between the proposed MR-DCGAN model and the MR-ConvLSTM model published in 2024. Both models use the same physical quantity inputs and are tested based on the Cangzhou radar echo test set data. The results show that the MR-DCGAN (u, v+rh) model outperforms the MR-ConvLSTM (u, v+rh) model in the 0-2 hour extrapolation time frame.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have made necessary changes, I have no more questions.
Author Response
Thank you very much for your meticulous review and professional suggestions on our manuscript. Your feedback is extremely valuable, as it not only helped us identify the shortcomings in the paper but also provided clear directions for its further improvement.Once again, we appreciate your precious comments despite your busy schedule, which are crucial for enhancing the quality of our work.
Round 3
Reviewer 1 Report
Comments and Suggestions for AuthorsAuthors have answered all questions reviewer concerns
