Review Reports - Convective–Stratiform Identification Neural Network (CONSTRAINN) for the WIVERN Mission

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article does not explain Figure 6. If it belongs to the U-Net network or an improved version, the figure does not show the skip connection, which should be explained.
Did this paper use data preprocessing methods when inputting data? It should be specified.
In Section 2.1, “Tropical Cyclone Simulation Dataset”, the authors mention mitigating the imbalance problem through loss reweighting and probability thresholds, as described in Section 3. Could you elaborate on this?
In 3.6.1, the author mentions that the task was redefined as a binary classification task and further data mapping processing was performed. Does this contradict the linear mapping rule in 3.2? In the subsequent figure legend, the Target Mask shows multiple color categories. Does this also contradict the binary classification task?
The figure in the text shows the excellent prediction results of the CONSTRAINN-Large network, but the prediction results of the Mini and Medium versions should also be shown to enhance persuasiveness.
5 The training setup should provide detailed information about the experimental setup, such as specific hyperparameter setting and the experimental environment.
The model in this paper is based on U-Net. Could the authors consider comparing it with other models to validate the superiority of the model in this paper?
According to MDIP format requirements, table names should be placed above the table, not below it.

Authors may consider adding subtitles to legends of figures to facilitate descriptions

Author Response

The article does not explain Figure 6. If it belongs to the U-Net network or an improved version, the figure does not show the skip connection, which should be explained.

It belongs to CONSTRAINN-medium. Skip connections were collapsed into the graph, the graph has been subsequently updated.
Did this paper use data preprocessing methods when inputting data? It should be specified.

Very minimal preprocessing has been applied, consisting in capping of the negative reflectivity values and in normalization, as reported in section 3.3
In Section 2.1, “Tropical Cyclone Simulation Dataset”, the authors mention mitigating the imbalance problem through loss reweighting and probability thresholds, as described in Section 3. Could you elaborate on this?

Tha segment has been rewritten. It was referring to using a weighted training loss function and to the postprocessing section computed on the inference output to reformulate the task as a binary classification problem and to consequently compute the thresholded metrics.
In 3.6.1, the author mentions that the task was redefined as a binary classification task and further data mapping processing was performed. Does this contradict the linear mapping rule in 3.2? In the subsequent figure legend, the Target Mask shows multiple color categories. Does this also contradict the binary classification task?

There is no contradiction. The network itself is trained to reconstruct continuous values, obtained through the aforementioned mapping rule. The continuous values obtained from model inference are then postprocessed, as described in now section 3.5, to assign to each pixel a value of either 0 or 1 (or NaN) and then the postprocessed output is treated as a binary classification. Case study figures with six plots show the model output (continuous), while the figures with three plots show the postprocessing results.
The figure in the text shows the excellent prediction results of the CONSTRAINN-Large network, but the prediction results of the Mini and Medium versions should also be shown to enhance persuasiveness.

An additional case study ran on all the three model sizes has been added to the paper.
5 The training setup should provide detailed information about the experimental setup, such as specific hyperparameter setting and the experimental environment.

The training setup section has been updated accordingly.
The model in this paper is based on U-Net. Could the authors consider comparing it with other models to validate the superiority of the model in this paper?

This would go beyond the scope of this publication. Future works section now mention possible extensions and comparison to different deep learning approaches.
According to MDIP format requirements, table names should be placed above the table, not below it.

Tables have been updated and fixed.

Reviewer 2 Report

Comments and Suggestions for Authors

1) The introduction is a bit long. It would have been wise to state the specific purpose of this work from the beginning of the paper, namely to classify convective and statiform zones, then to detail the specificities of the WIVERN instrument and to further emphasize the novelty brought by this study compared to previous work. 2) In 2.1 we understand that you trained your classifier on a dataset from a cyclonic event on October 6, 2024. What physical parameters were captured? With which instrument and how were these data processed to simulate data that could have been obtained by the WIVERN instrument? These last two points are important to judge the relevance of the simulation used for training. Paragraph 3 partly addresses this point but should be more concise before setting out some details of the different data sources. 3) The end of the paper (from paragraph 4 included) is well written and clear. Overall, the paper is interesting in its approach.

Author Response

1) The introduction is a bit long. It would have been wise to state the specific purpose of this work from the beginning of the paper, namely to classify convective and statiform zones, then to detail the specificities of the WIVERN instrument and to further emphasize the novelty brought by this study compared to previous work.

The introduction section has been revised and slightly reduced. It now states the specific purpose of the work from its beginning, following these considerations.

2) In 2.1 we understand that you trained your classifier on a dataset from a cyclonic event on October 6, 2024. What physical parameters were captured? With which instrument and how were these data processed to simulate data that could have been obtained by the WIVERN instrument? These last two points are important to judge the relevance of the simulation used for training. Paragraph 3 partly addresses this point but should be more concise before setting out some details of the different data sources.

We address this by including two new sentences in section 2.1

3) The end of the paper (from paragraph 4 included) is well written and clear. Overall, the paper is interesting in its approach.

Reviewer 3 Report

Comments and Suggestions for Authors

General Comments

This manuscript presents CONSTRAINN, a family of U-Net-based neural networks trained on WRF-driven simulations of WIVERN observations to estimate a continuous convective–stratiform (C/S) index. The topic is highly relevant, particularly in light of upcoming satellite missions such as ESA’s WIVERN. The study is technically sound, and the integration of multiple radar observables (Doppler velocity, reflectivity, and brightness temperature) is well motivated and appropriately implemented. However, several important issues require clarification before the manuscript can be recommended for publication: While the model adopts the established U-Net architecture, the manuscript should more explicitly delineate its novel contributions. Is the innovation primarily in the formulation of the continuous C/S index, the use of simulated Doppler velocity, or the adaptation to W-band radar geometry? The manuscript would benefit from incorporating more recent literature on AI-based radar retrievals—particularly those involving convective classification or Doppler-based diagnostics—to better situate this work within the current research landscape.

Specific Comments

Equation (2): Please elaborate on the rationale behind the choice of 1 m/s and 3 m/s as thresholds for the C/S mask. Are these values grounded in physical considerations, prior literature, or empirical tuning?
Sections 3.3–3.6: These subsections are relatively brief and could be consolidated for clarity. Additionally, since there is no Section 3.6.2, it may be unnecessary to label the existing subsection as 3.6.1.
Figures 7–9: The case studies are insightful, but a discussion of the model's failure modes—such as the occurrence of false-positive halos around convective cores—is needed. Could spatial smoothing, morphological filtering, or other post-processing techniques help mitigate these artifacts?
Conclusion: The authors are encouraged to add a paragraph outlining potential future work on applying CONSTRAINN to real satellite observations (e.g., EarthCARE or WIVERN). This should include discussion of anticipated challenges, such as instrument noise, calibration uncertainties, and domain adaptation between simulated and actual measurements.

Author Response

Equation (2): Please elaborate on the rationale behind the choice of 1 m/s and 3 m/s as thresholds for the C/S mask. Are these values grounded in physical considerations, prior literature, or empirical tuning?

1 m/s is the threshold commonly used in literature to separate stratiform and convective motions, while 3 m/s can be considered a value that corresponds to moderate convection. Thus, the region in between can be considered a region of transition between the two regimes.
Sections 3.3–3.6: These subsections are relatively brief and could be consolidated for clarity. Additionally, since there is no Section 3.6.2, it may be unnecessary to label the existing subsection as 3.6.1.

Sections have been slightly rearranged following these considerations
Figures 7–9: The case studies are insightful, but a discussion of the model's failure modes—such as the occurrence of false-positive halos around convective cores—is needed. Could spatial smoothing, morphological filtering, or other post-processing techniques help mitigate these artifacts?

Convection regions have high spatial variability, and therefore spatial smoothing is not a recommended processing. On the other hand, false-positive halos are considered an acceptable outcome for our considered task, since the main purpose of this classification model is to screen out convective pixels.
Conclusion: The authors are encouraged to add a paragraph outlining potential future work on applying CONSTRAINN to real satellite observations (e.g., EarthCARE or WIVERN). This should include discussion of anticipated challenges, such as instrument noise, calibration uncertainties, and domain adaptation between simulated and actual measurements.

The conclusions section has been updated following these considerations

The General Comments have been considered and the paper has been revised accordingly

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised paper addressed my concerns regarding the first version.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript has been revised in response to the previous comments and is recommended for acceptance.