Generative Algorithms for Wildfire Progression Reconstruction from Multi-Modal Satellite Active Fire Measurements and Terrain Height

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1.  P1–P2:The introduction provides a clear background and motivation for the study but lacks a clear justification for why the cWGAN was chosen over other generative models (such as Diffusion Models).
2.  P7:Regarding the data augmentation mentioning "each solution augmented 25 times," it is recommended to clarify whether this number was validated through ablation studies or if there is a risk of overfitting.
3.  P8–P9:The steps for constructing the observation operator are reasonable. However, for step 8, the uniform distribution U(0,2) for the "ignition time error" – is this based on actual GOES error statistics? Citing relevant error analysis studies would strengthen the persuasiveness.
4.  P10–P11:The network architecture is clearly described, but it does not mention whether ablation studies were conducted for hyperparameter tuning (e.g., learning rate, batch size). It is recommended to supplement the rationale behind the model structure choices.
5.  P12–P15:The experimental section includes rich case studies with good visualization. However, the prediction deviation is significant for some fires (e.g., Tennant) in Figure 5. It is recommended to provide a more in-depth analysis of the causes in the discussion.
6.  P19: The comparative analysis with Shaddy et al. (2021) is thorough, but it does not specify whether the previous method was replicated under the same hardware/environment. Adding this clarification would ensure comparability.
7.  P20–P21:The experimental design for analyzing terrain influence is clever. However, the conclusion in Figure 8 that "97% of pixels have differences less than 30 minutes" seems slightly general. It is recommended to further analyze the spatial distribution of these pixels (e.g., whether they are concentrated at the fire periphery).
8.  P24–P25: The discussion of the method's limitations in the conclusion is sincere, but computational efficiency (e.g., inference time) is not mentioned, which is important for practical deployment. It is recommended to add this.
9.  P25: Future work mentions "diffusion models." It is recommended to briefly explain their potential advantages compared to cWGAN.1.  **P1–P2:** The introduction provides a clear background and motivation for the study but lacks a clear justification for why the cWGAN was chosen over other generative models (such as Diffusion Models).
2.  P7: Regarding the data augmentation mentioning "each solution augmented 25 times," it is recommended to clarify whether this number was validated through ablation studies or if there is a risk of overfitting.
3.  P8–P9:The steps for constructing the observation operator are reasonable. However, for step 8, the uniform distribution U(0,2) for the "ignition time error" – is this based on actual GOES error statistics? Citing relevant error analysis studies would strengthen the persuasiveness.
4.  P10–P11:The network architecture is clearly described, but it does not mention whether ablation studies were conducted for hyperparameter tuning (e.g., learning rate, batch size). It is recommended to supplement the rationale behind the model structure choices.
5.  **P12–P15:** The experimental section includes rich case studies with good visualization. However, the prediction deviation is significant for some fires (e.g., Tennant) in Figure 5. It is recommended to provide a more in-depth analysis of the causes in the discussion.
6.  **P19:** The comparative analysis with Shaddy et al. (2021) is thorough, but it does not specify whether the previous method was replicated under the same hardware/environment. Adding this clarification would ensure comparability.
7.  P20–P21: The experimental design for analyzing terrain influence is clever. However, the conclusion in Figure 8 that "97% of pixels have differences less than 30 minutes" seems slightly general. It is recommended to further analyze the spatial distribution of these pixels (e.g., whether they are concentrated at the fire periphery).
8.  P24–P25:The discussion of the method's limitations in the conclusion is sincere, but computational efficiency (e.g., inference time) is not mentioned, which is important for practical deployment. It is recommended to add this.
9.  P25: Future work mentions "diffusion models." It is recommended to briefly explain their potential advantages compared to cWGAN.

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

 

General Comments

 

This manuscript introduces an innovative application of a conditional 

Wasserstein Generative Adversarial Network (cWGAN) for reconstructing 

wildfire progression, specifically fire arrival times, through the assimilation of 

multi-modal satellite data—including VIIRS active fire detections and GOES-

derived ignition times—combined with terrain elevation. The core 

methodological approach involves training a physics-aware deep learning 

model on high-fidelity simulations generated by the coupled atmosphere–fire 

model WRF-SFIRE. This represents a meaningful and timely integration of 

physically based modeling with data-driven techniques, holding clear relevance 

for advancing wildfire state estimation and operational data assimilation.

 

The research addresses a topic of considerable importance for enhancing 

wildfire monitoring and predictive capabilities. The proposed framework—which 

merges multi-source remote sensing inputs within a cWGAN architecture 

trained on physically consistent simulations—is methodologically sound and 

demonstrates significant potential. Nevertheless, the current presentation of 

the scientific contribution does not sufficiently articulate the distinct novelty 

and specific advances beyond foundational prior work, notably that of Shaddy 

et al. The narrative tends to prioritize a description of methodological steps 

over the formulation of a clear, testable hypothesis and a compelling statement 

of intellectual merit. To strengthen the manuscript's impact, a more focused and 

explicit framing of its core innovations is needed, coupled with improved clarity 

in the description of methodologically intricate components.

 

To solidify its impact, the manuscript requires a more focused and explicit 

framing of its core innovations alongside enhanced clarity in describing complex 

methodological components. For example:

 

a) The central claim of novelty could be concisely articulated as a numbered list 

of key contributions—such as a novel terrain-conditioned inference framework 

and a training paradigm grounded in high-fidelity historical simulations—in a 

dedicated summary positioned at the end of the Introduction. This would 

directly address the first specific comment.

 

b) The clarity and reproducibility of the synthetic observation operator, a critical 

methodological element, could be substantially enhanced by supplementing the 

textual description with a well-designed visual flowchart, thereby directly 

responding to the second specific comment.

 

Specific Questions

 

1. Clarify Innovation and Contributions

The advancements relative to Shaddy et al. could be summarized more 

concisely and forcefully in the Introduction.

 

While Section 1.3 (pages 4–5) provides a comparative discussion with prior 

studies, the key innovations are dispersed throughout the paragraph. It is 

recommended to incorporate a distinct, clearly highlighted paragraph at the 

conclusion of the Introduction (following Section 1.3, at the bottom of page 5) 

under the heading: **“The main contributions and innovations of this paper are 

as follows:”.

 

2. Improve Methodological Reproducibility and Clarity

The 11-step synthetic observation procedure detailed in Section 2.2.2 (pages 7

–8) is thorough but intricate. It is recommended that a flowchart diagram be 

added to visually summarize the sequence of transformations from simulated 

fire arrival times to synthetic measurement outputs.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Comments:

Shaddy et al. Generative Algorithms for Wildfire Progression Reconstruction from Multi-Modal Satellite Active Fire Measurements and Terrain Height

This manuscript presents a generative algorithm for reconstructing wildfire progression using satellite active fire observations. The approach is methodologically sound, and the results appear robust. Overall, the paper is clearly structured and has potential application value. I recommend acceptance after minor revisions.

1. Since WRF-SFire outputs are used to train the algorithm, their accuracy is essential. More clear and robust validation of the simulation results is needed to support the reliability of the training data.

2. Figures containing geographic coordinates currently show overly sparse latitude and longitude ticks. More detailed and continuous coordinate labels should be added to enhance readability and facilitate spatial interpretation.

3. Are the 140 CONUS wildfire simulations from 2023 sufficiently diverse in terrain, fuel types, and meteorological conditions to avoid training bias and ensure that the model generalizes well to atypical fire behaviors?

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript proposes a wildfire spread model that integrates physical models with deep learning models. While demonstrating methodological innovation and promising application results, it still presents several issues, specifically as follows:

1. The abstract needs to be further condensed. The latter half lacks a clear focus, and the overall structure appears somewhat loose.
2. The introduction section lacks comparison with the most recent studies. Many cited references are relatively old, and there is excessive self-citation in key parts.
3. The methodology section is overly redundant and difficult to follow. Most of the work has already been presented in the authors’ previous studies. It is recommended that the authors reorganize this section to highlight the core innovations. In particular, Section 2.2.2 reads more like an engineering report than an academic paper.
4. Sections 3.1 and 3.2 mainly describe the data, and it is not recommended to place these in the results section.
5. The manuscript lacks comparative experiments with mainstream models. Comparisons are made only with the authors’ previous work, which is insufficient to demonstrate the model’s advancement.
6. The authors use only cWGAN, and the generator is based on UNet—both relatively outdated models. A reasonable justification is needed.
7. Have the authors considered the discrepancy between the data distribution generated by the GAN and the real data distribution? This issue may cause the model to fail directly.
8. The discussion section appears superficial and should be separated from the conclusion.

Author Response

Please see the attachment. 

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

General Evaluation:

The authors have addressed the previous review comments thoroughly. The revisions have significantly enhanced the clarity and logical flow of the manuscript.

Regarding "Clarify Innovation and Contributions":
The authors have successfully implemented the recommendation to consolidate the paper’s contributions. By introducing a dedicated section titled “Main Contributions and Innovations” accompanied by a concise numbered list, the novelty of the terrain-conditioned framework and the physics-grounded training paradigm is now clearly and prominently articulated. This revision directly resolves the earlier concern regarding the previously scattered presentation of the innovation claims.

Regarding "Improve Methodological Reproducibility and Clarity":
In response to the suggestion for a flowchart in Section 2.2.2, the authors have restructured the complex 11-step synthetic observation procedure into a rigorous, sequential list. Although a visual aid was initially proposed, the current step‑by‑step textual breakdown provides a highly logical and unambiguous roadmap for the methodology. This systematic refinement effectively addresses the earlier complexity issue and ensures the reproducibility of the data generation pipeline.

Reviewer 4 Report

Comments and Suggestions for Authors

The author has effectively addressed my concerns, and the manuscript is acceptable for publication.