Review Reports - Forest Fire Risk Early Warning Based on Dynamic Fuel Moisture Content

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study assesses the marginal contribution of integrating dynamic FMC into fire prediction models. Concentrating on California, we developed a random-forest-based model that incorporates high-resolution FMC products retrieved by our team, along with meteorological, topographic, vegetation, and anthropogenic data.

It is unclear whether the model performance improvement stems from data quality or methodology.
Clearly state the validation details of the FMC product (RMSE, R², field data).
Spatial cross-validation (block CV) should be used.
The use of a single model (Random Forest (RF)) weakens the power of the article.
Meteorological variables can be expanded.
It can be tested with true class distribution.

Author Response

Original Manuscript ID: forests-4248901

Original Article Title: “Forest Fire Risk Early Warning Based on Dynamic Fuel Moisture Content”

To: Forests Editor

Re: Response to reviewers

Dear Reviewer,

Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.

We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), (b) an updated manuscript with red highlighting indicating changes, and (c) a clean updated manuscript without highlights (PDF main document).

Best regards,

Cui Zhou and Yongfeng Luo

School of Forestry & School of Low-Altitude Economy

Central South University of Forestry and Technology

Changsha 410004, China

Email: yfluo@csuft.edu.cn

Response to Reviewer 1 Comments

Comment 1: It is unclear whether the model performance improvement stems from data quality or methodology.

Response: Thank you for posing this crucial question. To differentiate the contributions of data quality and methodology to the improvement of the model's performance, we have supplemented our analysis and discussion in the revised manuscript from the following two aspects:

Data quality: The FMC data utilized in this study were sourced from the NRCBLA framework previously developed by our team (refer to Section 2.2.1 and Figure 3). This framework is founded on the CNN-BiLSTM-Attention architecture, incorporating MCD43A4 spectral data, ERA5 meteorological data, and SRTM topographic data, and employs distinct modeling strategies for diverse vegetation types. As presented in Table 1, the FMC product attains R² values spanning from 0.5467 to 0.7511 and RMSE values ranging from 12.73% to 21.74% across various vegetation types, indicating favorable accuracy and reliability.

Methodology: To independently validate the effectiveness of the methodology, a rigorous comparative experiment was designed in the revised manuscript (refer to Section 4.2, Tables 4, 5, 6, and Figures 8, 9). Under identical data input conditions (with all other variables remaining consistent except for FMC), the predictive performance of two random - forest models was compared: one incorporating dynamic FMC and the other without it. The results indicate that after the incorporation of FMC, the model's AUC increased from 0.9515 to 0.9668, the precision increased from 83.07% to 87.00% (a 3.93% improvement), and specifically, it increased from 86.35% to 89.94% (a 3.60% improvement). Given that both experiments employed the same algorithmic framework (random forest), the same hyperparameter configurations, and the same training/test set splits, with the sole difference being the inclusion or exclusion of the FMC variable, the performance improvement can be ascribed to the introduction of FMC as a feature variable, rather than to differences in data quality or algorithms.

Comment 2: Clearly state the validation details of the FMC product (RMSE, R², field data).

Response: We appreciate the reviewer for pointing this out. We have made changes according to your suggestion.

Comment 3: Spatial cross-validation (block CV) should be used.

Response: We sincerely appreciate your valuable suggestions. We have made changes according to your suggestions.

Comment 4: The use of a single model (Random Forest (RF)) weakens the power of the article.
Meteorological variables can be expanded.

Response: We express our gratitude to the reviewer for highlighting this issue.

It should be noted that the random forest (RF) method was selected as the core modeling approach in this study due to its maturity and interpretability in the field of wildfire risk prediction. As an ensemble learning method, RF can effectively handle high-dimensional features, nonlinear relationships, and interactions among features, and has been widely applied in ecological and fire modeling. The primary objective of this study is to validate the marginal contribution of dynamic FMC as a newly added feature variable, rather than to compare the performance of different algorithms. Under this research framework, the use of a single yet stable RF model, combined with strictly controlled experiments (where all inputs, hyperparameters, and data partitions remain consistent except for FMC), is sufficient to clearly attribute the source of performance improvement. Furthermore, the spatial block cross-validation and multi-metric evaluation system (including AUC, precision, specificity, etc.) adopted in this study ensure the robustness of the conclusions from a methodological perspective. Therefore, despite the use of a single model type, the rigor of the experimental design is sufficient to support the core findings.

This study adopts a prudent strategy in the selection of meteorological variables. In the FMC retrieval phase (NRCBLA framework), we indeed utilized a relatively rich set of meteorological variables, including U/V wind speed, 2-meter temperature, evapotranspiration, and vegetation interception (see Section 2.2.1 of the original manuscript for details). However, in the subsequent random forest-based wildfire risk prediction model, we introduced only temperature as a meteorological feature, primarily for the following reasons:

First, to avoid multicollinearity. FMC itself already comprehensively reflects the influence of multiple meteorological factors on fuel moisture content. Introducing several meteorological variables that are highly correlated with FMC (such as wind speed and evapotranspiration) into the prediction model could lead to severe multicollinearity among features, interfering with accurate estimation of the marginal contribution of individual variables and even causing overfitting.

Second, to maintain clarity of research focus. The core objective of this study is to quantify the marginal contribution of dynamic FMC as a feature variable to wildfire risk prediction. By including only temperature as a basic meteorological indicator while controlling for other variables, the predictive gain brought by FMC can be more clearly isolated, avoiding the dilution of the core conclusions due to excessive expansion of the meteorological variable set.

Comment 5: It can be tested with true class distribution.

Response: We express our gratitude to the reviewer for pointing this out. We have made corrections based on your comments.

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you very much. Please see my comments below.

Major revisions

1. FMC Data method Details, and Validation

Please provide a detailed and reproducible description of the NRCBLA retrieval algorithm, including input variables, models/formulae, processing steps, calibration and assimilation procedures, spatial and temporal coverage, and any underlying assumptions.

Additionally, report validation metrics for the FMC product within the study area, including bias, RMSE, and correlation against in-situ FMC observations or the U.S. National Fuel Moisture Database used in development.

2. Temporal Matching of FMC and Fire Occurrences

The correlation analysis and model depend on the temporal alignment between FMC (daily) and fire occurrences, specifically the MODIS MCD64A1 monthly burned area with Julian day. Please clarify the method of temporal matching: Were fire pixels matched to the FMC and other covariates on the same Julian day? If not, what temporal window was utilized? How were ambiguous burn dates addressed?

Discuss any potential temporal mismatches and their implications for both the correlation and the model.

3. Sampling Strategy and Class Imbalance Handling

Please provide further details on how non-fire points were selected spatially to prevent sampling bias (e.g., stratified by land cover, elevation, NDVI, or FMC ranges) and to ensure they are representative of the background environment.

4. Model Evaluation and Reporting

Please outline the hyperparameter tuning strategy or rationale for your choices. Were these values tuned via grid search or selected a priori? If they are fixed, please provide justification.

Please Include complete performance results for both baseline and integrated models across the 5-fold cross-validation and the independent test set for all reported metrics (AUC, accuracy, precision, recall, F1, and specificity). Please provide confusion matrices or at least counts of true positives, true negatives, false positives, and false negatives for the test set.

Present ROC curves for both models.

5. Please provide implementation details regarding SHAP.

6. The manuscript focuses on California and data from 2022 to 2023. Please discuss the applicability of your findings to other regions and longer historical periods. If feasible, provide a brief evaluation of model performance in different subregions or years to assess robustness.

Minor revisions

1. Please provide the exact product versions and acquisition dates (e.g., specify the version of MCD43A4 used, and while MODIS MCD64A1 Version 6.1 is mentioned, include the download dates as well). For LandScan and Esri World Roads, ensure that a version or year is included and maintain consistency in the citation format.

2. Please describe the gap-filling or cloud-masking procedures utilized for MODIS-derived NDVI and FMC.

3. Please provide maps of predicted risk for at least one representative day or month, along with an overlay of observed burned areas for a qualitative assessment.

Thanking you,

Best regards,

Author Response

Original Manuscript ID: forests-4248901

Original Article Title: “Forest Fire Risk Early Warning Based on Dynamic Fuel Moisture Content”

To: Forests Editor

Re: Response to reviewers

Dear Reviewer,

Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.

Best regards,

Cui Zhou and Yongfeng Luo

School of Forestry & School of Low-Altitude Economy

Central South University of Forestry and Technology

Changsha 410004, China

Email: yfluo@csuft.edu.cn

Response to Reviewer 2 Comments

Comment 1: 1. FMC Data method Details, and Validation

Response: We express our gratitude to the reviewer for highlighting this issue. We have made corrections in accordance with your comments.

Comment 2: 2. Temporal Matching of FMC and Fire Occurrences

Response: We sincerely express our gratitude for your valuable suggestions. We have made corrections based on your comments.

Comment 3: 3. Sampling Strategy and Class Imbalance Handling

Response: We sincerely express our appreciation for your valuable suggestions. We have made corrections based on your comments.

Comment 4: 4. Model Evaluation and Reporting

Please outline the hyperparameter tuning strategy or rationale for your choices. Were these values tuned via grid search or selected a priori? If they are fixed, please provide justification.

Present ROC curves for both models.

Response: We express our gratitude to the reviewer for highlighting this issue. Concerning the hyperparameter tuning strategy, the hyperparameter settings and the rationale for their selection regarding the random forest model in this study have been elaborated in the manuscript (refer to Section 3.1). With respect to model performance evaluation, detailed performance metrics and ROC curve information have been incorporated into the manuscript: we have appended the comprehensive performance metrics and ROC curve information for the 5-fold cross-validation on the training set, and the comprehensive performance metrics and ROC curve for the test set.

Comment 5: Please provide implementation details regarding SHAP.

Response: We sincerely express our appreciation for your valuable suggestions. We have made corrections in accordance with your comments.

Comment 6: The manuscript focuses on California and data from 2022 to 2023. Please discuss the applicability of your findings to other regions and longer historical periods. If feasible, provide a brief evaluation of model performance in different subregions or years to assess robustness

Response: We express our gratitude to the reviewer for highlighting this issue. Concerning the assessment of model robustness at various time points, apart from generating dynamic monthly risk maps for 2023, this study additionally selected a representative extreme high-temperature and drought day, September 9, 2020, to carry out a single-day fire risk analysis.

Comment 7: 1.Please provide the exact product versions and acquisition dates (e.g., specify the version of MCD43A4 used, and while MODIS MCD64A1 Version 6.1 is mentioned, include the download dates as well). For LandScan and Esri World Roads, ensure that a version or year is included and maintain consistency in the citation format.

Please describe the gap-filling or cloud-masking procedures utilized for MODIS-derived NDVI and FMC.

Response: We express our gratitude to the reviewer for highlighting this matter. In response to the issues you have raised, we have incorporated additional context.

Comment 8: Please provide maps of predicted risk for at least one representative day or month, along with an overlay of observed burned areas for a qualitative assessment.

Response: We express our gratitude to the reviewer for highlighting this issue. This study presents risk prediction maps for representative time periods and conducts an overlay analysis with the observed burned areas.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you for providing the revised version of the manuscript. I am pleased with your modifications. I appreciate the time and effort you dedicated to addressing my comments. Your attention to detail has significantly enhanced the clarity of the work.

Kind regards,