A Near-Real-Time Operational Live Fuel Moisture Content (LFMC) Product to Support Decision-Making at the National Level
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper focuses on the development of a near-real-time Live Fuel Moisture Content (LFMC). The research objective is clear and good for wildfire management decision-making. However, there are some issues in aspects such as formatting and content presentation that need to be revised and improved accordingly.
In the abstract and some lines, there are unnecessary blank lines, which affect the overall layout of the text. The redundant blank spaces should be removed to make the abstract more compact and aesthetically pleasing.
Meanwhile, when introducing the research work in the abstract, it is necessary to highlight the differences and improvements compared with previous works, so as to more clearly demonstrate the innovation and value of the research, enabling readers to quickly grasp the contributions.
Figure 1 lacks sufficient clarity. It is difficult to clearly present the key information of the sampling sites, which may interfere understanding of the sampling situation in the research area. It is good to enlarge and center the image. If the desired effect cannot be achieved, it should be replaced with a higher-quality and clearer image source to ensure that the image can accurately convey information and enhance the visual effect of the paper. In addition, subsequent images also need to be clear to avoid similar problems.
The layout of the tables and the text in the article is inconsistent, which affects the overall standardization and readability of the paper. Some tables can reduce the white space to make them more compact. Also, please consider unifying the alignment of the tables and the text so that the tables are integrated with the relevant text. This will facilitate readers to compare and understand the data, thus enhancing the overall quality of the paper.
There is a text formatting problem in lines 696 - 705. The author needs to carefully check and correct the formatting problem in this part to ensure the consistency of the paper's format and improve the quality of the paper from the details.
When explaining the reasons for sample selection in lines 137 - 144, only factors such as historical fire data and species diversity coverage are mentioned, and the explanation of the representativeness of the samples is insufficient. To enhance the persuasiveness, it is necessary to further explain the necessaries of these samples in terms of geographical distribution, vegetation type coverage, etc., and more comprehensively demonstrate the rationality of sample selection, so that readers can better understand the reliability of the research results.
Author Response
The paper focuses on the development of a near-real-time Live Fuel Moisture Content (LFMC). The research objective is clear and good for wildfire management decision-making. However, there are some issues in aspects such as formatting and content presentation that need to be revised and improved accordingly.
Comment #1: In the abstract and some lines, there are unnecessary blank lines, which affect the overall layout of the text. The redundant blank spaces should be removed to make the abstract more compact and aesthetically pleasing.
Response #1: The blank spaces have been removed.
Comment #2: Meanwhile, when introducing the research work in the abstract, it is necessary to highlight the differences and improvements compared with previous works, so as to more clearly demonstrate the innovation and value of the research, enabling readers to quickly grasp the contributions.
Response #2: We believe it is better to avoid an excessively large abstract and keep it within limits. The innovation and value of the research is stated in these lines of the Abstract: “The ability to estimate LFMC is important to improve our capability to predict when and where large and intense wildfires can occur. Currently, there is a gap in providing reliable near-real time LFMC estimates useful for operational purposes that can contribute for better operational decision-making. The objective of this work was to develop a near-real time LFMC estimates for operational purposes in Portugal”
Comment #3: Figure 1 lacks sufficient clarity. It is difficult to clearly present the key information of the sampling sites, which may interfere understanding of the sampling situation in the research area. It is good to enlarge and center the image. If the desired effect cannot be achieved, it should be replaced with a higher-quality and clearer image source to ensure that the image can accurately convey information and enhance the visual effect of the paper. In addition, subsequent images also need to be clear to avoid similar problems.
Response #3: Regarding Figure 1, we have made some changes and hope these have improved its understanding and readability. We have also added snapshots of each site and increase the quality of the figure. Regarding the other figures, we would need additional information to understand which figures and what aspects need to be improved.
Comment #4: The layout of the tables and the text in the article is inconsistent, which affects the overall standardization and readability of the paper. Some tables can reduce the white space to make them more compact. Also, please consider unifying the alignment of the tables and the text so that the tables are integrated with the relevant text. This will facilitate readers to compare and understand the data, thus enhancing the overall quality of the paper.
Response #4: Tables, and the entire manuscript, were created using Fire’s template and guidelines. However, there seems to be a problem with the order of the columns of the Tables. The files we submitted originally are correct, but for some reason that we are not aware of, the tables have been “flipped”. We will not change the tables at this stage to avoid repeating the same problem, that probably concerns FIRE’s manuscript processing system.
Comment #5: There is a text formatting problem in lines 696 - 705. The author needs to carefully check and correct the formatting problem in this part to ensure the consistency of the paper's format and improve the quality of the paper from the details.
Response #5: If the text formatting problem refers to the web link, we have solved it. Otherwise, please specify.
Comment #6: When explaining the reasons for sample selection in lines 137 - 144, only factors such as historical fire data and species diversity coverage are mentioned, and the explanation of the representativeness of the samples is insufficient. To enhance the persuasiveness, it is necessary to further explain the necessaries of these samples in terms of geographical distribution, vegetation type coverage, etc., and more comprehensively demonstrate the rationality of sample selection, so that readers can better understand the reliability of the research results.
Response #6: We apologize for the confusion, there was an error in our initial text. After consulting back with the Forest Service we were informed that the selection of sampling locations was done mostly based on operational factors. We have significantly changed this text (now in Lines 113-134).
Reviewer 2 Report
Comments and Suggestions for AuthorsThe abstract mentions that there is a gap in the availability of reliable LFMC estimates, but it could be better reinforced how this study fills this specific gap. Some important aspects, especially between "Introduction" and "Materials and Methods", could have a more fluid transition to better connect the problem with the proposed methodology. The modeling section with Random Forests is well described, but it would be useful to better explain the choice of hyperparameters and how validation may have influenced the results. There is mention of limitations of the available data (e.g.: ERA5-Land was not used). A brief discussion on the impact of the absence of these variables could strengthen the investigation section. The discussion could deepen the comparison with similar studies and stimulate the practical impact of the model. Some specifications could be mentioned, but there is greater detail on the impact of the lack of data in southern Portugal. Some figures present important data, but the captions could be more descriptive to facilitate quick understanding without having to consult the main text. Highlight the most relevant metrics to facilitate visualization of the results.
Author Response
Comment #1: The abstract mentions that there is a gap in the availability of reliable LFMC estimates, but it could be better reinforced how this study fills this specific gap.
Response #1: We believe that the 3rd and 4th sentences in the abstract provide sufficient information to the reader. The 3rd sentence states the gap and its importance. The 4th states the objective and implicitly how this fills the gap. We could make it more explicit but that would repeat text in a word-limited Abstract.
Comment #2: Some important aspects, especially between "Introduction" and "Materials and Methods", could have a more fluid transition to better connect the problem with the proposed methodology.
Response #1: We thank the reviewer for the comment but would need more specific information to improve the “flow” of the manuscript.
Comment #3: The modeling section with Random Forests is well described, but it would be useful to better explain the choice of hyperparameters and how validation may have influenced the results.
Response #3: We have revised the text in the Methods and added a table to Appendix A with the values of the parameters used to build the grid search for the tuning of the model. This adds transparency and reproducibility to the modeling approach. The validation step has not determined a different selection of parameters, as for the result of the grid search, although we opted to increase the number of trees to 250, to enhance generalization of the model, as stated in the manuscript.
Comment #4: There is mention of limitations of the available data (e.g.: ERA5-Land was not used). A brief discussion on the impact of the absence of these variables could strengthen the investigation section.
Response #4: The objective of the work was to develop near-real time LFMC estimates. The selection of variables and datasets depends on the objective. Therefore we did not select ERA5 Land data because it is not (yet) available for near-real time predictions. Section 4.4 already provides several insights on potential improvements, including the use of ERA5 Land data. Adding more information would be based on speculation, because we have no evidence if including other data sources would improve the LFMC estimates.
Comment #5: The discussion could deepen the comparison with similar studies and stimulate the practical impact of the model. Some specifications could be mentioned, but there is greater detail on the impact of the lack of data in southern Portugal.
Response #5: We mentioned in detail the lack of data for southern Portugal because it is a clear limitation of the work. We also showed that, based on the possible analysis, LFMC estimates in the south seem to be consistent with the rest of the country. Regarding the comparison with other studies: section 4.2 has an extensive comparison with different modelling approaches and model performances presented in other studies.
Comment #6: Some figures present important data, but the captions could be more descriptive to facilitate quick understanding without having to consult the main text. Highlight the most relevant metrics to facilitate visualization of the results.
Response #6: We have improved the captions and corrected the captions of Figures 8 and 11 that were initially wrong.
Reviewer 3 Report
Comments and Suggestions for Authorsfire-3544327 – A near-real-time operational Live Fuel Moisture Content (LFMC) product to support decision-making at national level – a Review
I have completed my review of the paper fire-3544327, “A near-real-time operational Live Fuel Moisture Content (LFMC) product to support decision-making at national level”. In this paper the authors use ~1,000 measurements of Live Fuel Moisture Content (LFMC) from sixteen sampling sites located in different regions in Portugal. Then they combined various variables from the Google Earth Engine (GEE) and developed a Random Forest (RF) model to predict LFMC model, while separating the data to training and validation cohorts. Next, they present descriptive statistics of wildfire data with respect to LFMC values in the same region.
The topic is interesting and the paper is overall well written and easy to follow. However, I currently have major concerns which I believe should be addressed before the paper can be considered for publication in Fire.
My primary concern is to the generalizability of the model. The site-level assessment shows that the performance of the model is extremely low in some regions. Also, did the authors include the station number as an input in the general analysis? If so, this would make the model not generalizable to other stations.
Figure 4 – how do the authors account for different vegetation types than those used in the measurement sites?
The authors include several fire weather indices and subindices as predictors. However, don’t the subindices (FFMC for example) include data on LFMC?
The authors miss out on some very important variables that could potentially improve the performance of their models. For example, I think precipitation and humidity have to be included in such model.
How was the Machine Learning (ML) model chosen? The authors should compare its performance with additional models such as XGBoost and LightGBM, which have been shown to have better performance in tabular data predictions.
How were hyperparameters chosen for the ML model? The authors should provide these details for the study to be replicable.
The authors use figures excessively, sometimes with no reasons. The number of figures should be reduced and some of them should be moved to the appendix. Instead, the authors should provide more summary statistics. For example, the authors should use a linear or logistic regression to show how changes in LFMC affect wildfire risk, instead of the current descriptive statistics.
“These discrepancies were often due to improper handling, mistakes during collection (i.e. samples containing flowers, fruits, or woody material) or post-processing errors, causing anomalously high or low LFMC values.” – the authors should clarify how outlier removal was performed.
Were all measurements performed in the same hour? If not, how do the authors account for this discrepancy?
The authors should clarify when they are referring to percentages, and when they are referring to percentage points. This is confusing as LFMC is measured in %.
Author Response
I have completed my review of the paper fire-3544327, “A near-real-time operational Live Fuel Moisture Content (LFMC) product to support decision-making at national level”. In this paper the authors use ~1,000 measurements of Live Fuel Moisture Content (LFMC) from sixteen sampling sites located in different regions in Portugal. Then they combined various variables from the Google Earth Engine (GEE) and developed a Random Forest (RF) model to predict LFMC model, while separating the data to training and validation cohorts. Next, they present descriptive statistics of wildfire data with respect to LFMC values in the same region.
Comment #1: The topic is interesting and the paper is overall well written and easy to follow. However, I currently have major concerns which I believe should be addressed before the paper can be considered for publication in Fire.
My primary concern is to the generalizability of the model. The site-level assessment shows that the performance of the model is extremely low in some regions. Also, did the authors include the station number as an input in the general analysis? If so, this would make the model not generalizable to other stations.
Response #1: We agree that the generalization of the model was a concern, given the uneven spatial distribution of the sampling sites, mainly for the south part of the country. This was the reason why we increased the number of trees to 250 trees, as stated in the LFMC modelling section, even if this was not identified as necessary by the fine tuning step of the hyperparameters. We also reserved 25% of records to the test set, more than in the usual 80-20 in the train-test split, to ensure that any bias or overfitting could be detected. We caution readers of these limitations of the generalization in the results in the first paragraph of the discussion.
Regarding the site-level assessment. We disagree with the reviewer’s opinion. Validation RMSE was between 3.06 and 16.43 which is very reasonable for operational purposes and compares well with the performances of other studies (section 4.2). R2 is low for two sampling sites (Monsanto, Oleirinhos) that have a very small sample size. For remaining it is between 0.53 and 0.89, which is between reasonable and good.
Regarding the station number as input: We did not include it as a variable (see Table 3).
Comment #2: Figure 4 – how do the authors account for different vegetation types than those used in the measurement sites?
Response #2: The variables selected in the LFMC model do not include vegetation type (section 3.1). LFMC estimates are produced for the entire country and later masked for areas with potential shrub existence (section 2.3).
Comment #3: The authors include several fire weather indices and subindices as predictors. However, don’t the subindices (FFMC for example) include data on LFMC?
Response #3:The fire weather index and subindices do not include data on LFMC, as described here:
van Wagner, C.E.; Stocks, B.J.; Lawson, B.D.; Alexander, M.E.; Lynham, T.J.; McAlpine, R.S. Development and Structure of the Canadian Forest Fire Behavior Prediction System. Fire Danger Group, Forestry Canada: Ottawa, ON, Canada, 1992.
Comment #4: The authors miss out on some very important variables that could potentially improve the performance of their models. For example, I think precipitation and humidity have to be included in such model.
Response #4: Precipitation and relative humidity are included in the model through the Drought Code (DC) variable.
Comment #5: How was the Machine Learning (ML) model chosen? The authors should compare its performance with additional models such as XGBoost and LightGBM, which have been shown to have better performance in tabular data predictions.
Response #5: Although the modelling component of the study is important, it was not the main focus of it. For these reasons, we selected a model algorithm which is of wide adoption and robust for this case, as stated in the manuscript. The performance metrics for Random Forest showed to be robust enough to support the use of its results in the current application.
Comment #6: How were hyperparameters chosen for the ML model? The authors should provide these details for the study to be replicable.
Response #6: The hyperparameter values were selected combining a grid search approach with a cross validation of the train set. We added as a table in Appendix A with the values used in the grid search, and changed the text on the section 2.2.2 to make this further clear to the readers
Comment #7: The authors use figures excessively, sometimes with no reasons. The number of figures should be reduced and some of them should be moved to the appendix. Instead, the authors should provide more summary statistics. For example, the authors should use a linear or logistic regression to show how changes in LFMC affect wildfire risk, instead of the current descriptive statistics.
Response #7: We respectfully disagree with the reviewer's opinion. We believe that all figures bring added insights to the Methods, but most importantly to the Results. The number of figures compares well with other papers published in Fire.
Regarding summary statistics, we believe that summary statistics are less appealing than figures for the readers. It is probably a question of style.
Regarding the use of linear\logistic regression: the objective of the work was not to define mathematical relations between LFMC and fire size (or other fire related variable). We use the comparison as a way to show that the dataset is robust and can be used for fire risk monitoring. Furthermore, linear\logistic regressions would not be the best options considering the scatterplots shown throughout the several figures of the work.
Comment #8:“These discrepancies were often due to improper handling, mistakes during collection (i.e. samples containing flowers, fruits, or woody material) or post-processing errors, causing anomalously high or low LFMC values.” – the authors should clarify how outlier removal was performed.
Response #8: We refer in the text that “AGIF and ICNF identified anomalous values during sample collection that were screened”. These anomalous values were identified and screened by comparison with prior and subsequent LFMC values and by comparison with nearby sampling sites. We added more information in Lines 142-143.
Comment #9: Were all measurements performed in the same hour? If not, how do the authors account for this discrepancy?
Response #9: All measurements were made between 14h30 and 15h00. Sample integrity was guaranteed by the Forest Service.
Comment #10: The authors should clarify when they are referring to percentages, and when they are referring to percentage points. This is confusing as LFMC is measured in %.
Response #10: In every part of the manuscript were “%” was referred to, we identified if we were referring to LFMC, DFMC or percentage of the (sub) sample. If the reviewer identifies parts of the manuscript that need clarification we kindly ask to provide the line number(s).
Round 2
Reviewer 3 Report
Comments and Suggestions for Authorsfire-3544327R1 – A near-real-time operational Live Fuel Moisture Content (LFMC) product to support decision-making at national level – a Review
I have completed my review of the revised version of the paper fire-3544327, “A near-real-time operational Live Fuel Moisture Content (LFMC) product to support decision-making at national level”. I thank the authors for their detailed and elaborate responses. I am pleased with all the responses given, and have no further major comments. I believe the manuscript is ready for publication.
Two last things:
- I still believe that the number of figures is too high, and that some should be moved to the appendix. But of course this is a subjective issue, so I leave this to the decision of the authors and editors.
- I assume/hope the authors make the data available to readers. I believe this is crucial to increase the impact of this study.
I congratulate the authors for their great results.
Author Response
Comment 1: I still believe that the number of figures is too high, and that some should be moved to the appendix. But of course this is a subjective issue, so I leave this to the decision of the authors and editors.
Response: We thank the reviewer for the comments. We have reviewed the figures and believe all of them are necessary to convey the message.
Comment 2: I assume/hope the authors make the data available to readers. I believe this is crucial to increase the impact of this study.
Response: The data is available at Zenodo and the link is provided in the data availability statement section.
Comment 3: I congratulate the authors for their great results.
Response: We thank the reviewer for the kind words and effort provided in the review of the manuscript.