A Recurrent Neural Network for Forecasting Dead Fuel Moisture Content with Inputs from Numerical Weather Models
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a recurrent neural network (RNN) model for real-time forecasting of dead 10-hour fuel moisture content (FMC), incorporating both weather inputs and geographic predictors. The RNN’s performance is benchmarked against several common baselines: a physics-based ordinary differential equation model, an XGBoost machine learning model, and an hourly climatology. The results demonstrate substantial improvements in forecasting accuracy over all baseline methods. The following issues should be addressed to enhance the manuscript’s quality:
- Novelty: The authors note that this work builds upon Mandel et al. [25], who also employed an RNN for FMC forecasting. However, the specific methodological or conceptual advancements relative to that study are not clearly articulated. The novelty of the present approach must be explicitly stated.
- Hyperparameter rationale: The justification for the RNN hyperparameters listed in Table 2 should be clarified—e.g., whether they were selected via cross-validation, grid search, or domain knowledge.
- Missing conclusion: The manuscript currently lacks a conclusion section. A concise summary of key findings, implications, and potential future work should be added.
Author Response
We thank the reviewer for the valuable feedback. Additions to the manuscript are highlighted in red, removed text is stricken through in red, and revised figures and tables are marked with (Revised Figure) in red text. We respond to individual reviewer concerns below, with the topic from the reviewer report in bold font.
Novelty: We acknowledge this was lacking and we have revised the manuscript, see Lines 773-778 in the Discussion.
Hyperparameter Rational: We acknowledge the confusion that came from the presentation of hyperparameters. We added clarification on Lines 374-376 when introducing the core architecture that additional hyperparameters will be discussed later. Then we added the missing core hyperparameters to Table 2 and extra clarification on Lines 399-409.
Missing Conclusion: We acknowledge this concern and have revised the manuscript.
We hope the revised manuscript adequately addressed your concerns.
Reviewer 2 Report
Comments and Suggestions for AuthorsI recommend that a Introduction section should be separated from the literature review becasue they are combined together which makes the flow of the research paper vague. Two different sections (Introduction and literature review) would improve the quality of the manuscript. Also present a table that include existing literature relevant to the topic and highlight the methods and tools already been used and then present the research gap and propose the current research approach.
In the Data Method Section: I recommend that a description on the Geographic location with demographic details are important along with geographic information. This is missing from the manuscript. I strongly recommend this to be considered. Also a cartographically sound map is required for the study area.
Discussion Section: The discussion section is very shallow and needs improvement in the context of other fire research that has already been published and how this study results are different from them.
Finally, the manuscript is missing conclusion section. I recommend that the discussion section should be separated from the conclusion and future recommendation should be under the conclusion section.
Author Response
We thank the reviewer for the valuable feedback. Additions to the manuscript are highlighted in red, removed text is stricken through in red, and revised figures and tables are marked with (Revised Figure) in red text. We respond to individual reviewer concerns below, with the topic from the reviewer report in bold font.
Introduction Section: We added subsections in red text to break up the flow, and we made additional revisions to the literature review.
Data Method Section: We added geographic details on lines 247-255. Additionally, Figure 2, Figure 9, and the new Figure 11 added cartographic details to the maps, including information on the projection and longitude/latitude labels.
Discussion Section: Additional details have been added, including details on the novelty of the paper.
Missing Conclusion: We acknowledge this concern and have revised the manuscript.
We hope the revised manuscript adequately addressed your concerns.
Reviewer 3 Report
Comments and Suggestions for AuthorsGeneral comments
The manuscript presents an LSTM-based network for 48 h forecasts of 10-h dead fuel moisture content across the Rocky Mountains, driven solely by HRRR output. A rigorous spatiotemporal cross-validation (2023 training, 2024 prediction, fully independent stations) yields an RMSE of 3.02%, 39% lower than the operational ODE+KF scheme inside WRF-SFIRE. All code and raw-processed data are released, ensuring full reproducibility. These strengths make the study immediately relevant to operational fire-weather centres.
Major issues
My central concern is the incomplete treatment of uncertainty. The authors quote 3.02% ± 1.12%, but this standard deviation reflects only the 500-fold random split of training vs test stations. It omits the dominant error sources: RAWS sensor accuracy (≈1.5% FMC), HRRR systematic biases in temperature, humidity and precipitation, and network weight initialisation. I therefore suggest a new “Total Uncertainty” subsection/part that (i) decomposes variance into observation, meteorological input and model components, (ii) provides P50, P90, and P95 RMSE values, and (iii) explicitly warns readers that the quoted interval is not a predictive confidence range.
The manuscript lacks a case-study validation during an extreme fire-weather day. Providing a 72 h time series (predicted vs observed FMC) for a 2024 critical episode; e.g., the East Troublesome re-ignition on 17 March 2024, would demonstrate behaviour near the 6–8% combustion threshold, which can be placed as the Supplement.
While the paper is application-oriented, a brief ablation table (four configurations) is needed to show the individual contribution of geographic static features and of each meteorological input. A short statement on why Transformer or ConvLSTM architectures were not explored should also be added to the Discussion. Phrases such as “forecasts for all of 2024” should be qualified in both Abstract and Discussion to avoid the impression that error remains 3% for multi-day horizons.
The manuscript’s most glaring structural gap is the complete absence of a stand-alone Conclusion section. Key findings such as final error level, operational applicability, and the 48 h lead-time ceiling, are dispersed across the Abstract, Results, and Discussion, contravening WMO guidance that forecast products must present a concise, self-contained summary of performance and limitations. A brief, unified conclusion should explicitly state the achieved RMSE, confirm the network’s readiness for plug-and-play deployment within WRF-SFIRE, and reiterate that accuracy guarantees apply strictly to forecasts ≤ 48 h.
Issues in literature review
The introductory review, while logically structured, suffers from three notable gaps: (1) it overlooks recent large-scale ML advances such as ConvLSTM and Transformer-based FMC products (2022–2024), (2) it under-represents foundational operational schemes (e.g., Canadian FWI, NFDRS-88) that form the historical baseline for the current ODE+KF system, and (3) it disproportionately cites the authors’ own work while omitting key physics-guided deep-learning studies published in the last two years. Expanding the literature coverage to include these missing references will both strengthen the claim of “first nationwide, forecast-ready AI” and demonstrate fair comparison with the state-of-the-art.
Issues in visual presentation
The figure set is generally clear, but several compositional adjustments are required to improve the quality of the figures and the essentials of the figure captions.
- Fig. 1: Add a miniature full-year strip (greyscale) in the upper left corner and a synchronized precipitation bar along the right y-axis so that humidity peaks/valleys can be instantly linked to rainfall events. In addition, provide more details for the presentation of the data in the figure caption.
- Fig. 3/4: Tensor shapes and dropout rates should be labelled inside each block; the cell-state arrow of the LSTM must be drawn separately from the hidden-state arrow to avoid the common “single-loop” misconception.
- Fig. 8: The font size is too tiny to show the parameters for two co-ordinates. In addition, the figure caption omits axis definitions, sample size and reference-line explanations, forcing readers to guess which panel shows observations versus predictions and obscuring the very comparison the figure is meant to highlight.
- Fig. 9: Switch to a diverging colormap (white at 5%) and overlay 5% and 10% RMSE contours; station dots should be binned into low (≤1800 m) and high (>1800 m) elevation sizes so that the alpine error cluster is immediately visible in print.
- Fig. 10: Introduce shape + line-style dual coding (solid circle for RNN, open triangle for ODE, dashed for climatology) and add a dual x-axis (UTC + Mountain Time) so that fire-duty forecasters can read the plot without mental time-zone conversion.
- New Fig. 11 recommended: Provide a 72 h critical fire case with predicted/observed FMC in the upper panel and synchronous HRRR rain + wind in the lower panel; include computed ROS in a side text box to demonstrate model behaviour near the 6–8% combustion threshold.
Once these revisions are implemented, the paper will offer a robust, transparent and immediately deployable replacement for the operational FMC module within WRF-SFIRE. I look forward to a revised version.
Author Response
We thank the reviewer for the valuable feedback. Additions to the manuscript are highlighted in red, removed text is stricken through in red, and revised figures and tables are marked with (Revised Figure) in red text. We respond to individual reviewer concerns below, with the topic from the reviewer report in bold font. We hope the revised manuscript adequately addressed your concerns.
Uncertainty: The different sources of error and uncertainty are important to consider. First, we note that the variability from network weight initialization was already accounted for and discussed in the original manuscript, see Lines 554-557. Next, what the reviewer describes as “systematic biases” will in part constitute the errors presented in the paper, and an ML model can learn systematic biases between inputs and outputs.
Uncertainty in the RMSE and Bias estimates resulting from uncertainty in the weather inputs is an important issue. However, the HRRR weather model provides only a single deterministic forecast, not a probabilistic forecast nor an ensemble. Quantifying the uncertainty in the FMC forecasts resulting from uncertainty in the weather inputs would require a complex analysis that would involve using estimates of uncertainty for each weather variable at each location in space and each time of the year. We do not know of such a study, and that is outside the scope of the paper. We discuss this issue on Lines 561-570 to let the reader know the uncertainty estimates do not include uncertainty from the weather.
Accuracy and uncertainty from the RAWS sensors are important to consider. As currently constructed, the model predicts what an FMC sensor will say in the future, and it will recreate any systematic errors or biases that those sensors have. FMC sensors are already used operationally, and the model presented in the paper could be used directly in place of the sensor readings. Our approach is thus agnostic to any attempts to correct for sensor errors. Using the output of an FMC sensor to characterize the moisture content of real-world fuel landscapes is a much broader research area. Having an accurate prediction of what FMC sensors will say in the future is a crucial step in that broader research area. We added clarification on Lines 162-167 and on Lines 780-784 about the goals of the model and the conservative way to interpret the forecasts.
In the Results section on Lines 620-624, we clarify that the uncertainty in overall RMSE is uncertainty in the aggregated error metric. This contrasts with the RMSE metrics we present in the new Case Study section which presents the median, min, and max RMSE values for individual predictions, see Lines 745-749.
Case-Study: We thank the reviewer for this suggestion. A case study has been added as Section 3.2. We chose to use the Alexander Mountain Fire, the largest in Colorado in 2024. We analyze the forecast accuracy by comparing them to the observations at 4 nearby RAWS stations in the pre-fire period. Further, we choose the pre-fire period, since during the fire the weather may be sufficiently modified by the local fire, and necessitate using a coupled atmosphere-fire model, which is outside the scope of this paper.
Predictor Importance: We thank the reviewer for this suggestion. A sensitivity analysis for different predictors and groups of predictors is presented in Section 3.1.
Other Methods, Transformer / ConvLSTM: We omitted these topics initially for brevity and because to our knowledge these techniques been applied in the context of live fuel moisture content (LFMC), rather than dead. We add a brief literature review of methods used in LFMC forecasting on Lines 141-151, which touches on the methods that the reviewer highlights. We discuss that the model is constructed point-wise in space and discuss the reasons for doing so on Lines 797-807. Further, we discuss the theoretical reasons why we did not use those other modeling techniques on Lines 814-822.
48-Hour Forecast Windows: The paper has been modified to include qualification about the forecasting window in all relevant places.
Missing Conclusion: We acknowledge this concern and have revised the manuscript.
Literature Review: We agree that it is valuable to contextualize live FMC modeling along with dead FMC. We add a brief scientific background on Lines 50-55 to discuss introduce LFMC and discuss the main modeling differences. Then we include a brief literature review on ML topics in LFMC on Lines 141-151. Living plants attempt to maintain homeostasis, so there is less short-term variability in LFMC than dead. Plant physiology is affected by historical drought, so the relevant timescales for LFMC are much longer than for dead. Finally, LFMC is affected by the broader ecological environment that the plant is embedded within, such as the soil composition, soil moisture, nearby species composition, canopy structure, water table depth, and other factors.
We note that we already cited references to the NFDRS on Lines 89-90, but we acknowledge this could be more complete, and we add another citation to the broader NFDRS research goals on Lines 95-96.
We acknowledge the issue with self-citation and removed several portions of text, highlighted in red with strike-through.
Figure 1: Acknowledged and revised. We added exact times and important weather variables to the figure, along with detail in Lines 74-78 about the FMC time series and interaction with the weather.
Figure 3/4: We respectfully disagree with the reviewer suggestions for these figures, and have elected to keep the figures unchanged. Figures 3/4 are intended to introduce the reviewer to recurrence and explain the functional mapping in time. We do not think that adding tensor shapes and dropout rates would add to this goal, and it could be distracting. Tensor shapes are partially determined by the hyperparameter of batch size, which is a tuned hyperparameter that doesn’t affect the functional mapping of the model. Dropout is a regularization technique used during training and inactive during prediction. We don’t think it would add to the clarity or purpose of the figure to include these elements in the figure.
Figure 8: Acknowledged and revised.
Figure 9: Acknowledged and revised with additional details on the High vs Low altitude model accuracy.
Figure 10: Acknowledged and revised.
New Figures: We thank the reviewer again for the suggestion to add a case-study. Additional figures have been added from the case-study, including a map of the Alexander Mountain fire and nearby RAWS (Figure 11), histograms of forecast errors at those RAWS (Figure 12), and time series of forecasts and weather at one representative station (Figure 13).
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI have no comment at this round of review.
Author Response
We thank the reviewer
Reviewer 3 Report
Comments and Suggestions for AuthorsThe manuscript presents an LSTM-based model for 48-hour forecasts of 10-h dead fuel moisture content (FMC) across the Rocky Mountains, driven solely by High-Resolution Rapid Refresh (HRRR) output. The model is trained on 2023 data from 161 Remote Automated Weather Stations (RAWS) and evaluated on 151 independent stations throughout 2024, generating roughly 46 million hourly predictions. A case study of the 2024 Alexander Mountain Fire is included to assess performance under pre-fire conditions.
This study appears to be the first real-time, continent-scale LSTM for 10-h dead FMC, and the RMSE drops from roughly 5% to just above 3%—a sizable reduction that supports its potential as a ready-to-use replacement for the current ODE+KF scheme. The code and data are fully open, so fire weather centres can slot it straight into operations with minimal additional effort. In the first round of revision the authors responded constructively to review comments: all visuals now meet readability standards (dual time axes, diverging colormaps, elevation binning). The Conclusions section clearly defines the 48-hour forecast horizon, and sensitivity analyses quantify the relative importance of meteorological and geographic inputs. The Alexander Mountain case shows a 72-hour dry-down to FMC levels commonly linked to critical fire danger. It should be better to split the one paragraph of this section into two, though.
Meanwhile, a few scientific wrinkles still remain. The new uncertainty section only reflects sampling variance from 500 station splits—sensor accuracy (~1.5% FMC) and HRRR systematic biases are not propagated; RMSE values are also absent. The selected episode involves relatively moderate wind speeds and thus does not fully represent the most extreme fire weather conditions where model performance is most critical. In addition, two major Transformer-based FMC studies in the years of 2023–24 are still not cited and compared so far. Addressing these points should further improve the quality of the paper.
Several figures would benefit from minor editorial adjustments to align with MDPI formatting guidelines. At the movement, some units are set in italic tyle (e.g., “m s⁻¹”, “%”) in axis labels and colorbars. They should be presented in upright (Roman) type, as required by SI and MDPI style guidelines. Font sizes remain inconsistent: axis titles and legend text in Figs. 6, 7, 9, and 10 are too small, while those in Figs. 11–13 even exceed 12 pt. Please standardize to 8–9 pt for all axes and legends, as required by the journal template and exemplified in Fig. 8, to ensure visual hierarchy and print readability.
Once these minor scientific and presentation issues are addressed, this paper should be suitable for publication in Fire. I recommend minor revision.
Author Response
We thank the reviewer for the valuable feedback. Text that was highlighted in red from the first round of revisions has been changed to normal color. New revisions are highlighted in red, with modified figures noted in red with “(Revised Figure)”. Our replies to Round 2 are organized below.
Uncertainty from sensors:
We do not attempt to correct for sensor errors, and we discuss why in the manuscript. The RNN model is intended to be used in place of sensor observations. The model will reproduce any errors associated with the sensors. The sensors are already in operational use, and different methods to account for systematic errors and uncertainty already exist in operational frameworks.
Uncertainty from weather:
Systematic biases in the weather inputs contribute to the forecast RMSE. But if they are systematic in the sense that they are fixed and consistent, then they will not contribute to uncertainty across replications. Uncertainty in the inputs would propagate to the accuracy metrics, but not systematic biases.
Quantifying the uncertainty from weather inputs from HRRR would be a valuable analysis, but it would necessitate a large follow-up study. The HRRR does not produce an ensemble nor any probabilistic forecasts, so propagating the uncertainty from the weather inputs would require an extensive study that estimated the uncertainty associated with each weather input at each location in the study area and across all times. We clarify in the manuscript the sources of uncertainty that we are and are not accounting for.
Transformer-Based FMC Studies (2023-2024):
We thank the reviewer for this suggestion. We attempted to locate the studies using the information provided and through standard literature searches but were unable to identify the specific works in question. We would be happy to include these references if additional details (e.g., full citations or alternative titles) can be provided.
Figure Formatting:
We thank the reviewer for identifying these inconsistencies in the visual presentation. Text sizing in Figures has been modified to be as standardized as possible given sizing constraints.
