Using artificial intelligence to monitor live fuel moisture content across France, based on a high-resolution land surface analysis

Baehr, Yann; Vanderbecken, Pierre; Bonan, Bertrand; Robert, Catherine; Regimbeau, Mathieu; Pimont, François; Raynal, Kevyn; Liu, Xiangzhuo; Savazzi, Remi; Garouani, Moncef; Mothe, Josiane; Rodriguez-Fernandez, Nemesio; Jarlan, Lionel; Calvet, Jean-Christophe

doi:10.5194/egusphere-2026-1247

Preprints

https://doi.org/10.5194/egusphere-2026-1247

Preprints

16 Apr 2026

| 16 Apr 2026

Status: this preprint is open for discussion and under review for Natural Hazards and Earth System Sciences (NHESS).

Using artificial intelligence to monitor live fuel moisture content across France, based on a high-resolution land surface analysis

Yann Baehr, Pierre Vanderbecken, Bertrand Bonan, Catherine Robert, Mathieu Regimbeau, François Pimont, Kevyn Raynal, Xiangzhuo Liu, Remi Savazzi, Moncef Garouani, Josiane Mothe, Nemesio Rodriguez-Fernandez, Lionel Jarlan, and Jean-Christophe Calvet

Abstract. Live Fuel Moisture Content (LFMC) is one of the most critical variables for understanding fire dynamics, particularly within forest environments. This study intends to develop a daily LFMC indicator to support operational fire danger management services in France. The product is based on in situ observations from the French National Forest Office and is generated using a lightweight expressive neural network model. The network has been designed to generalise well over time and space. It can be integrated directly into land surface models to enable real-time monitoring of vegetation’s hydric status. The modelling framework combines outputs from a physically based land surface model and satellite-derived leaf area index (LAI) observations, providing high-resolution, spatially consistent estimates of land surface over France. To evaluate the model’s generalisation capacity, we implemented complementary cross-validation strategies to test interannual robustness, spatial transferability, and to simulate an operational deployment scenario. Additionally, we performed a robustness analysis to quantify the sensitivity of predictions to training variability. The results demonstrate a strong ability to estimate the range and dynamics of LFMC across most of France. They also identify regions where additional in situ sampling or improved representation could reduce epistemic uncertainty and enhance the reliability of the model.

Received: 05 Mar 2026 – Discussion started: 16 Apr 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 1950 KB)

Supplement (1145 KB)

Download & links

Status: open (until 08 Aug 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-1247', Krishna Rao, 20 Jul 2026 reply
Understanding fire activity in Mediterranean ecosystems is paramount, especially under a changing climate that has driven recent increases in fire activity and associated hazards. Live Fuel Moisture Content (LFMC) is a critical indicator of fire danger, yet it remains poorly understood in spatial detail due to limited observational data and a lack of continuous, wall to wall maps. In this context, this paper is a very welcome contribution to the literature.
I view this manuscript as a "3 in 1" contribution: the authors not only produce daily LFMC estimates across France and adjacent regions, but they also provide actionable insights for selecting future sampling sites and evaluate the relationship between estimated LFMC and regional fire outbreaks at the canton level. The manuscript is well written, and the discussions regarding the physical mechanisms linking leaf area index modulated by soil water capacity are insightful. I commend the authors for extending their work beyond map generation to actively inform future sampling strategies.
However, I have several major questions and suggestions regarding methodological details that currently hinder a full evaluation of the findings. Addressing these points, I think, will significantly strengthen the clarity and impact of the paper.

1. Cross Species LFMC Averaging
While I recognize the constraints imposed by available data on species distribution, it is unclear why LFMC was averaged across species, especially given the authors' own observation that "the species are too varied" (L. 90). Extensive literature demonstrates significant inter species variation in LFMC responses under identical climate conditions (e.g., Martin StPaul et al., 2018; Pimont et al., 2019; Yebra et al., 2024).
Could an existing land cover or species map (e.g., Theia or IGN Tree Species Maps) be used to identify the dominant species at a given sampling site and isolate the corresponding LFMC time series?

Please clarify the rationale for averaging across species, or discuss the uncertainty this introduces into the downstream model targets.

2. Physical Mechanism Linking LAI/WFC to LFMC
Table 2 shows that the top feature linked to LAI is LAI/WFC (LAI normalized by soil water field capacity), with Section 4.6 providing details of a plausible mechanism where leaf growth under limited soil capacity leads to competing forces on plant water status, thus lowering live fuel moisture. While this mechanism is plausible for certain conditions, does it hold universally across different plant functional types?
How does this relationship perform in drought adapted ecosystems or for species with deep root systems that draw water well below the topsoil layer (i.e., deeper than the depth of 10 cm used here)?

For species with internal stem or leaf water storage or lower sensitivity to immediate soil water deficits, is $\text{LAI}/\text{WFC}$ still an adequate proxy?

In Section 4.4, I see mention of other indices including microwave data, but a brief discussion about the potential limitations of relying primarily on canopy leaf volume and shallow soil water parameters would help.

I am not familiar with the shrub types in France, so feel free to point me (and the manuscript's readers) to existing literature if such variability does not exist in France.

3. Derivation of the "Interest Score"
The "Interest Score" concept presented in Figure 10 is creative and practical for future sampling strategy design. However, the methodology behind it is insufficiently detailed.
How were the weighting coefficients ($\alpha$, $\beta$, and $\gamma$) determined . I only see in the supporting section 7 that they were "empirically set". What does that mean?

What are the exact values assigned to these coefficients?

Was any ground truth or validation heuristic used (e.g., constraining the score toward zero at existing, well sampled monitoring stations)?

4. Linking LFMC to Ignitions (Figure 9)
Figure 9 presents an intriguing relationship between simulated LFMC and ignition timing, but several critical details and controls are missing:
Dataset & Terminology: Please clarify what constitutes a "fire outbreak" versus an "ignition" (L. 380 vs. Figure 9 axis). Where did you obtain the fire outbreak from? Is this based on the Prométhée database?

Lack of Counterfactual Control: L. 379 states that this analysis tests whether simulated LFMC aligns with actual ignitions. However, because the evaluated ignitions occur during summer (June to September), LFMC naturally declines during this window. To prove that LFMC carries an ignition specific signal, I think a counterfactual baseline is required:

Suggested Test: For each ignition in Canton $A$ on Day $T$, select a paired control Canton $B$ that had no fire outbreaks on Day $T$ (or earlier in that fire season). Calculate the canton averaged LFMC for $T-10$ to $T+4$ days for Canton $B$. Repeat for each ignition and produce a similar box plot for the counter factual case. Does the non fire control curve look substantially different from Figure 9? If both curves look identical, the trend reflects seasonal drying rather than an ignition specific LFMC threshold.

Post Ignition LFMC Trajectory: Figure 9 shows that LFMC remains relatively flat following Day 0. In many burned landscapes, post fire burn scar pixels experience a sharp drop in visible canopy moisture. Is the lack of a post ignition drop an artifact of spatial averaging over an entire canton? Even with spatial averaging, one might expect at least a slight post fire decline relative to pre ignition levels..

Spatial Aggregation Level: Why were LFMC values aggregated across entire cantons rather than extracted at the specific $2\text{ km} \times 2\text{ km}$ ignition pixel or over a fixed spatial radius (e.g., 5 km)? Because cantons vary dramatically in size (from $4\text{ km}^2$ to $>1000\text{ km}^2$), this introduces spatial heterogeneity. Is the underlying ignition dataset restricted to canton level resolution?

As I write this, I acknowledge that this depth of analysis for an experiment that is not the central focus of the paper might be considered too much. But I think half-results are more dangerous that no results. Since there is already plenty of other contributions in the paper, I leave it to the authors to decide whether they include full details of the analysis and make the experiment more robust or leave it out from the paper and follow up with a future study.

Minor Comments
Figure 12: The current layout makes it difficult to quantitatively compare feature importances. While the larger size of the LAI rhombus implies higher importance, it is hard to discern the exact impact. For e.g., if one were introduced in learning how much performance metrics degrade when LAI/WFC is neutralized, how would they find out? Does the figure show that? I see the 1 - KGE = 1.0 for LAI/WFC. Does that mean upon neutralization, KGE becomes zero? Surely, that's not right?

L. 401: The text refers to "Supplementary Section 8" for the impact of the day of the year analysis, but this appears to correspond to Supplement 8 in the supplementary material. Please verify and correct the section reference.

L. 10 (Abstract): "The results demonstrate a strong ability to estimate the range and dynamics of LFMC across most of France" The term "strong ability" is subjective. Please replace or support this statement with concrete performance metrics such as RMSE or NSE.

L. 71 to 85: I appreciate the authors for including a thorough and clear description of the field sampling protocols. This was very helpful in understanding the labels for the neural net.

Krishna Rao

References
Martin-StPaul, N., Pimont, F., Dupuy, J.L. et al. Live fuel moisture content (LFMC) time series for multiple sites and species in the French Mediterranean area since 1996. Annals of Forest Science 75, 57 (2018). https://doi.org/10.1007/s13595-018-0729-3

Pimont, F., Ruffault, J., Martin-StPaul, N. K., & Dupuy, J. L. (2019). Why is the effect of live fuel moisture content on fire rate of spread underestimated in field experiments in shrublands?. International journal of wildland fire, 28(2), 127-137.

Yebra, M., Scortechini, G., Adeline, K. et al. Globe-LFMC 2.0, an enhanced and updated dataset for live fuel moisture content research. Sci Data 11, 332 (2024). https://doi.org/10.1038/s41597-024-03159-6

Reply
Citation: https://doi.org/10.5194/egusphere-2026-1247-RC1

Supplement

https://doi.org/10.5194/egusphere-2026-1247-supplement

Data sets

Supplementary data: Using artificial intelligence to monitor live fuel moisture content across France, based on a high-resolution land surface analysis Yann Baehr, Pierre Vanderbecken, Bertrand Bonan, Catherine Robert, Mathieu Regimbeau, François Pimont, Kevyn Raynal, Xiangzhuo Liu, Remi Savazzi, Moncef Garouani, Josiane Mothe, Nemesio Rodriguez-Fernandez, Lionel Jarlan, and Jean-Christophe Calvet https://zenodo.org/records/18850161

Model code and software

Viewed

Total article views: 476 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
343	106	27	476	52	31	25

HTML: 343
PDF: 106
XML: 27
Total: 476
Supplement: 52
BibTeX: 31
EndNote: 25

Views and downloads (calculated since 16 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	134	67	13	214
May 2026	133	20	3	156
Jun 2026	27	9	4	40
Jul 2026	49	10	7	66

Cumulative views and downloads (calculated since 16 Apr 2026)

Month	HTML	PDF	XML	Total
Apr 2026	134	67	13	214
May 2026	133	20	3	156
Jun 2026	27	9	4	40
Jul 2026	49	10	7	66

Viewed (geographical distribution)

Total article views: 462 (including HTML, PDF, and XML) Thereof 462 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Aug 2026

Short summary

This study produced vegetation water content maps to help manage the risk of forest fires in France. Artificial intelligence was used alongside land surface model outputs, satellite data and in situ observations to monitor vegetation stress in real time. The methodology tested has been shown to be robust and spatially consistent.


Total:	0
HTML:	0
PDF:	0
XML:	0