Drought indices predict changes in live fuel moisture content (LFMC) across French Mediterranean shrublands: A multisource and machine learning approach
| Authors: | Xiangzhuo Liu, Nicolas Martin-StPaul, Julien Ruffault, Gengke Lai, Kevyn Raynal, RodriguezSuquet Raquel, Huan Wang, Russell A. Parsons, Jean-Luc Dupuy, Francois Pimont |
| Year: | 2026 |
| Type: | Scientific Journal |
| Station: | Rocky Mountain Research Station |
| DOI: | https://doi.org/10.1016/j.agrformet.2026.111093 |
| Source: | Agricultural and Forest Meteorology |
Abstract
The French Mediterranean region is among the most wildfire-prone areas in Europe, and fire activity is projected to intensify further under climate change. Live fuel moisture content (LFMC), a key driver of fire behavior, is critical for fire danger monitoring. However, the scarcity of high-quality LFMC data in this region hampers understanding of wildfire drivers and limits robust fire danger assessment. To address this gap, we developed an LFMC prediction model for the French Mediterranean fire season using the extreme gradient boosting (XGB) algorithm, hereafter referred to as the LFMC-XGB model. The input variables of the LFMC-XGB model were selected from the remote sensing (both optical and microwave) and weather-related variables using the leave-one-site-out cross-validation method. The results showed that the best performance was achieved by combining the Drought Code and Duff Moisture Code calculated from SAFRAN weather reanalysis data, the root zone soil moisture from the Soil Moisture Active Passive product, and the non-tree percentage from the Moderate Resolution Imaging Spectroradiometer product. When evaluated against in-situ LFMC measurements, the LFMC-XGB model yielded an R² of 0.56 with an RMSE of 13.95 %. The model also performed well in simulating site temporal fluctuations (R2 = 0.58 and RMSE = 12.21 %) and site means (R2 = 0.52 and RMSE = 6.77 %). Shapley additive explanation (SHAP) analysis identified the Drought Code and Duff Moisture Code as the most influential predictors, exhibiting monotonically decreasing contributions with increasing LFMC. Non-tree cover and root zone soil moisture followed, showing stepwise response patterns. Using the LFMC-XGB model, we generated daily LFMC maps at 250 m resolution for 2016–2022. The analysis of LFMC maps highlighted the potential of the LFMC-XGB model to strengthen operational fire danger prediction and improve understanding of fire regime drivers in the French Mediterranean.