New ML framework fuses satellite data to predict wildfire fuel moisture with 71% accuracy

A team led by Valerio Pampanoni from the University of Alcalá has published a paper on arXiv (2607.16714) detailing a unified multisensor machine-learning framework for retrieving live fuel moisture content (LFMC), a critical variable for wildfire management. The framework integrates data from Terra and Aqua MODIS, VIIRS, Landsat 8/9, Sentinel-2, and Sentinel-3, along with meteorological variables, topography, and cyclic day-of-year predictors. To harmonize across different optical sensors, the authors restrict predictors to a common red, near-infrared, and shortwave-infrared feature space and perform empirical reflectance calibration toward a Sentinel-2 reference domain.

The models were trained separately for grass, shrub, and tree land cover types using Random Forest and XGBoost regressors. Under a stringent validation design that withholds observation dates from training sites, the best models achieved pooled R² values of 0.715, 0.693, and 0.700 for grass, shrub, and tree, respectively. Preliminary experiments showed that weather, topography, and seasonal predictors provide most of the predictive gain beyond vegetation indices alone. The framework is designed to incorporate additional optical sensors as long as compatible reflectance bands and sufficient overlap observations are available. The study, submitted to JSTARS, represents a significant step toward real-time, large-scale LFMC estimation for fire danger assessment.