Estimation of Live fuel moisture content via spectral–meteorological time-series fusion: a TCN–BiGRU deep learning framework

Live Fuel Moisture Content (LFMC) is a critical physiological indicator governing wildfire ignition probability and spread rate. To overcome the limitations of existing methods in capturing long-term dependencies and spatiotemporal dynamics, this study proposes a novel deep learning framework: the Temporal Convolutional-Bidirectional Gated Recurrent Unit network (TCN-BiGRU). The model uniquely integrates satellite spectral data from MODIS with high-resolution meteorological drivers from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) to reconstruct daily LFMC dynamics without relying on prior land cover classification. By leveraging a 365-day historical time series, the TCN-BiGRU effectively captures the cumulative effects of weather and phenology. The model was rigorously trained and validated using the Global-LFMC database across the Contiguous United States (2002–2018). Results demonstrate superior predictive performance, achieving an overall correlation coefficient (R) of 0.75 and a Root Mean Square Error (RMSE) of 25.02% in independent temporal validation (Scenario a). Furthermore, the model exhibited robust fire risk stratification capability, classifying high-fire-danger conditions with accuracies of 75.4% for forests, 79.8% for shrublands, and 82.5% for grasslands. These findings confirm that the proposed multi-source deep learning approach offers a generalized, high-precision solution for large-scale operational LFMC monitoring and wildfire risk management.