Abstract
<jats:p>Live fuel moisture content (LFMC) is a key determinant of fire behavior, but measurements are sparse. Machine learning can address this by creating wall-to-wall LFMC maps from remote sensing. However, LFMC models face a tradeoff between performance and record length. Microwave synthetic aperture radar (SAR)-informed models achieve strong performance but are limited by short and sparse SAR records. Alternatively, algorithms using long-term optical and meteorological datasets can support multi-decadal mapping, but generally perform worse. Here, we combine these strengths in a machine learning framework that fuses optical, hydroclimatic, land cover, and static predictors while using Sentinel-1 SAR backscatter observations as auxiliary training targets. This allows SAR to shape the learned LFMC representation without being required to make LFMC predictions. The model uses separate input branches for optical imagery, hydroclimatic history, and static information, then combines features with a transformer to predict LFMC. Using this framework, we produced daily, 500 m LFMC estimates across the western United States since 2001. Under spatial cross-validation, the model reproduced LFMC measurements with R2 = 0.65, RMSE = 26.2%, and bias = -1.65%. The model captured anomalies from the site mean better than the site mean itself (R2 = 0.67 and R2 = 0.56, respectively), indicating skill at predicting temporal variability. Performance varied by land cover, with strong skill in shrublands (R2 = 0.72) and weaker skill in evergreen forests (R2 = 0.18). These results demonstrate that combining long-term predictors with short-term observational datasets used as auxiliary training data can improve long-term geospatial products.</jats:p>