Forest surface fuels are a key factor influencing fire behavior, and their spatial heterogeneity is strongly affected by forest type, climate, and topography. In this study, we innovatively introduced Active Learning into the surface fuel estimation framework. By integrating multi-source satellite remote sensing data—including optical spectral features, radar-derived variables, and ancillary topographic and climatic data—we constructed Extreme Gradient Boosting (XGBoost) and Support Vector Regression (SVR) models to estimate four classes of surface fuels (1h, 10h, herbaceous, and shrub) in forest ecosystems along the China–Mongolia border. Model training was first conducted using Stratified Ten-fold Cross-Validation (ST10CV), followed by optimization through the Euclidean-Based Diversity Active Learning (EBD-AL) strategy. The results showed that the XGB model optimized with EBD better explained the variability of dead 1h fuels, with the mean R 2 increasing from 0.20 to 0.53. The EBD optimization yielded relatively weak improvements for 10h dead woody fuels. For herbaceous fuels, the mean R 2 of the XGB-EBD model increased from 0.20 to 0.50. For shrub fuels, the XGB-EBD optimization improved the mean R 2 from 0.07 to 0.42, indicating a substantial enhancement in the model’s explanatory power. These findings demonstrate that combining multi-source remote sensing data with Active Learning in a Machine Learning framework can significantly improve the spatial predictive accuracy of forest surface fuel load, providing reliable data support for forest fire risk assessment and management. • Multi-source remote sensing and field measurements mapped spatial distributions of fuels. • ST10CV was used to reduce estimation bias in fuel regression tasks. • EBD-AL active learning optimized sample selection for surface fuel estimation. • Active learning substantially improved model performance for all fuel categories.
Hu et al. (2026) studied this question.