This study aims to systematically elucidate the distribution patterns and main influencing factors of fuel loads in three fuel types (herb layer, litter layer, and humus layer), construct high-precision prediction models for fuel load, and to accurately identify high-fire-risk areas in Chinese fir forest regions and develop differentiated forest fire management strategies.

Based on 702 standard sample plots of Chinese fir forests in Hunan Province, surveyed from 2020 to 2022. Fifteen factors, including aspect, slope position, slope gradient, and canopy closure were selected. One-way analysis of variance was used to examine the distribution characteristics of fuel loads among the three fuel types, and Pearson correlation analysis was applied to investigate the correlations between 15 factors and 3 types of fuel loading. Prediction models for fuel load, including linear regression, random forest, gradient boosting, and XGBoost, were developed using Python 3.8.

1) The fuel loads in the litter layer (0.615–21.110 t/hm²) and humus layer (0–44.210 t/hm²) of Cunninghamia lanceolata forests in Hunan Province exhibited extremely wide ranges. The mean fuel loads, in descending order, were: humus layer (11.647 t/hm²), litter layer (6.407 t/hm²), and herb layer (0.638 t/hm²), with significant differences among the three (P<0.01); 2) The fuel load of the herb layer was primarily driven positively by herb cover (r=0.673) and significantly suppressed by canopy closure (r=-0.290) and stand density (r=-0.240). For the litter layer, litter thickness was the core influencing factor (r=0.657), while elevation (r=0.098) and slope position (r=-0.119) indirectly regulated its accumulation by affecting decomposition rates. The fuel load of the humus layer was significantly positively correlated with humus thickness (r=0.780), elevation (r=0.167), and total growing stock (r=0.180); 3) The gradient boosting model performed best for the herb layer (R²=0.682 6) and litter layer (R²=0.751 8), while the XGBoost model achieved the best performance for predicting the humus layer fuel load (R²=0.665 5). The machine learning models (random forest, gradient boosting, and XGBoost) significantly outperformed the traditional linear regression model and demonstrated good generalization capability.

1) The extremely high fuel loads in the litter and humus layers of some C. lanceolata forest stands in Hunan Province pose a significant risk for forest fire occurrence and rapid spread; 2) Canopy closure has opposite effects on the herb layer (highly significant negative correlation) and the litter layer (highly significant positive correlation); therefore, when implementing thinning operations, managers should balance the fire risks of both layers, and promptly remove logging residues after thinning to avoid a sharp increase in litter input. In addition, litter decomposition is slower and accumulation is greater at high-elevation upper slope positions; thus, prescribed burning or mechanical removal of thick litter layers should be carried out before the dry winter-spring season. Moreover, some stands have developed deep humus layers, and areas with thick humus accumulation should be prioritized for ground-fire prevention and regularly inspected; 3) The gradient boosting and XGBoost prediction models exhibit high accuracy and strong stability, providing technical support for dynamic fuel monitoring, precise fire-risk zoning, and scientific fire prevention decision-making in C. lanceolata forest regions.