UAV-based aboveground biomass estimation via trait-mediated pathways in a cultivated Leymus chinensis grassland

Aboveground biomass (AGB) in cultivated Leymus chinensis grasslands is strongly influenced by irrigation, nitrogen fertilization, and mowing, yet many UAV-based AGB models rely mainly on spectral indices and random data splits, which can overestimate generalization under spatiotemporal dependence. Here we test whether adding management information and ground-measured structural traits improves UAV-informed AGB estimation in a plot-based, management-intensive system. Using a 3-year factorial experiment (12 water-nitrogen-mowing treatments) with UAV multispectral imagery, we built LightGBM models integrating spectral indices, management factors, and structural traits. A plot- and year-independent, target-optimized split was used to balance AGB and treatment distributions between training and test data. Mixed-effects models and structural equation modeling were used to quantify management interactions and trait-mediated pathways. Nitrogen fertilization increased AGB by 40-80%, while frequent mowing weakened the synergistic effect of irrigation and nitrogen. The best model achieved R² = 0.73 on the fixed test set; external validation performance declined (temporal R² = 0.54; spatial transferability CV R² = 0.56) when key structural traits (canopy height and leaf area index) were unavailable, highlighting that transferability depends on feature availability. Structural traits contributed 52% of total importance, management main effects 24%, and spectral indices 20%. These results support management-relevant AGB monitoring in cultivated grasslands while clarifying current scalability limits.