A source-side trait-mediated framework for dynamic maize yield prediction using UAV multispectral imagery

Dynamic in-season maize yield prediction is essential for cultivation management and precision decision-making. However, traditional yield prediction models based on vegetation indices often lack physiological interpretability, and integrating multitemporal dynamic information within a unified framework remains challenging. This study proposes a source-side trait-mediated approach for dynamic maize yield prediction using UAV multispectral imagery. Leaf area index (LAI), leaf chlorophyll content (SPAD), and aboveground biomass (AGB) were used as intermediate variables. A cascaded framework linking canopy spectra, source-side agronomic traits, and yield was constructed to improve agronomic interpretability. In addition, the temporal dynamic factor (TDF) was introduced to model stage-dependent trait-yield relationships, enabling dynamic yield prediction. The results showed that increasing planting density significantly enhanced LAI and AGB, whereas yield first increased with planting density and then decreased slightly. For agronomic trait prediction, XGBoost generally outperformed RF. In the independent multi-cultivar test, the XGBoost models achieved coefficients of determination (R 2 ) of 0.8185, 0.8758, and 0.8113 for LAI, SPAD, and AGB, respectively. The TDF fitting results for the 2025 dataset showed that the linear relationship between yield and the Comprehensive Yield Index (CYI) had R 2 values ranging from 0.67 to 0.86 and RMSE values from 0.428 to 0.648 t/ha across growth stages. In the 2024 interannual independent test, the model maintained an R 2 of 0.7803 and an RMSE of 0.6076 t/ha. Furthermore, in a multi-cultivar scenario, yield prediction achieved an R 2 of 0.6824 and an RMSE of 0.3746 t/ha. The proposed framework improved the agronomic interpretability and interannual and cross-cultivar applicability of maize yield prediction using UAV-based multitemporal imagery.