A dual-branch perception and hybrid attention integrated framework for temporal remote estimation of wheat leaf biomass.

Introduction: Wheat leaf biomass is a key indicator of crop growth, nitrogen status, and yield potential, and its accurate estimation is essential for precision agriculture. Unmanned aerial vehicle (UAV) remote sensing provides multi-stage phenological observations for non-destructive biomass monitoring. However, existing approaches often fail to capture the superimposed temporal patterns inherent to crop phenology, including short-term physiological fluctuations driven by management events and long-term seasonal growth trends, as well as the cumulative causal effects of early-stage conditions on final biomass accumulation. Methods: This study proposed a dual-branch perception and hybrid attention integrated framework (DBAFN) for temporal estimation of wheat leaf biomass from UAV multi-temporal observations across key growth stages. Results and discussion: Experimental results demonstrated that the DBAFN achieved the best performance, with the coefficient of determination (R²) of 0.87, root mean square error (RMSE) of 38.41 g/m², mean absolute error (MAE) of 27.77 g/m², and relative RMSE (RRMSE) of 17.29%. Overall, the proposed framework provided an effective solution for temporal biomass estimation and demonstrated strong generalization capability, as further validated by independent experiments across different ecological regions and wheat genotypes (R² = 0.816-0.820). Compared with conventional machine learning models, the DBAFN showed consistently higher accuracy and lower prediction error. Multi-source feature analysis indicated that the combination of reflectance, vegetation indices, and canopy height provides the most accurate estimation. Ablation experiments further confirmed the effectiveness of each module in improving model performance. The SHapley Additive exPlanations (SHAP) analysis revealed that the canopy height and key spectral features contribute most to biomass prediction, highlighting the importance of integrating structural and physiological information. This study demonstrates that integrating multi-scale temporal dynamics, hybrid attention mechanisms, and transformer-based dependency modeling significantly improves the reliability of UAV-based biomass estimation. It offers a practical, data-driven pathway for intelligent crop monitoring and precision nitrogen management.