UAV-based temporal synergistic estimation of multiple alfalfa qualities integrating physics-informed network and 3D...

Accurately monitoring alfalfa nutritional quality is essential for optimal pasture management. Yet, current UAV remote sensing methods rely on single-temporal imagery and empirical indices, limiting their ability to handle multi-stage growth dynamics, canopy spectral saturation, and canopy-to-whole-plant scale differences. Furthermore, small sample sizes often cause purely data-driven models to overfit correlations, yielding biologically unrealistic results. Overcoming these challenges, we designed a comprehensive quality estimation framework using 127 alfalfa core germplasms, combining high-dimensional spectral mining, a physics-informed network, and a 3D allometric transfer operator. After screening 14,960 spectral operators across original and log-transformed spaces, we applied a dual dimensionality reduction strategy to isolate optimal features. Four-band dual-difference structures proved highly sensitive to fiber components (ADF/NDF, |r| = 0.896), while logarithmic decoupling operators accurately isolated protein and nitrogen signals (CP/N, |r| = 0.868). We then engineered a Physics-Informed Sparse Shallow Network (PI-SSN). By leveraging temporal attention decoupling, it adaptively assigns growth-stage weights to different components and uses carbon-nitrogen metabolic constraints to maintain biological accuracy during multi-task retrieval. Multi-stage temporal data significantly boosted accuracy over single-period spectra. PI-SSN delivered exceptional test set coefficients of determination ( $R^2$ ) of 0.812-0.848 and RPDs >2.0 for N, CP, ADF, and NDF, easily outperforming standard baselines. To bridge the canopy-only observation gap, we introduced a 3D allometric transfer operator that incorporates canopy coverage and plant height. This effectively corrected vertical stem-leaf observation biases, enhancing Relative Feed Value (RFV) predictions. Ultimately, this approach offers a powerful new framework for high-throughput forage phenotyping.