English Abstract
A rapid and nondestructive assessment of photosynthetic capacity is valuable for physiological phenotyping and has potential applications in precision crop management. Hyperspectral sensing combined with machine learning provides an efficient approach for estimating leaf physiological traits, yet most existing methods rely on single observations and rarely exploit temporal information contained in sequential measurements. In this study, we developed a trajectory-informed hyperspectral modeling framework for estimating the maximum Rubisco carboxylation rate standardized to 25 °C (Vcmax,25), a key indicator of leaf photosynthetic capacity. The analysis comprised 408 date-level observations derived from 1149 eligible leaf-level measurements from 32 species across 10 sites, organized into intra-seasonal spectral sequences. Conventional snapshot models based on partial least squares regression and support vector regression (SVR) were compared with temporal SVR models incorporating day of year, principal component analysis (PCA)-derived descriptors of recent spectral trajectories, or both. Incorporating spectral trajectory information produced the best performance (R2 = 0.806, RMSE = 15.244 μmol m−2 s−1), compared with R2 = 0.740 and RMSE = 17.649 μmol m−2 s−1 for the snapshot SVR model. Day of year alone yielded only a limited improvement and did not further improve the trajectory-based model. A temporal feature misalignment experiment reduced mean R2 to 0.749 and increased mean RMSE to 17.310 μmol m−2 s−1, indicating that the predictive gain depended on preserving the correspondence between each observation and its temporal context rather than simply increasing predictor dimensionality. SHAP analysis further showed that instantaneous spectral information, particularly in the red-edge and shortwave-infrared regions, remained the dominant source of prediction, while temporal descriptors provided complementary information on recent spectral change. These results indicate that incorporating temporal spectral trajectories can extend conventional snapshot-based machine learning and provide a useful strategy for physiological phenotyping of leaf photosynthetic capacity across the growing season.