Estimating Leaf Chlorophyll Fluorescence Parameters Using Partial Least Squares Regression with Fractional-Order Derivative Spectra and Effective Feature Selection

Chlorophyll fluorescence (ChlF) parameters serve as non-destructive indicators of vegetation photosynthetic function and are widely used as key input parameters in photosynthesis–fluorescence models. The rapid acquisition of the spatiotemporal dynamics of ChlF parameters is crucial for enhancing remote sensing applications and improving carbon cycle modeling. While hyperspectral reflectance offers a promising data source for estimating ChlF parameters, previous studies have relied primarily on spectral indices derived from specific datasets, which often lack robustness. In this study, we simultaneously monitored ChlF parameters and spectral reflectance in leaves from different species, growth stages, and canopy positions within a temperate deciduous forest. We developed a data-driven partial least squares regression (PLSR) model by integrating fractional-order derivative (FOD) spectral transformation with multiple feature selection methods to predict ChlF parameters. The results demonstrated that FOD spectra effectively improved prediction accuracy compared to conventional PLSR attempts. Among the feature selection algorithms, the least absolute shrinkage and selection operator (LASSO) and stepwise regression (Stepwise) methods outperformed others. Furthermore, the LASSO-based PLSR model that used low-order (<1) FOD spectra achieved high predictive performance for NPQ (R2 = 0.60, RPD = 1.60, NRMSE = 0.16), ΦP (R2 = 0.73, RPD = 1.94, NRMSE = 0.11), ΦN (R2 = 0.62, RPD = 1.62, NRMSE = 0.12), and ΦF (R2 = 0.54, RPD = 1.48, NRMSE = 0.15). These findings suggest that the integration of FOD spectral transformation and appropriate feature selection enables the simultaneous estimation of multiple ChlF parameters, providing valuable insights for the retrieval of ChlF parameters from hyperspectral data.