Decomposing Illumination-Induced Variability in UAV Hyperspectral Canopy Reflectance Improves Leaf Chlorophyll Content Estimation

Abstract

Accurate estimation of leaf chlorophyll content (LCC) from centimeter-level uncrewed aerial vehicle (UAV) hyperspectral imagery is often challenged by illumination-induced variability within crop canopies, particularly in row crops where directional illumination interacts with complex canopy architecture. This variability reshapes the distribution of within-canopy reflectance and introduces spectral mixing that confounds biophysical retrieval. This study addresses this issue by decomposing the within-canopy reflectance distribution using a Gaussian mixture model (GMM), enabling the separation of illumination-consistent reflectance components under heterogeneous lighting conditions. The component with the highest mean reflectance was selected to represent canopy spectra less affected by shadow-induced mixing. LCC was estimated using vegetation indices (VIs), an inverted radiative transfer model (PROSAIL), and partial least squares regression (PLSR). The results show that the selected high-reflectance component consistently improves LCC estimation relative to canopy-averaged reflectance. In PROSAIL-based inversion, the coefficient of determination (<inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula>) increased from 0.67 to 0.84, while relative RMSE decreased from 16.91% to 10.17% (a 40% reduction). Across six VIs, the selected component improved VI-LCC relationships relative to binary sunlit-shaded thresholding (<inline-formula> <tex-math notation="LaTeX">$\Delta R^{2} \approx ~0.051$ </tex-math></inline-formula>–0.118) and increased <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> by 0.063–0.177 compared with using all vegetation pixels. In data-driven PLSR modeling, the component-based input raised the validation <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> from 0.46 to 0.56 and reduced RMSE from 2.23 to <inline-formula> <tex-math notation="LaTeX">$2.03~\mu $ </tex-math></inline-formula>g<inline-formula> <tex-math notation="LaTeX">$\cdot $ </tex-math></inline-formula>cm−2 (double-split means), with consistent improvements confirmed under repeated double-split validation and predicted residual error sum of squares (PRESS)-based component selection. These results indicate that decomposing illumination-induced reflectance variability enhances the robustness of LCC estimation from UAV hyperspectral canopy reflectance under heterogeneous illumination conditions.