Using Multi-Angular Spectral Reflection of Dorsiventral Leaves to Improve the Transferability of PLSR Models for Estimating Leaf Biochemical Traits

Leaf biochemical traits are crucial for understanding plant physiological status and ecological dynamics. Partial least squares regression (PLSR) models have been widely used to estimate leaf biochemical traits from spectral reflectance information. However, variations in sun–sensor geometry, the sensor field of view, and the random orientation of leaves can introduce multi-angular reflection properties that differ between leaf sides. In this context, the transferability of PLSR models across different leaf sides and viewing zenith angles (VZAs) remains unclear. This study investigated the potential of multi-angular spectral reflection from dorsiventral leaves to improve the transferability of PLSR models for estimating the leaf chlorophyll content (LCC) and equivalent water thickness (EWT). We compared models trained using multi-angular data from both leaf sides with models trained using nadir data (from the adaxial side, abaxial side, or their combination). The results show that the PLSR models trained with multi-angular data from both leaf sides outperformed the models trained with nadir data, achieving the highest accuracy in estimating biochemical traits (LCC: R2 = 0.87, RMSE = 7.17 μg/cm2, NRMSE = 10.71%; EWT: R2 = 0.86, RMSE = 0.0015 g/cm2, NRMSE = 10.00%). In contrast, the PLSR models trained using single-angle reflection from either the adaxial or abaxial side showed a lower estimation accuracy and greater variability across leaf sides and VZAs. The superior performance across datasets obtained under different measurement conditions (e.g., integrating spheres and leaf clips) further confirmed the improved generalizability of the PLSR model trained with multi-angular data from dorsiventral leaves. These findings highlight the potential of the multi-angular spectral reflection of dorsiventral leaves to enhance the estimation of biochemical traits across various leaf sides, viewing angles, and measurement conditions. They also underscore the importance of incorporating spectral diversity into model training for improved transferability.