Hyperspectral remote sensing has been successfully used to retrieve parameters such as chlorophyll and nitrogen content. However, effective models for inverting dynamic physiological processes like gas exchange are lacking, and the predictive accuracy and applicable limits of such inversions remain unclear. In a controlled water and nitrogen stress experiment, 350–1150 nm leaf reflectance spectra were obtained along with simultaneous measurements of 12 physiological parameters. PLSR, RF, XGBoost, and LightGBM models were constructed, and SHAP was utilized for model interpretation. The results showed that predictive accuracies could be categorized into high, medium, and low tiers (R2 approximately 0.8, 0.5, and <0.3, respectively), corresponding to leaf water status and vapor pressure, E and Fm′, and gsw and Fs. LightGBM performed best for six high-tier water-related parameters (R2 = 0.75–0.81), while PLSR achieved the best performance for the medium-tier parameters E (R2 = 0.49) and Fm′ (R2 = 0.51). However, all models failed to predict gsw, suggesting that the relevant signal in the 350–1150 nm range is either absent or too weak to detect in our dataset. This study outlines the practical limits of estimating dynamic photosynthetic processes using VNIR spectra, offering a reference for future sensor configuration and model development.
Ren et al. (Wed,) studied this question.