Physiology-informed hyperspectral retrieval of leaf Vcmax across wheat and maize

Leaf hyperspectral reflectance can provide a scalable way to estimate photosynthetic capacity (Vcmax), but models trained in one species or measurement context often lose accuracy in another. This transfer problem limits the use of spectral approaches in multi-species crop phenotyping and carbon-cycle applications. Here, we tested physiology-informed inputs for leaf-level Vcmax25 retrieval using paired gas-exchange and reflectance data from wheat (C₃; n = 198) and maize (C₄; n = 81) grown under contrasting nitrogen supply. The four input configurations were raw spectra (Mod1), spectra scaled by a PPFD–absorptance proxy (Mod2), scaled spectra augmented with radiative-transfer-derived traits (Mod3), and scaled spectra augmented with a spectral coordination proxy (Mod4). Within datasets, the best models reached R² = 0.82 in wheat, 0.41 in maize, and 0.76 in the combined dataset. In a matched comparison with a common random-forest learner, the spectral coordination proxy Mod4 improved accuracy only slightly over Mod2 in wheat (RMSE −0.51%; p = 0.0058) and maize (RMSE −1.26%; p = 0.0011) but not in the combined dataset (RMSE −0.15%; p = 0.074), and the trait-based Mod3 showed no consistent benefit. When wheat models were tested on measurement dates not used in training, accuracy remained moderate (R² = 0.563; RMSE = 15.07 µmol m⁻² s⁻¹). Despite this within-dataset performance, models applied to the other species without calibration failed in both directions (negative R²), and adding source-species data did not improve prediction even when a few samples of the new species were used for calibration. These results show that physiology-informed input design provides at most small within-dataset gains, and that reliable prediction across C₃ and C₄ crops requires calibration data from the target crop.