Assessing the feasibility of up-scaling grapevine physiological trait estimation using multispectral UAV imagery

Scalable grapevine physiological assessment is relevant for precision viticulture, but direct physiological measurements are labor-intensive and leaf-specific. This study evaluated whether a machine-learning framework developed for leaf-level hyperspectral measurements can be applied to UAV-derived multispectral imagery. In situ hyperspectral observations from 2024 and 2025 were combined and restricted to the visible and near-infrared (VNIR) range from 350 to 1000 nm to align with common UAV spectral coverage and assess multi-temporal performance and spectral resampling. UAV-derived multispectral observations from 2025 were analyzed for sunlit, shaded, and combined canopy extraction strategies. First-derivative preprocessing improved in situ performance, yielding best R² values of 0.811 for net assimilation rate A , 0.792 for electron transport rate ETR, and 0.890 for effective quantum yield of photosystem II Φ PSII . Moderate spectral resampling caused limited performance losses for A and ETR. UAV-based models yielded R² up to 0.765 for A , 0.768 for ETR, and 0.901 for Φ PSII . Shaded and combined observations improved A and ETR predictions relative to sunlit observations, whereas Φ PSII remained robust across extraction strategies. Plant-grouped validation supported generalization to unseen vines, whereas leave-one-date-out validation failed to demonstrate robust generalization to unseen measurement dates, highlighting temporal generalization as a key limitation of the UAV-based approach.