Seeing nutrients in a leaf: How hyperspectral AI reveals grapevine health

By analyzing how leaves interact with light across hundreds of wavelengths and learning shared patterns among nutrients and pigments, the approach delivers a fast, non-destructive alternative to labor-intensive tissue sampling.

Grapevines require balanced supplies of nutrients such as nitrogen, phosphorus, potassium, calcium, magnesium, and micronutrients to maintain healthy growth. Traditional tissue testing can identify deficiencies early, but it is costly, slow, and provides limited spatial resolution. Remote sensing has long promised a solution, yet many methods rely on simple vegetation indices or single-trait models that struggle with the complex, overlapping spectral signatures of plant traits. Advances in hyperspectral sensing and machine learning now make it possible to move beyond single indicators toward integrative models that reflect how plant traits co-vary and jointly influence leaf reflectance.

A study (DOI: 10.1016/j.plaphe.2025.100142) published in Plant Phenomics on 12 November 2025 by Alireza Pourreza’s team, University of California, points to a scalable pathway for more precise, timely, and spatially detailed nutrient monitoring in vineyards.

The study applied a stepwise modeling framework to improve hyperspectral estimation of grapevine leaf traits by first correcting model bias, then characterizing spectral–trait relationships, completing missing data, and finally comparing predictive strategies. Chlorophyll (Chl) estimates from the PROSPECT-PRO radiative transfer model were first validated against 327 fully labeled samples, revealing systematic overestimation (regression slope 1.33, NRMSE 0.41). A regression-based rescaling (multiplying estimates by 0.75) effectively corrected this bias, producing close agreement with measurements (NRMSE reduced to 0.17) while preserving realistic variability, with mean Chl values of 35.9 μg/cm² in the fully labeled set and 27.2 μg/cm² in the partially labeled set. Spectral importance analysis then identified key wavelength regions for different traits: nitrogen and chlorophyll were most sensitive in the visible range (around 450–550 nm) and near 2200 nm, whereas equivalent water thickness (EWT) and leaf mass per area (LMA) showed stronger responses in the shortwave infrared (around 1200, 1700, and 2200 nm). Correlation and spectral overlap analyses revealed strong inter-trait linkages, such as Mg–Ca (ρ = 0.69) and LMA–Ca (ρ = 0.68), negative relationships between K and both N and Chl, and substantial overlap of informative bands among nutrients and pigments, indicating shared spectral controls. Principal component analysis showed that the first two components explained 94.3% of total spectral variance, confirming strong low-dimensional structure across datasets. To address missing labels, a neural-network imputation model using 23 spectral PCs achieved high accuracy for P, Ca, and Mg (R² ≈ 0.72–0.78), but lower performance for Zn and Mn, reflecting weaker spectral signals. Final trait prediction comparisons demonstrated that a multi-trait CNN–LSTM model consistently outperformed single-trait models across most of 16 traits, with large gains for Mn (R² 0.30→0.62) and leaf structural parameter Nstruct (0.25→0.90) and lower prediction errors overall. Uncertainty analysis showed higher residuals for spectrally dissimilar samples, and filtering low-confidence imputations yielded a robust training set of 925 observations, supporting reliable multi-trait prediction.

The findings demonstrate that multi-trait hyperspectral modeling can deliver accurate, non-destructive assessments of grapevine nutrition at the leaf level. For vineyard managers, this opens the door to earlier detection of nutrient imbalances, more targeted fertilization, reduced input costs, and lower environmental risks. Beyond viticulture, the framework is adaptable to other crops where multiple physiological traits interact to shape spectral signals, supporting broader advances in precision agriculture and crop monitoring.

###

References

DOI

10.1016/j.plaphe.2025.100142

Original Source URl

https://doi.org/10.1016/j.plaphe.2025.100142

About Plant Phenomics

Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.