Grapevine nitrogen retrieval by hyperspectral sensing at the leaf and canopy level

Abstract

Context and purpose of the study – Grapevine nitrogen (N) monitoring is essential for efficient N management plans that optimize fruit yield and quality while reducing fertilizer costs and the risk of environmental contamination. Unlike traditional vegetative-tissue sampling methods, remote sensing technologies, including hyperspectral imaging, have the potential to allow monitoring of the N status of entire vineyards at a per-vine resolution. However, differential N partitioning, variable spectral properties, and complex canopy structures hinder the development of a robust N retrieval algorithm. The present study aimed to establish a solid understanding of vine spectroscopic response at leaf and canopy levels by evaluating the different nitrogen retrieval approaches, including the radiative transfer model.

Material and methods – At the leaf level, N content and its relative position within a shoot were measured along with the proximal hyperspectral reflectance (350nm-2500nm) from ‘Flame Seedless’ vines grown in pots as well as ‘Solbrio’ vines in a vineyard. At the canopy level, leaf nitrogen concentration, and hyperspectral images (400nm-1000nm) of ‘Valley Pearl’ vines were collected using a hyperspectral camera mounted on an uncrewed aerial vehicle. At leaf and canopy levels, we evaluated the N retrieval performance of several spectral analytics approaches, including empirical data-driven models, a physical-based model (radiative transfer model), and hybrid models. 

Results – At the leaf level, the performance of data-driven approaches using the entire 350-2500 nm spectrum (chemometrics and machine learning) outperformed (R²=0.76-0.78) the use of vegetation index, physical-based modeling, and hybrid approaches. However, collecting and analyzing hyperspectral data within visible, near-infrared, and shortwave infrared is unrealistic for large-scale monitoring. Protein, one of the variables retrieved by a physical-based approach, showed high potential to be used as a predictor of N content because protein, unlike chlorophyll, remained consistently correlated with N content regardless of leaf age. At the canopy level, the performance of data-driven and hybrid approaches was competitive (R²=0.61-0.69) except for the combination of physical-based parameters and random forest regression (R²=0.50). However, the performance of N content retrieval models varies widely across datasets, and it is not yet clear what factors determine the performance of models. Further data processing and calibration to extract more reliable spectral features from hyperspectral images are required to scale N retrieval from the leaf level to the canopy level by leveraging the knowledge acquired at the leaf level analysis.