Nitrogen retrieval in grapevine (Vitis vinifera L.) leaves by hyperspectral sensing

Abstract

Nitrogen (N) is a critical macronutrient that directly affects grapevine yield and quality but is also highly mobile in soil and can cause environmental contamination when over-applied. Monitoring leaf N content is useful for assessing vine N status and thereby improving N management plans, and the ability to assess N by remote sensing is desirable. Remote sensing technology has been utilized for retrieving N in agronomic crops since the 1990s. However, remote sensing has not been widely adopted for N management mainly due to variability in spectral patterns that emerge when training a model on one dataset and then applying it to a different dataset. Differences in environmental conditions, phenological stages, and plant varieties contribute to variability. This study aimed to understand the major factors limiting the generalizability of N-level retrieval algorithms in grapevine: the impact of leaf age and how calibrating nitrogen content, using either leaf dry mass (Nmass) or leaf area (Narea), influence the outcomes. This study also compares the performance among five major N retrieval approaches. We used spectral data from a hand-held hyperspectral sensor measuring wavelengths from 350 to 2500 nm from 664 individual leaf samples obtained from two grapevine varieties on three different sampling campaigns. Our findings indicate that while machine learning and chemometric approaches offer high accuracy levels (R2 = 0.78), the physical modeling approach utilizing the PROSPECT radiative transfer model (RTM) is capable of consistently retrieving N independent of specific dataset conditions and with an acceptable level of accuracy (R2 = 0.45) while using only 50% of the hyperspectral data. The estimation of Nmass using Proteinmass, which was calculated from Proteinarea retrieved by RTM and ground truth LMA, demonstrates a high potential for consistent retrieval of Total N (TN) in grapes. We also found that while calibrating a model by Narea can eliminate the negative impact of leaf age on prediction accuracy, Nmass still performed better in four out of five approaches tested. This study shows that a hybrid model that combines machine learning and RTM can alleviate the impact of leaf age on prediction accuracy while reducing 50% of spectral bands required and maintaining an acceptable level of accuracy (R2 = 0.54).