NIR attribute selection for the development of vineyard water status predictive models

Near-Infrared spectroscopy (NIR) returns full spectra in the region between 750 and 2500 nm. Although a full spectrum provides extremely informative data, sometimes this enormous amount of detail is redundant and does not bring any additional information. In this work, different attribute selection methods for the development of vineyard water status predictive models are presented. Spectra from grapevine leaves were collected on-the-go (from a moving vehicle) along nine dates during the 2015 season in a commercial vineyard using a NIR spectrometer (1200–2100 nm). Contemporarily, the stem water potential (Ψ) was also measured in the monitored vines. A manual selection, based on Variable Importance in Projection scores (VIP scores) to choose the spectrum intervals including the most important wavelengths (interval selection), the locally most important wavelengths in the spectrum (peak selection), as well as the Interval Partial Least Squares (IPLS) were tested as attribute selection methods. The results obtained for the estimation of Ψ using the whole spectrum (R = 0.84, RMSEP = 0.167 MPa) were comparable to those yielded by the three attribute selection methods: the interval selection method (R = 0.80, RMSEP = 0.186 MPa), the peak selection method (R = 0.77, RMSEP = 0.201 MPa) and the IPLS (R ∼ 0.62–0.79, RMSEP ∼ 0.186–0.252 MPa). The highest simplification was provided by two IPLS models with three wavelengths and bandwidths of 20 and 4 nm that yielded R∼0.78 and RMSEP∼ 0.190 MPa. These results corroborate the suitability of a highly reduced selection of NIR wavelengths for the prediction of grapevine water status, and its utility to develop simpler multispectral devices for vineyard water status estimation.Grant PID2019-108330RA-I00 funded by MCIN/AEI/10.13039/501100011033 (Spain)