Rapid monitoring of grapevine reserves using ATR-FT-IR and chemometrics

Calibrations employing SVM-regression provided the optimum predictive models for nitrogen (R2 = 0.98 and RMSEP = 0.07%DW) compared to PLS regression (R2 = 0.97 and RMSEP = 0.08%DW). The best predictive models for starch was obtained using PLS regression (R2 = 0.95 and RSMEP = 1.43%DW) compared to SVR (R2 = 0.95; RMSEP = 1.56%DW). The RMSEP for both nitrogen and starch is below the reported seasonal flux for these analytes in Vitis vinifera. Nitrogen and starch concentrations in grapevine tissues can thus be accurately determined using ATR-FT-IR, providing a rapid method for monitoring vine reserve status under commercial grape production.Predictions of grapevine yield and the management of sugar accumulation and secondary metabolite production during berry ripening may be improved by monitoring nitrogen and starch reserves in the perennial parts of the vine. The standard method for determining nitrogen concentration in plant tissue is by combustion analysis, while enzymatic hydrolysis followed by glucose quantification is commonly used for starch. Attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FT-IR) combined with chemometric modelling offers a rapid means for the determination of a range of analytes in powdered or ground samples. ATR-FT-IR offers significant advantages over combustion or enzymatic analysis of samples due to the simplicity of instrument operation, reproducibility and speed of data collection. In the present investigation, 1880 root and wood samples were collected from Shiraz, Semillon and Riesling vineyards in Australia and Germany. Nitrogen and starch concentrations were determined using standard analytical methods, and ATR-FT- IR spectra collected for each sample using a Bruker Alpha instrument. Samples were randomly assigned to either calibration or test data sets representing two thirds and one third of the samples respectively. Signal preprocessing included extended multiplicative scatter correction for water and carbon dioxide vapour, standard normal variate scaling with second derivative and variable selection prior to regression. Excellent predictive models for percent dry weight (DW) of nitrogen (range 0.10'2.65%DW, median 0.45%DW) and starch (range 0.25'42.82%DW, median 7.77%DW) using partial least squares (PLS) or support vector machine (SVM) analysis for linear and nonlinear regression respectively, were constructed and cross validated with low root mean square errors of prediction (RMSEP).