TY - JOUR T1 - Visible-Near Infrared Reflectance Spectroscopy for Nondestructive Analysis of Red Winegrapes JF - American Journal of Enology and Viticulture JO - Am. J. Enol. Vitic. DO - 10.5344/ajev.2015.15035 SP - ajev.2015.15035 AU - Michael Fadock AU - Ralph B. Brown AU - Andrew G. Reynolds Y1 - 2015/09/03 UR - http://www.ajevonline.org/content/early/2015/08/31/ajev.2015.15035.abstract N2 - Composite samples of intact grape berries were collected weekly from veraison until harvest. Each sample comprised ≈400 berries selected following the pre-harvest row sampling protocol specified by the vineyard manager. The grape cultivars, and corresponding number of samples of each collected in 2009 and 2010 were as follows: Cabernet Sauvignon (43, 36), Cabernet franc (83, 80), and Syrah (38, 36). Reflectance spectra for the composite samples in a wavelength range of 350–850 nm were collected with a diode array spectrometer. Chemical analyses for soluble solids content (Brix), pH, titratable acidity (TA), total phenols, and total anthocyanins were carried out for all samples. Chemometric calibrations for corresponding reflectance data were developed using trained partial least squares regression (PLSR) models with several preprocessing methods (smoothing, normalization, differentiation) and subject to variable selection by recursive feature elimination. Trained models were externally validated with data from the alternate year. Best performing models for Brix, pH, TA, phenols, and anthocyanins in 2009 had root mean square errors (RMSEP) of 0.65, 0.05, 0.59 g/L, 31.2 mg/L and 75 mg/L, respectively, with corresponding R2 values of 0.84, 0.58, 0.56, 0.27 and 0.65. The best 2010 models had RMSEP of 0.65, 0.05, 0.86 g/L, 27.9 mg/L and 111 mg/L, and R2 values of 0.89, 0.81, 0.58, 0.25 and 0.17. The 2009 calibrations were used for estimating Brix and pH from spectral data of the samples collected in the next growing season and yielded RMSEP performance of 0.87 and 0.05 with R2 values of 0.71 and 0.56 respectively. Principal component analysis decomposition of 2009 and 2010 reflectance data showed similarities in the resultant loadings, indicating a similar underlying data structure. ER -