Precision viticulture is increasingly being applied to automate and optimize grape production in the vineyard. This paper describes the development of a method for automatic selection of regions of interest from hyperspectral images obtained of a row of vines and intended for prediction of soluble solids content. For this purpose, a dataset consisting of hyperspectral images of a row of ‘Sangiovese’ wine grapes was adopted. Hyperspectral images were acquired directly in the field by means of a hyperspectral imaging Vis/NIR system (400–1000 nm) mounted on a ground-based vehicle. The analyses were carried out on 17 different days, under clear or partly cloudy conditions, in the period between post-veraison and harvest. The vineyard row of Sangiovese vines was divided into 11 sections and a hyperspectral image for each section for each day of analysis was acquired. The regions of interest of the hyperspectral images, comprising the areas representing the grapes, were selected using a PLS-DA-based method. The best PLS-DA model provided excellent results, with sensitivity and specificity values of 0.991 and 0.996, respectively. The mean spectra of the selected regions of interest (ROI) were finally used to predict the soluble solids content (SSC) of the grapes by PLS regression to a primary reference analysis. The results of SSC predictions using the automatic selection of ROIs (R²_CV = 0.74 and RMSECV = 0.86 °Brix) were on par with similar regression based on carefully manual selection of ROIs (R²_CV = 0.73 and RMSECV = 0.87 °Brix).