To address the limitations of current methods for detecting physicochemical indicators of wine grapes, such as sample destruction, time-consuming, energy-consuming, and not suitable for large-scale detection, this study used three varieties (‘Dunkelfelder’, ‘Marselan’, and ‘Petit Verdot’) with different ripening times to establish prediction models for the soluble solids content (SSC) and titratable acid (TA) content of wine grapes using hyperspectral imaging (HSI) combined with chemometric methods. Back propagation neural network (BPNN) and random forest (RF) were found to be the best models for predicting the SSC and TA content, respectively. The coefficients of determination for prediction (R_P²) of the BPNN and RF models were 0.9559 and 0.9253, with root mean square error of prediction (RMSEP) of 0.9965 °Brix and 2.0451 g/L, and residual predictive deviation (RPD) of 4.3602 and 3.0081, respectively. These results demonstrate that the combination of HSI and chemometric methods enables rapid and non-destructive detection of the SSC and TA content of wine grapes, thereby providing a solution for the rapid and non-destructive detection of physicochemical indexe of wine grapes.