Abstract: The geographical origin was one of the important factors affecting the quality of hawthorn. To discriminate the geographical origin of hawthorn rapidly and nondestructively, hawthorns from five different provincial production areas were used as samples, and visible-shortwave infrared (410~2500 nm) band hyperspectral data were obtained for the pedicel face (G), side (C), and bottom (D) of each sample by using a near-infrared hyperspectral imaging system. Partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and random forests (RF) classification models were built by multivariate scattering correction (MSC), first derivative (D1), SG smoothing (Savitzky-Golay, SG), and standard normal transform (SNV) four preprocessing methods. The results showed that the D-D1-SVM model discriminated optimally, with 100% accuracy in both the training and prediction sets. To simplify the model, successive projections algorithm (SPA) and competitive adaptive reweighted sampling algorithm (CARS) were applied to select feature wavelength. The multivariate data analysis revealed that the D-SPA-SVM model had the best performance, with an accuracy of 95.2% and 93% for the training and prediction sets, respectively. This study could provide technical support for rapid and non-destructive identification of hawthorn origin.