Saffron (Crocus sativus L.), the dried stigma, is an extremely valuable spice and medicinal herb, whose economic value is affected by geographical origin, age and grade. In this study, we proposed a method to identify saffron from different Chinese origins, ages and grades, which was based on visible-near infrared hyperspectral imaging (VNIR-HSI), machine learning and data fusion strategies. Firstly, saffron samples were graded according to lSO2011/2010 standards, with age having a greater influence on grade than geographical origin. By comparing the effectiveness of different classification algorithms with different preprocessing methods, the results showed that MSC-CARS-SVM was an effective spectral classification algorithm to determine saffron origin and FD-CARS-SVM was an effective spectral classification algorithm to determine saffron age and grade. Finally, image and spectral features were fused at a mid-level to establish classification models for origin, age and grade, and the results showed that origin and age models were more effective after fusion than the initial spectral information, with prediction accuracies of 98.3% and 97.9%. However, the spectral FD-CARS-SVM model was found to be the most discriminative with a prediction accuracy of 89.6% for grade identification. This study provides a theoretical basis and technical support to characterize saffron quality for industry and consumers.