A three-vigor gradient classification model for maize seeds was constructed using machine learning and deep learning algorithms along with hyperspectral imaging technology to solve the problems of time-consuming and seed damage in the traditional method for detecting seed vigor and to realize the rapid， non-destructive detection of maize seed vigor. 1 012 maize seeds were divided into three vigor gradient samples by artificial aging. The hyperspectral noise was removed with convolution smoothing （SG） and multivariate scattering correction （MSC） after collecting the hyperspectral data of maize seeds. Principal component analysis （PCA） and continuous projection algorithm （CPA） were used for dimensionality reduction of spectral feature， respectively. Three bands including 1 156 nm， 1 191 nm， and 1 463 nm were extracted from the reduced dimensionality band to synthesize a false color image. The texture features of region of interest （ROI） were extracted using local binary mode （LBP） and fused with pure spectral features. Machine learning models including decision tree （DT） and support vector machine （SVM） models constructed based on pure spectral features and the random forest （RF）， SVM and extreme gradient lifting tree （XGBoost） models constructed based on fused features were established. Maize seed vigor was predicted by inputting the false color images into five deep learning models including ResNet18， MobileNetV2， DenseNet121， Efficientb0， and Efficientb2. The results showed that the PCA-SVM model performed best for pure spectral features， with a test set accuracy of 92.5% in terms of machine learning methods. The SVM model performed best for fusion features， with a test set accuracy of 93.1%. In terms of deep learning methods， the lightweight MobileNet achieved the highest test set accuracy of 99.5%. The classification activation mapping method based on interpretable gradient indicated that the classification network focused on the bottom or basal region of maize seeds. It will provide some references for the nondestructive detection of maize seed vigor in terms of data sources， deep neural network visual interpretation and machine learning， and deep learning performance analysis.