Abstract: Affected by the rainy weather during the harvest period, wheat is prone to sprouts, and severely, a large-scale reduction in production might occur. Failure to dry in time or poor management during storage can directly cause normal wheat to sprout and deteriorate wheat quality. During purchase, too much malt will cause economic losses. At this stage, the judgment of wheat germination is generally by naked eyes, but it is still hard to detect early stage of germination especially when the changes are subtle in appearance. If the early detection was conducted by a chemical method, it is usually time consuming and the steps will be cumbersome, and it will be destructive to the sample. Therefore, it is necessary to screen a fast and nondestructive method to detect the germination degree of wheat, and to monitor and early-warn the germination of wheat. Hyperspectral imaging technology can not only reflect the external characteristics of the sample, but also reflect the internal structure and chemical composition. It is a fast and non-destructive detection technology. Such early judgement can properly serve the storage, processing and utilization of wheat. This study uses hyperspectral spectral information to classify and identify wheat with different germination degrees. First, collect the hyperspectral data of sprouted wheat (bubble, cracked skin, and white) and sprouted wheat (the bud length is half the length of the wheat grain, the bud length is the same as wheat kernel length). After processing, the average spectrum of each grain of wheat is extracted, and then a least squares-support vector machine (LS-SVM) classification model was established by four methods, including Savitzky-Golay convolution smoothing method, Savitzky-Golay convolution derivation method (1ST, 2ND), standard normal variable transformation (SNV). Among them, the recognition accuracy of SNV was the highest (91.70%). Secondly, particle swarm optimization (PSO) was used to optimize the parameters of the LS-SVM model and the kernel function parameters, the accuracy of the optimized model was 93.14% with an increase of 1.57%. Finally, in order to further reduce the amount of calculation and improve the accuracy, Competitive Adaptive Re-weighting Algorithm (CARS) was adopted to select 49 characteristic wavelengths, and the accuracy of this classification model was 94.13%. After determining the optimal model as PSO-CARS-SNV-LS-SVM system, it could quickly and nondestructive detect the wheat at different germination stages for classification, and finally achieve the purpose of visualization of classification results.