Tea is an important cash crop in China. The early detection and diagnosis of tea diseases will help agricultural producers to take effective protective measures in time. In order to achieve accurate discrimination of tea diseases, the spectral characteristics of tea were studied using chlorophyll fluorescence spectrum. A total of 90 samples of healthy tea leaves, 90 samples of the early stage of Cephaleuros virescens Kunze leaf disease and 90 samples of the severe stage of Cephaleuros virescens Kunze leaf disease were collected in the experiment and were accordance with the Kennard-Stone algorithm divided into the training set and prediction set according to the proportion of 3:1 for each kind. Adopt the chlorophyll fluorescence spectrum collection system to collect the spectrum of tea leaf spot disease and normal leaves and set the collection parameters: integration time 20 ms and laser power 40 mW. The spectral response characteristics of the diseased and normal leaves were analyzed separately. In general, there are differences in the absorption intensity of the three types of leaves, and the spectrum trends are the same. There is a chlorophyll fluorescence peak near 685 and 740 nm. The difference is mainly reflected in the difference in fluorescence peak intensity. Then the polynomial smoothing(Savitzky-Golay)method was carried out for smoothing and noise reduction on the original spectral, the establishment of partial least squares discriminant model (PLS-DA), in the PLS-DA modeling set model, the number of misjudged samples is 3, the false positive rate is 3%; in the PLS-DA prediction set model, the number of false positive samples is 5, and the false positive rate is 7.1%. Then the support vector machine model established by 4 different kernel functions is compared. RBF is used as the kernel function. The SVM model established by PCA has the lowest misjudgment rate, and the accuracy rate reaches 95.72%. Finally, the model established by principal component analysis (PCA) and linear discriminant analysis (LDA) has the best effect, and the accuracy rate reaches 98.9%. The selection of the optimal number of principal components is obtained by the leave-one-out verification method. When the first 10 principal components are selected for modeling, the cross-validation accuracy rate is the highest, reaching 98%. Through model comparison, the accuracy of the PLS-DA modeling set and prediction set is more than 90%. Among the support vector machine models built with four kernel functions, the radial basis kernel function model is the best, reaching 95.72%, the linear discriminant model (LDA) established after principal component analysis has the best effect, and the recognition rate is 98.9%. This study uses chlorophyll fluorescence spectroscopy combined with chemometrics to identify tea diseases, providing a new method for rapid and accurate prediction of tea diseases.