Low-temperature freezing injury is one of the most serious natural disasters in tea plantations. Quantitatively detecting tea leaves under low-temperature stress is of great significance for evaluating the degree of freezing injury in tea plantations and taking timely measures. The traditional detection of tea plants under low-temperature stress is mainly through observing gardeners and determining physicochemical parameters. However, this method has some problems, such as low accuracy, low efficiency and strong subjectivity, which seriously affects the management of tea plants in the later stage of disasters. A method for quantitatively judging the freezing degree of tea plants based on hyperspectral imaging was proposed. First, the hyperspectral imaging equipment was used to collect spectral data on tea leaves in the early and later stages of the non-freezing periods. Moreover,the average reflectance of tea leaves was extracted. The physicochemical parameters such as relative electrical conductivity (REC), chlorophyll (SPAD) and malondialdehyde (MDA) in the corresponding leaves were determined. Secondly, the collected original hyperspectral data were preprocessed by using multivariate scattering correction (MSC), first derivative (1-D) and Savitzky-Golay (S-G) algorithms, and the characteristic bands of the preprocessed hyperspectral data were screened by using the uninformative variable elimination (UVE) and successive projections algorithm (SPA) algorithms. Finally, the quantitative prediction models of SPAD, REC and MDA content were established by using a convolutional neural network (CNN), support vector machine (SVM) and partial least squares (PLS).The results showed that: (1) The spectral curve preprocessed by the MSC+1-D+S-G algorithm had more prominent peaks and troughs than the original spectral curve, which improved the resolution and sensitivity of the spectrum and helped to improve the accuracy of the regression model established later. (2) The number of feature bands screened by the UVE algorithm was the largest, and the later modeling effect was good. The number of feature bands screened by the SPA algorithm was the least, and it was more suitable for building regression models with traditional machine learning methods. (3) The best prediction models of SPAD, REC and MDA were SPAD-UVE-CNN (R2P=0.730, RMSEP=3.923), REC-UVE-SVM (R2P=0.802, RMSEP=0.037) and MDA-UVE-CNN (R2P=0.812, RMSEP=0.008). In this study, the combination of hyperspectral imaging technology and a variety of algorithms can non-destructively, accurately and quantitatively monitor the degree of low temperature stress in tea leaves, which is of great significance for quickly predicting the occurrence of freezing damage in tea plantations and taking necessary measures.