Before the samples were cryopreserved, the chlorophyll content of each leaf was measured by a portable chlorophyll meter (SPAD-502, Japan). Each leaf was measured at 6 different positions and leaf veins were avoided manually. The SPAD value of tea leaves was the mean value of the 6 measured points.

Tea leaves were frozen by liquid nitrogen and ground into powder. The contents of SS and MDA and the activities of protective enzymes CAT, SOD and POD were determined according to the kit of Suzhou Grace Biotechnology Co., Ltd. SS, MDA content and protective enzyme CAT, SOD, POD activity kits are all microplate method, and the product numbers are G0501W, G0109W, G0105W, G0103W and G0107W respectively.

The establishment of LTRI. To quantitatively and comprehensively analyze the low temperature stress of tea leaves, LTRI was constructed. Firstly, the 6 major physiological indexes (i.e., SPAD, SS, MDA, CAT, POD, SOD) related to low temperature stress were standardized and analyzed by SPSS 25.0 software. Then, the linear combination coefficient (LCC) and the comprehensive score coefficient (CSC) were calculated by the formulas (1) and (2) respectively, and the LTRI was finally obtained.

In formula (1), LC_ij is the load coefficient of the j-th principal component and the i-th index, λ is the initial eigenvalue of j-th. In formula (2), VC is the variance contribution rate, and CVC is the cumulative variance contribution rate.

The acquisition and analysis process of hyperspectral data was shown in Figure 1. In this study, HSI system (GaiaField-Pro, Jiangsu Dualix Spectral Image Technology Co. Ltd) was used to collect spectral information of tea leaves in the range of 397-1008 nm. The composition, specifications and parameters of the HSI system were the same as those in reference (Mao et al., 2022). The spatial resolution of hyperspectral images collected in this study was 960 × 1101 (Space × Spectrum), the spectral resolution was 3.5 nm and the number of spectral channels was 176. In order to reduce the influence of dark current noise and external environment, the original hyperspectral image (R) was corrected to reflectance hyperspectral image (C) by equation (3) after spectral data acquisition.

Where, C is the corrected image, R is the original image, D is the black reference image obtained by covering with a lens cover with a reflectance of about 0%, and W is the white reference image collected using a pure white standard whiteboard with a reflectance of about 100%.

Due to the influence of hyperspectral equipment and environmental factors, the spectrum of tea leaves has problems such as scattering effect and noise, which will weaken the spectral signals of the internal physical and chemical indicators of tea materials and affect the accuracy of the regression model. Therefore, in order to eliminate the spectral differences caused by different scattering levels, the MSC (Cheng et al., 2014) algorithm was used to remove artifacts or defects in the data matrix; In order to effectively reduce the random noise of the average reflection spectrum, the S-G (Kong et al., 2014) algorithm was used to “average” or “fit” each point within a certain width window, so as to obtain the best estimated value of the smooth point; In order to eliminate the process of baseline shift and split overlapping spectral peaks, the spectral band characteristics were enhanced by the 1^stD (Feng and Sun, 2013). Among them, the formula of differential method-1^stD is shown in formula (4).

Where, y_i is the spectrum of the i-th sample, and Δx is the wavelength interval.

Although the preprocessed spectral data has removed some noise, it still contains too much band information. This not only increased the amount of data operations, but more importantly, the redundancy of the band variables would affect the prediction accuracy and stability of the model (Ouyang et al., 2016). Therefore, SPA, CARS and UVE algorithms were used to select representative bands in the full-band spectral data as the “feature bands”, and compare them with the full-band (NONE) data, and finally select the best feature variable selection algorithm. Among them, the SPA algorithm can find the variable combination containing the least redundant information from the spectral information, and select a set of representative spectral variables with the minimum collinearity, thus reducing the complexity of the model (Araújo et al., 2001). The CARS algorithm is a variable screening method based on the principle of “survival of the fittest” in Darwin’s evolution theory, which can effectively find the optimal spectral combination (Yuan et al., 2021). The UVE algorithm can remove irrelevant variables with more noise, optimize model variables, and improve the predictive ability of the model (Li et al., 2016b). The basic parameters of SPA, CARS and UVE algorithms used in this study were shown in the Table 2.

In order to find an optimal prediction model, a regression model between the spectral data of tea leaves and its physical and chemical indexes was established by using four machine learning methods of PLS, SVM, RF, BP and two deep learning methods of CNN and LSTM, and the effects of six models were compared.

Each of these six methods has its own advantages. Among them, the PLS algorithm is the widely used prediction model at present, it can not only overcome the collinearity problem, but also remove the influence of unhelpful noise on regression (Geladi and Kowalski, 1986). The optimal number of potential components in PLS was determined by minimizing the RMSE of leave-one-out cross-validation (Liu et al., 2021). The SVM can solve the classification and regression problems of high-dimensional features (Cortes and Vapnik, 1995). When applying a SVM to the regression problem, the model can be optimised by adjusting the polynomial kernel of a kernel function. The RF algorithm adopts random sampling, and the trained model has small variance and strong generalization ability (Breiman, 2001). The BP neural network has a high degree of self-learning and self-adaptive ability, and has the ability to apply learning results to new knowledge (Sun et al., 2019). The CNN network is one of the most effective networks for data feature extraction (Li et al., 2016a). The LSTM network can not only discover the mutual dependence of the data in the time series data, but also automatically detect the best mode for the relevant data (Hochreiter and Schmidhuber, 1997).

The network structures of PLS, SVM, RF, BP, CNN and LSTM were determined with reference to our two recent studies (Li et al., 2022a; Li et al., 2022b). Because the network structure applied in previous studies have made better progress in the monitoring agronomic traits of tea plants, and the established models have a strong generalization ability and are applicable to tea plants. Specifically, the input layer of the CNN is a two-dimensional matrix of spectral data. The hidden layer mainly consists of three convolutional layers, three activation functions, one maximum pooling layer, and one fully connected layer. In addition, the optimal combination of parameters was determined by continuously adjusting the network parameters. The final determined specific parameters were shown in Table 3.

Based on the characteristic bands screened by CARS, SPA, UVE algorithms and all the bands, a total of 168 different indexes models were established by using PLS, SVM, RF, BP, LSTM and CNN algorithms. The accuracy evaluations (R², RMSE and RPD) of all models were shown in Supplementary Table 2. Figure 5 showed the validation results of the model with R p 2 evaluation test set samples. The results indicated that in the prediction of SPAD, SS, MDA, CAT, POD, SOD and LTRI, the models with the highest prediction accuracy were SPAD-UVE-BP, SS-UVE-LSTM, MDA-UVE-SVM, CAT-UVE- CNN, POD-UVE-CNN, SOD-SPA-SVM and LTRI-UVE-CNN, the R p 2 of the models were 0.879, 0.807, 0.779, 0.760, 0.577, 0.698 and 0.890, respectively. Figure 6 showed the prediction results of the best model for 7 indicators. Among them, the red line was the 1:1 line, and the orange line was the regression line between the predicted and actual values. The predicted values of the samples have been distributed around the regression line at a relatively close distance, indicating that the 7 models have good robustness.

To sum up, the LTRI comprehensive model had higher accuracy and better efficacy than the 6 single index models. This demonstrated that the relationship between LTRI and spectrum was closer than other single physicochemical indicators. LTRI can more comprehensively and objectively evaluate the low temperature stress of tea plants, and effectively evaluate the cold resistance of tea plants.

Since LTRI was more representative in each indicator, this section only described the prediction model based on LTRI. Figure 7 showed the comparison of the results of the LTRI models established based on the different number of characters bands. The results showed that the spectral information of the bands selected by the UVE algorithm was more abundant and effective, and the regression effect of the model based on the characteristic bands screened by the UVE was better than the model based on the full-band and the models based on the feature bands screened by CARS and SPA. The study of Yang et al (Yang et al., 2021). also proved that the characteristic wavelength model established after variable screening was better than the full-band model. The reasons may be: compared with the full-band, UVE deletes some collinearity variables; compared with CARS and SPA algorithms, UVE algorithm retains more information related to LTRI in spectral data. In addition, the CARS may cause unstable wavelength selection results due to Monte Carlo sampling (Yun et al., 2014). While the SPA greatly reduces the number of variables and simplifies the model, it may also delete some key variables, thus reducing the accuracy of predictions. The research results of Ji et al (Ji et al., 2022). also showed that CARS extracted fewer feature bands than UVE-SPA, but also lost more useful bands, resulting in a worse subsequent regression effect. This was consistent with the findings of Guo et al (Guo et al., 2020). They compared the results of the UVE-SPA-PLS and CARS-SPA-PLS models with those of the UVE-PLS and CARS-PLS models, and found that the predictive performance of the SPA engagement model was slightly lower than that of UVE-PLS and CARS-PLS.

In order to compare the prediction effects of different modeling algorithms on LTRI, 18 prediction models of LTRI were analysed and compared. As shown in Figure 7, among the 6 modeling algorithms, the CNN model had the best performance (UVE-CNN, R p 2 = 0.890). Interestingly, the other two combined models of CNN (CARS-CNN, R p 2 = 0.770; SPA-CNN, R p 2 = 0.727) achieved poor results. On the one hand, because these three CNN models all use different types of input vectors, and the feature bands obtained based on UVE screening were more than that of SPA and CARS. Thus, the UVE-CNN model included more characteristic information, which indirectly improved the accuracy of CNN model prediction (Hsieh and Kiang, 2020). On the other hand, because deep learning was very suitable for a large number of data sets, there were only 192 samples in this paper, which limited the use of deep learning to a certain extent. Even so, the R p 2 of all models in this study exceeded 0.70, indicating that model performance can be improved in future studies. Perhaps, the expansion of the sample set can expand the advantages of the deep learning method. In addition, other deep learning models (Stacked Auto-Encoders, Recurrent Neural Network, and Deep Belief Networks) may be interesting explorations for the evaluation of tea plants frost damage.

In summary, the LTRI-UVE-CNN model achieves the best results ( R c 2 = 0.957, RMSEC=0.247, RMSEP=0.249, R p 2 = 0.890, RMSEP=0.325, RPD=2.904), which proves that the comprehensive evaluation model is better than the single physical and chemical index model, it is also proved that the UVE-CNN architecture can exhibit higher prediction performance.