In this study, a method for detecting the moisture content of withered leaves for black tea was proposed based on micro-near infrared spectroscopy (NIR). An NIR spectrometer developed in our lab was used to collect diffuse reflectance spectra of the withered leaf samples, whose moisture content was measured using a moisture meter at various time points. Pretreatment methods, variable screening methods and principal component analysis (PCA) were adopted to perform optimization and dimensionality reduction of spectral data to establish a discriminant model for the withering degree and a prediction model for the moisture content of withered leaves for Yimeng black tea. The results showed that the random forest (RF) discriminant model achieved an accuracy rate of 99.4% on the test set, exhibiting extraordinary classification performance. The model developed using bootstrapping soft shrinkage combined with support vector regression (BOSS-SVR) exhibited good prediction performance with correlation coefficient of calibration (r_c) of 0.994, correlation coefficient of prediction (r_p) of 0.984, root mean square error of calibration (RMSEC) of 0.730, root mean square error of prediction (RMSEP) of 1.198, and relative percent deviation (RPD) of 4.485. This research provides a theoretical basis and data support for the standardized and digital production of Yimeng black tea.

Black tea is the second largest type of tea in China. Four stages are involved during black tea processing, such as the withering, rolling, fermentation, and drying. Among them, the fermentation can be closely related to the color and taste of black tea. Furthermore, the quality of black tea after fermentation can dominate its market value. In addition, the black tea with moderate fermentation is less prone to spoilage during storage. Mild fermentation or excessive fermentation can decrease in the preservation of tea, making them susceptible to external influencing factors on the unique quality of black tea. Therefore, it is often required for the precise control over the fermentation quality of black tea. At present, the fermentation quality of black tea can rely mainly on the experience of tea makers in "observing color, smelling aroma, and touching texture", which is highly arbitrary and subjective task. The quality of black tea has been limited to the standardization in production. In this study, an improved deep learning model was proposed to accurately evaluate the fermentation quality of black tea using image features and machine vision technology. The deployment of the model was considered under actual production environments. Firstly, the experimental comparisons were conducted on the seven models of convolutional neural network under the same conditions. Among them, the Ghostnet model shared the best discriminative performance, in order to select as the teacher model. Mobilenetv3_small was used as the student model after comparison, considering both the discriminative performance and complexity of the model. Secondly, a series of experiments were conducted to compare the discriminative performance, taking the AdamW, SGD, and RMSProp as the research objects. After that, the student and teacher model were replaced with RMSProp optimizer. The discriminative performance of the model was balanced between its complexity or speed. Afterwards, the loss functions of the student model Mobilenetv3_small and the teacher model Ghostnet were simultaneously changed into the CE Loss. Their discriminative performance was further improved for the less complexity of the models. There was the range of knowledge distillation loss ratio between 0.1 and 2.0. Knowledge distillation experiments were performed on the student model under SoftTarget using the teacher model. The results showed that the discrimination performance was significantly improved to maintain the complexity and speed, when the knowledge distillation loss ratio was 0.4, 1.7, and 1.9. In contrast, the discriminative performance of the model was improved the most, when knowledge distillation loss ratio was 1.9. The improved model was achieved in the Accuracy, Precision, Recall, and F1 of 96.93%, 95.15%, 95.79%, and 95.46%, respectively. These scores increased by 2.01, 2.67, 3.72 and 3.19 percentage points, respectively, compared with the original model. The fermentation quality of black tea can be precisely controlled to fully meet the fermentation quality of the black tea. The finding can also provide the strong technical support for the digital and intelligent processing of black tea.

In order to rapidly detect the moisture content in tea leaves during green tea fixation, a quantitative prediction model for moisture content changes during green tea fixation was constructed by using machine vision and near infrared spectroscopy. Spectral and image information of samples at different stage of fixation was collected. The characteristic wavelengths were extracted by four different variable selection methods, competitive adaptive reweighted sampling (CARS), variable combination population analysis (VCPA), variable combination population analysis, iterative retained information variable algorithm (VCPA-IRIV), and random frog (RF) algorithm, and the prediction models were developed by using linear partial least squares regression (PLSR) or non-linear support vector regression (SVR) as well as fusing 15 color and texture features in the image. The results showed that the model based on data fusion had improved prediction accuracy compared with that based on single data. The SVR model developed based on the spectral feature wavelengths extracted by CARS algorithm and fusion of 15 color feature using normalization pretreatment combined with principal component analysis (PCA) was the best among all established models. The correlation coefficient of the calibration set (R_c) for the model was 0.9742, prediction set correlation coefficient (R_p) value 0.9719, and relative percent deviation (RPD) value 4.1546, indicating the model had excellent prediction performance. In conclusion, this study proved the feasibility of integrating spectroscopy and imaging technology to predict the moisture content during the process of green tea fixation, which overcomes the problem of the low prediction accuracy of a single sensor, and lays a theoretical foundation for rapid nondestructive detection of the moisture content of fixed tea leaves for green tea and accurately controlling tea fixation quality.

The tea yield estimation provides information support for the harvest time and amount and serves as a decision-making basis for farmer management and picking. However, the manual counting of tea buds is troublesome and inefficient. To improve the efficiency of tea yield estimation, this study presents a deep-learning-based approach for efficiently estimating tea yield by counting tea buds in the field using an enhanced YOLOv5 model with the Squeeze and Excitation Network. This method combines the Hungarian matching and Kalman filtering algorithms to achieve accurate and reliable tea bud counting. The effectiveness of the proposed model was demonstrated by its mean average precision of 91.88% on the test dataset, indicating that it is highly accurate at detecting tea buds. The model application to the tea bud counting trials reveals that the counting results from test videos are highly correlated with the manual counting results (R² = 0.98), indicating that the counting method has high accuracy and effectiveness. In conclusion, the proposed method can realize tea bud detection and counting in natural light and provides data and technical support for rapid tea bud acquisition.