Applications of deep learning in tea quality monitoring: a review

Abstract

Tea is a popular beverage which can offer numerous benefits to human health and support the local economy. There is an increasing demand for accurate and rapid tea quality evaluation methods to ensure that the quality and safety of tea products meet the customers’ expectations. Advanced sensing technologies in combination with deep learning (DL) offer significant opportunities to enhance the efficiency and accuracy for tea quality evaluation. This review aims to summarize the application of DL technologies for tea quality assessment in three stages: cultivation, tea processing, and product evaluation. Various state-of-the-art sensing technologies (e.g., computer vision, spectroscopy, electronic nose and tongue) have been used to collect key data (images, spectral signals, aroma profiles) from tea samples. By utilizing DL models, researchers are able to analyze a wide range of tea quality attributes, including tea variety, geographical origin, quality grade, fermentation stage, adulteration level, and chemical composition. The findings from this review indicate that DL, with its end-to-end analytical capability and strong generalization performance, can serve as a powerful tool to support various sensing technologies for accurate tea quality detection. However, several challenges remain, such as limited sample availability for data training, difficulties for fusing data from multiple sources, and lack of interpretability of DL models. To this end, this review proposes potential solutions and future studies to address these issues, providing practical considerations for tea industry to effectively uptake new technologies and to support the development of the tea industry.

Introduction

Tea is one of the most widely consumed non-alcoholic beverages in the world. It is highly favored for its numerous health benefits. According Food and Agriculture Organization (FAO), the global production of tea is approximately 6 million tons per year, highlighting the importance of the tea industry in supporting the local economy (Karwowska et al. 2019). It is reported that daily consumption of green tea can help prevent and treat various chronic diseases, including cardiovascular diseases and certain types of cancer (Khatoon 2023). Additionally, the bioactive components in green tea exhibit synergistic effects on metabolic systems (Sidhu et al. 2024). Clinical studies have further demonstrated that tea consumption can enhance our endogenous antioxidant defense system (Liu et al. 2024). The quality of tea plays a key role on its offerings for health benefits and market value. Tea quality depends not only on various sensory attributes such as flavor, color, and aroma, but also on nutritional and functional components such as tea polyphenols, theaflavins, and theanine (Li et al. 2013). However, these quality attributes are influenced by various factors such as regions, climate, cultivation, and processing techniques. It is of great importance to develop automated, standardized, and precise tea quality assessment methods to ensure the mutual benefits for tea producers and consumers.

The tea processing procedure generally involves withering, fixation, rolling, fermentation, post-fermentation, and drying (or roasting). Tea varieties can be classified into six categories based on different processing methods: green tea (unfermented), yellow tea, white tea, oolong tea, black tea (fermented to various degrees), and dark tea (post-fermentation) (Kaushal et al. 2022). Different processing techniques can significantly impact the aroma, taste, and color of the final tea product (Turgut et al. 2021), making it challenging to evaluate the overall tea quality.

The overall tea quality assessment involves sensory evaluation and chemical composition analysis, flavor analysis, and others. Traditionally, the sensory evaluation is conducted by professional tea tasters who assess key quality attributes (including color, aroma, taste, etc.) to score the tea. This method is subjective and lacks a standardized assessment process (Moreira et al. 2024). Chemical composition analysis can be carried out to identify components associated with tea flavor and nutritional value, such as tea polyphenols, amino acids, and caffeine. However, the concentration of those compounds varies due to factors such as plucking time and processing methods (Xu et al. 2021). High-performance liquid chromatography (HPLC) and gas chromatography-mass spectrometry (GC-MS) are two widely used analytical techniques for comprehensive chemical profiling of tea constituents, enabling precise quality assessment. While these methods offer high-precision analytical results, they often involve complex sample preparation procedures and the use of large amount of chemicals (Zhi et al. 2017). Unlike traditional chemical analysis methods, sensor-based approaches can provide rapid, real-time and non-destructive analysis. For example, the electronic nose simulates human olfaction by detecting volatile compounds of tea products through a gas sensor array, enabling the rapid evaluation of tea aroma quality. The electronic tongue uses a liquid sensor array to capture the fingerprint response of tea samples by simulating human taste perception to assess tea flavor. Spectroscopic-based sensors, on the other hand, study the interaction between the light and the tea samples by analyzing the absorbance characteristics across different spectral ranges. The spectra obtained normally contain spectral signals at hundreds of wavelengths, allowing for both qualitative and quantitative analysis of the corresponding components (Calvini And Pigani 2022; Tang et al. 2023; Tibaduiza et al. 2024).

The above-mentioned sensor-based methods can generate a large amount of data which consists of useful information that can be linked to tea qualities. Effective data processing methods are required to extract meaningful data features from such a complex data structure. Traditional data processing methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), typically require complex preprocessing steps and are not able to achieve real-time comprehensive analysis. Moreover, these conventional approaches often need to manually construct the data features, which may compromise the accuracy and generalizability of the results. In addition, as a result of variations in the detection environment as well as the heterogeneity of different tea samples, the acquired sensor data can contain both feature information and certain levels of noise, which increases the difficulty of feature extraction by traditional data-processing methods.

DL is a subset of machine learning (ML) methods that processes data by simulating the human brain to learn patterns and features within the data. More specifically, DL can automatically learn effective features from a large amount of raw data without manual feature extraction. Compared to traditional methods, DL offers better model generalization and demonstrates strong model robustness to heterogeneous data. When dealing with data variations caused by objective factors such as sensor noise, instrument variability and environmental conditions, deep neural networks can effectively capture local and global features from the samples through multi-layer feature extraction.

In recent years, the continuous advancement of artificial intelligence has driven the wide application of DL methods to solve complex problems across various fields. For instance, Mamoshina et al. (2016) employed deep neural networks (DNNs) to automatically extract features from high-dimensional omics data, facilitating the discovery of disease biomarkers and drug development. Christin et al. (2019) applied convolutional neural networks (CNNs) to classify and identify wildlife images and audio, significantly advancing ecological monitoring and conservation. Mahdavifar and Ghorbani (2019) utilized CNNs and recurrent neural networks (RNNs) for effective intrusion detection and malware classification. Alzubi et al. (2022) proposed the FDL-CADIS framework, which integrates MobileNetV2 with either a gated recurrent unit (GRU) or a long short-term memory (LSTM) network to visualize malicious code files, enabling efficient identification and classification of sophisticated malware. Additionally, Alzubi et al. (2024) developed a hybrid CNN-LSTM deep learning intrusion detection system (IDS) for edge devices, achieving high-accuracy network threat recognition. In the field of materials science, DL techniques such as CNNs and graph neural networks (GNNs) have been employed to analyze crystal structures and microscopy images, enabling accurate material property prediction and image-based structural recognition (Choudhary et al. 2022). In agriculture, DL has also demonstrated promising results. Ahmad et al. (2023) leveraged CNNs and transfer learning to analyze leaf images for the accurate identification of multiple plant diseases and their severity levels. Furthermore, Mukhiddinov et al. (2022) developed a real-time freshness classification system for fruits and vegetables, which is based on an improved YOLOv4 model with the Mish activation function and data augmentation techniques.

DL has also achieved promising research outcomes in the tea industry, where DL methods are applied to sensor data for accurate prediction of the intrinsic quality traits of tea (Ren et al. 2024). In addition, DL combined with computer vision has been widely applied to evaluate the external quality of tea. CNNs are used to process tea image data, and to assess tea sensory attributes (Bhargava et al. 2022). As a powerful tool for tea quality detection, DL has shown great potential in both qualification and quantification tasks for tea quality evaluation (Gharibzahedi et al. 2022).

Prior to this study, several reviews related to tea quality detection have been published. For example, Yu et al. (2020) discussed emerging technologies such as spectroscopy and electrochemical methods in tea product quality evaluation and safety testing. Kaushal et al. (2022) reviewed the application of electronic noses and traditional intelligent pattern recognition methods in tea quality detection. Moreira et al. (2024) summarized tea quality evaluation methods including analytical techniques and human sensory assessment methods. However, most existing review publications do not cover the application of DL algorithms. This review has three main contributions. Firstly, it introduces the application of DL methods coupled with sensing technologies for tea quality evaluation (including internal and external quality attributes) as well as tea authentication and tea adulteration detection. Secondly, the scope of the application in the tea industry covers 3 key areas: tea cultivation, tea processing, and product evaluation, thus providing readers with a comprehensive overview of the application-based DL solutions for tea quality detection. Finally, this review discusses several key challenges in the practical applications of DL algorithms and proposes future research directions to enhance the development of DL-based methodology for tea industry.

Apart from the Introduction section, this review is structured into six remaining sections. Section 2 describes the methodology used for literature selection, including data sources, search strategies, and inclusion criteria. Section 3 introduces traditional ML methods in tea quality detection and its limitations. Section 4 introduced commonly-used DL algorithms that have been applied to tea quality evaluation. Section 5 elaborates on the practical applications of these DL algorithms in various stages of the tea industry. Section 6 discusses current challenges in applying DL approaches in the tea industry and proposes future studies to enhance DL models applicability and explainability. Finally, Sect. 7 concludes the limitations of current research on the practical implications of DL in tea and future research directions.

Methods for literature search

To collect relevant studies for this review, a systematic literature search was conducted in accordance with the general review protocols. The databases used for searching included Web of Science (WoS), ScienceDirect, Scopus, IEEE Xplore, MDPI, ACM Digital Library, and arXiv, supplemented by the academic search engine Google Scholar to ensure a comprehensive coverage of the topic. Among these, WoS covers over 15,000 journals and more than 90 million documents, ScienceDirect covers nearly 4000 journals and 43,000 books, and Scopus indexes covers over 20,000 active journals. Given that the focus of this review is on the recent application of DL in tea quality detection, and based on prior bibliometric statistics and systematic reviews on tea quality detection, a combination of keywords “deep learning” and “tea quality” was used to capture relevant literature. While the keyword includes the term “deep learning”, our research still focuses on tea quality detection, and the additional terms related to “deep learning” were used to filter relevant literature based on this technological context. The search string in Scopus was as follows: TITLE-ABS-KEY (“tea quality” OR “tea grading” OR “tea evaluation”) AND TITLE-ABS-KEY (“deep learning” OR “neural network” OR “artificial intelligence”). The Boolean operator “OR” was used to combine these keywords, and “AND” was used to ensure that both search conditions were met. To further optimize the literature search, only English-language articles were considered, and as with other related literature reviews, only articles published in peer-reviewed academic journals were selected to ensure that the chosen literature contained the most influential and validated knowledge. Additionally, the search was limited to articles published in the past ten years, from 2015 to 2025. Among these databases, the Web of Science yielded the highest number of relevant articles, with publication years mainly between 2019 and 2025. Articles related to tea picking and pest detection, as well as review articles, were excluded from this review study. In cases of uncertainty, full-text review was conducted. Ultimately, 50 research papers were included in this review. The distribution of the publications by year is shown in Fig. 1, which illustrates that the number of related articles has gradually increased over the past few years.

[See PDF for image]

Fig. 1

Number of literatures collected from 2019 to 2025

Traditional machine learning methods in tea quality detection

Traditional ML methods can be used to process data acquired from various sensor-based devices (e.g., electronic nose, electronic tongue, spectrometers) at different stages during the entire tea production. ML algorithms such as unsupervised k-nearest neighbor (KNN) and support vector machine (SVM) can automatically analyze the data to reveal patterns for tea quality assessment. Traditional ML methods integrate principles from statistics and linear algebra, offering a high degree of interpretability. For instance, partial least squares regression (PLSR) constructs a linear regression model by projecting both predictor and response variables into a new latent space. Random forest (RF), on the other hand, is an ensemble classifier consisting of multiple decision trees, in which the final output is determined by the majority vote of individual trees. These ML models have relatively simple architectures and can perform well on small-sample datasets (Lin et al. 2023). Several studies have demonstrated the robustness of traditional ML models for classification tasks in tea quality evaluation. Chen et al. (2020) combined near-infrared spectroscopy (NIRS) and the RF algorithm to develop a low-cost solution for tea grade evaluation. Liu et al. (2019) proposed a multi-task model based on backpropagation neural network (BPNN) to effectively grade organic green tea. Based on image features obtained through computer vision, this BPNN method employs knowledge-driven manual feature selection, resulting in good classification performance. However, the entire tea classification process involves multiple steps (including image preprocessing, manual feature extraction and selection for modeling), which cannot meet the requirements of an end-to-end learning process for real-time detection (Wang et al. 2023). In addition, during tea production and processing, the physical and chemical characteristics of tea undergo significant changes. As a result, it is challenging for traditional ML methods to achieve accurate results in tea quality detection. More specifically, the high-dimensional data obtained from sensors such as NIRS, electronic noses, and electronic tongues often exhibit a nonlinear relationship with the target prediction attributes. In addition, dimensionality reduction techniques (i.e., PCA) or feature selection methods (i.e., competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) are often required to deal with high-dimensional data, which may potentially lead to the loss of crucial sample information. Moreover, differences in the physical and chemical properties of different teas can impact the sensor signals acquired. Noise can sometimes be included in the sample data due to the sensor device itself and the changing detection environment. Traditional ML models based on conventional data preprocessing and variable selection algorithms often struggle to distinguish between noise and relevant features, which can negatively impact the accuracy of the prediction results (Ye et al. 2021; Zhang et al. 2021).

On the other hand, the overall quality assessment of tea cannot be achieved by focusing on one single quality attribute. The internal (e.g., moisture content and catechin levels) and external (e.g., color and surface texture) quality indexes of tea are affected by multiple factors, such as geographical origin, variety, planting conditions, processing techniques and storage time. Therefore, accurately determining the overall tea quality requires the integration of both external features and internal chemical composition characteristics. By collecting data using multiple detection technologies (e.g., computer vision and NIRS), strategies such as data fusion can be applied to provide more comprehensive sample information (Li et al. 2023b).

DL-based approaches can be used to overcome the above challenges due to its powerful feature extraction capabilities, nonlinear mapping abilities, and end-to-end learning model. As an advanced data analytical tool, deep neural networks offer new opportunities to achieve a more precise and efficient result for tea quality detection. For example, attention mechanisms can enhance the focus on key features, generative adversarial network can improve the dataset by generating high-quality sample data, the advancement of DL models can help optimize the processing performance of large-scale data. All of these contribute to enhancing model generalization capabilities, enabling DL models to address the issues which can occur during tea quality detection, for example, the diverse tea varieties, multiple quality characteristics, and different production conditions. In recent years, DL-based methods for tea quality detection have gained widespread attention and demonstrated significant advantages compared to traditional ML methods. Figure 2 shows the structure of four commonly used ML methods (in Fig. 2a) and the deep neural learning network (in Fig. 2b).

[See PDF for image]

Fig. 2

Graphical explanation of machine learning and deep learning methods. a Principle of four commonly used traditional machine learning algorithms; b the process flow of deep learning algorithms based on multiple layer neural network

Deep learning algorithms

With the development of computer science, intelligent hardware and software, more advanced ML models consisting of more complex architectures and computing units are developed. These advanced ML methods fall under the category of DL. DL models can automatically learn complex data patterns through multilayer nonlinear feature extraction. In combination with loss function design, backpropagation, and gradient-based optimization. DL has become a powerful tool for tea quality assessment (Guo et al. 2024b). For tea quality detection, the most widely used DL frameworks include four aspects: data augmentation, attention mechanisms, model classification, and regression. In terms of data augmentation, a commonly used method is generative adversarial network (GAN). For attention mechanisms, channel attention mechanism (CAM) and spatial attention mechanism (SAM) are widely applied. In model classification, convolutional neural network (CNN), DNN, and LSTM are extensively used for tea quality classification. Among them, CNN-based models include a series of architectures such as AlexNet, VGGNet, residual neural network (ResNet), MobileNet, EfficientNet, Inception, and RepSet. In the regression domain, deep belief network (DBN), CNN, and LSTM are the primary methods for quantifying the concentration of key tea quality attributes. Figure 3 shows six different DL models and their architectures. Each DL model is introduced in detail in the following sections.

Data augmentation

GAN is a type of unsupervised learning method that involves a learning process where two neural networks compete against each other. It is commonly used in image generation, data augmentation, and other applications such as super-resolution imaging and video generation. The network consists of two components: a generator and a discriminator. The generator takes the input from the latent space to produce synthetic data, attempting to deceive the discriminator, while the discriminator is trained to continuously improve its ability to differentiate between the real and generated sample data. Through adversarial training, both networks are optimized, allowing the generator to produce data that closely resembles the real data distribution (Goodfellow et al. 2020). In tea quality detection, the data typically consists of images of tea samples, outputs from sensor-based devices, and other small datasets. Using GANs to generate new high-quality data to augment the training dataset is an effective way to address the limitation of small datasets. Deep convolutional generative adversarial network (DCGAN) and Wasserstein generative adversarial network (WGAN), as two common variants of GAN, have been widely applied in data augmentation tasks for tea quality detection (Zhu et al. 2025b).

Attention mechanism

Channel attention mechanism

CAM is a DL technique primarily applied in image classification to improve the performance of CNNs by focusing on the most important channel features of the input data. CAM consists of a global pooling layer, a multilayer perceptron (MLP), and a sigmoid activation function. The input feature map is firstly processed through global max pooling and global average pooling along the spatial dimensions, thus obtaining two aggregated feature maps. These feature maps are then processed by a shared MLP with two fully connected layers. The outputs of the MLP are element-wise summed to obtain a fused feature representation. Finally, the sigmoid activation function is applied to generate attention weights for each channel. These attention weights are applied to the original feature map channels, thereby enhancing important features while suppressing irrelevant ones (Guo et al. 2022). In addition, the CAM can dynamically adjust the focus of the network on different channel features, enhancing the ability of the model to identify tea quality. In recent years, the combination of CAMs and CNNs has become an effective method for tea quality detection (Kang et al. 2023).

Spatial attention mechanism

The SAM is designed to enhance critical spatial information in feature maps by assigning weights to different spatial positions, hence enabling the models to focus on key regions. Unlike the CAM method, SAM applies global max pooling and global average pooling through the channel dimension of the input feature map. The resulting two feature maps are then concatenated along the channel axis and undergo a convolution operation for feature fusion. Finally, the fused feature map is passed through a sigmoid activation function to generate a spatial attention map, which is used to reweight the spatial information of the original feature map (Mao et al. 2022). SAM is commonly integrated with CAM to enhance the performance of tea quality detection models, particularly in quantitative analyses of spectral data. This combined attention mechanism, known as the convolutional block attention module (CBAM), can effectively emphasize channel features that are relevant to target components while adjusting the significance of different spectral band regions (Zhang et al. 2024a).

Regression model

Deep belief network

DBN is a DL model composed of multiple stacked restricted Boltzmann machines (RBMs). An RBM is a generative model consisting of a visible layer and a hidden layer, where inter-layer connections are bidirectional and there are no intra-layer connections. DBN employs an unsupervised, layer-wise training strategy, where each RBM is trained using the contrastive divergence (CD) algorithm. The output of each layer serves as the input of the next layer, enabling the model to learn hierarchical feature representations. After pretraining, a supervised fine-tuning process is applied, where network parameters are optimized using the backpropagation algorithm to enhance performance in classification and regression tasks (Chen et al. 2015). DBN can automatically extract deep-level features from data while reducing dependence on labeled samples, making it a powerful tool for regression modeling.

Convolutional neural network

CNN is a specialized type of artificial neural network and a feedforward neural network primarily used for image processing and recognition. The structure of a CNN consists of convolutional layers, pooling layers, and fully connected layers. The input data, such as image features or spectral features, undergoes convolution operations through multiple filters to extract local patterns. Subsequently, pooling layers (commonly max pooling or average pooling) are applied to perform down-sampling to reduce the dimensionality of the data. For regression tasks, CNN flattens the extracted features and feeds them into the fully connected layer, which outputs a continuous value as the final regression result. The mean squared error loss function is typically used to measure the difference between the predicted and actual values. Unlike traditional handcrafted feature extraction methods, CNN can automatically learn features from raw data and progressively capture multi-level information, ranging from low-level features to high-level representations, through hierarchical convolutional learning (Alzubaidi et al. 2021). Recent studies have demonstrated that the combination of CNN and sensing technology has great potential in quantitative analysis of tea quality attributes (Luo et al. 2022b).

Long short-term memory network

The LSTM is a type of RNN designed to deal with sequential or time-series data. The network consists of multiple interconnected units, each comprising a cell state, an input gate, an output gate, and a forget gate. The forget gate determines which information from the previous cell state is being retained or discarded, while the input gate regulates the flow of current input data into the cell. The cell state is updated based on the current input data and the previous hidden state through the forget gate and input gate. The output gate then generates the current hidden state based on the updated cell state. Both the current cell state and hidden state are passed to the next time point for continuous processing. The hidden state of the final unit serves as the feature representation of the entire input sequence, which is subsequently mapped to the final regression output through a fully connected layer. LSTM introduces a gating mechanism and a long-term memory mechanism, enabling it to capture long-range dependencies in sequential data more effectively than traditional ML models (Dong et al. 2024). Some researchers have applied LSTM to perform data analysis for tea composition quantification (Hu et al. 2025).

Classification model

Deep neural network

DNN is an extension of traditional neural networks and consists of multiple layers of interconnected nodes. The network structure of a DNN model is composed of an input layer, hidden layers, and an output layer. The number of hidden layers determines the depth of the network, with each hidden layer consisting of multiple neurons which are connected to neurons from adjacent layers through weighted links. Raw data enter the network through the input layer and are subsequently processed through multiple hidden layers, where weighted summation and nonlinear transformations via activation functions take place. The final output is generated through the output layer. The multiple hidden layers in a DNN enable the model to learn and extract complex nonlinear relationships in the data, resulting in more robust and accurate model performance (Liang et al. 2020).

Long short-term memory network

LSTM can effectively solve the issues of vanishing and exploding gradients that are commonly encountered in traditional RNNs by introducing memory units. Within the memory unit, the cell state stores long-term information, while the gating units (i.e., forget gate, input gate, and output gate) control the retention of information. This mechanism allows an effective extraction of long-range dependencies in sequential data. Spectral data exhibits the characteristic of continuity and sequentiality which is similar to time series, as the wavelengths are in a continuously increasing order with the spectral range. LSTM can model the dependencies between different wavelengths in the spectrum. The hidden state at the final time step serves as the feature representation of the spectral data, which is then mapped to the class space through a fully connected layer. The final classification result is obtained via a softmax activation function by converting the output vector into a probability distribution over predefined classes (Huang et al. , 2024b).

Convolutional neural network

CNN is widely used in classification tasks for tea quality assessment. Through multiple layers of convolution and pooling operations, CNN can efficiently extract complex features from the input data and pass the extracted features into the fully connected layer, where the softmax activation function outputs the probability distribution for each class. The final classification result is obtained by comparing these probabilities. Cross-entropy is typically used as the loss function for classification tasks. Since the data obtained from different sensor devices can be different in terms of data type and data dimensions, one-dimensional convolutional neural networks (1D-CNN), two-dimensional convolutional neural networks (2D-CNN), and three-dimensional convolutional neural networks (3D-CNN) are employed to meet the practical requirements for tea quality detection (Liu et al. 2021). CNN-based models include a number of architectures such as AlexNet, VGGNet, ResNet, MobileNet, EfficientNet, Inception, and RepSet. These architectures will be introduced in detail as follows.

AlexNet is a deep convolutional neural network comprising five convolutional layers and three fully connected layers. It introduces the rectified linear unit (ReLU) activation function to address the vanishing gradient problem and applies overlapping max-pooling by using a stride smaller than the kernel size to reduce blurring and enhance feature representation. Dropout is employed in the fully connected layers to improve generalization by randomly deactivating neurons during training (Lu et al. 2021). VGGNet, extending AlexNet, increases network depth by stacking multiple 3 × 3 convolutional kernels instead of larger ones. This design facilitates the extraction of more fine-grained and abstract features, especially in complex image analysis tasks (Sun et al. 2024).

ResNet introduces a residual learning framework to mitigate the vanishing gradient problem in deep CNNs. Instead of directly mapping inputs to outputs, the core idea of ResNet is to learn the residual functions between them. The network is constructed from multiple residual blocks, each employing skip connections that add the input x to the output of a convolutional branch F(x), resulting in the residual formulation:

1

when F(x) = 0, the block performs an identity mapping such that H(x) = x (Shafiq And Gu 2022).

The Inception network consists of stacked Inception modules, each module extracts multi-scale features via parallel convolutions with different kernel sizes and pooling, followed by channel-wise concatenation (Yu et al. 2025). Classical Inception architectures range from V1 to V4, with Inception V3 being commonly used in the studies reviewed. Inception V3 reduces spatial dimensions to optimize computation while increasing feature channels to mitigate representational bottlenecks. It also applies both symmetric and asymmetric convolutional factorization to improve efficiency and feature extraction.

MobileNet is a lightweight neural network. It replaces standard convolutions with depthwise separable convolutions which are composed of depthwise (DW) and pointwise (PW) convolutions. DW performs convolution on each input channel independently, while PW conducts pointwise weighting and combination across channels. This structure significantly reduces the number of tunable parameters with minimal impact on model accuracy. The classic MobileNet models include versions V1, V2, and V3, where MobileNetV2 introduces inverted residuals and linear bottlenecks, and MobileNetV3 incorporates squeeze-and-excitation (SE) modules and optimizes the activation functions (Zhao et al. 2022). EfficientNet adopts a compound scaling strategy to jointly balance the network depth, width, and resolution. The architecture is optimized through neural architecture search (NAS) and is constructed by stacking multiple mobile inverted bottleneck convolution (MBConv) blocks. MBConv, derived from MobileNetV2, includes SE modules and the swish activation function. It expands channels with PW convolution, extracts spatial features with DW convolution, adjusts channel importance via SE, and restores dimensions with a final PW convolution, enabling residual connections with the input and completing feature extraction (Wen And He 2024).

Unlike traditional CNNs, RepSet is a neural network architecture designed for unordered and variable-sized input sets. It consists of permutation-invariant layers and standard fully connected layers. Each permutation-invariant layer maps the input features to a hidden set space, and performs pair-wise feature comparison to generate a new matrix. Subsequently, it applies a bipartite matching (BM) algorithm to compute the maximum matching between the input set and the generated matrix. This result is used as a feature vector input to the fully connected layers for classification (Chang et al. 2022).

[See PDF for image]

Fig. 3

Representative deep learning models and their architectures. a Generative adversarial network (GAN); b Channel attention mechanism (CAM); c Spatial attention mechanism (SAM); d Convolutional neural network (CNN); e Long short-term memory (LSTM) network; f Residual neural network (ResNet)

Deep learning methods in tea quality detection

Tea quality determines both its health benefits and market value of tea products. Sensor-based technologies for rapid and accurate evaluation of tea quality have become a research priority to meet the growing demand for high-quality tea products. However, traditional algorithms alone are insufficient to deal with the vast amounts of data generated, given their complexity and high dimensionality. DL is capable of learning effective knowledge automatically from high-dimensional raw data via hierarchical feature extraction, thus eliminating the need for manual feature extraction. Recent advancements in DL have widened its applications to agriculture and food domain, including tea quality detection (Yang et al. 2023). The applications of DL in tea quality assessment are mainly in three tea production stages: tea cultivation, tea processing, and tea product evaluation, as shown in Fig. 4. In the cultivation stage, research focuses on tea leaf by monitoring the chlorophyll and carotenoid content, assessing its nitrogen fertilizer status, and detecting pesticide residues. During the processing stage, DL is applied to analyze the data gathered from different processes such as withering, rolling, fermentation, and drying. In the final product evaluation stage, DL is applied to tasks such as tea variety classification (distinguishing different tea cultivars), quality grade classification (preventing mixing of different grades), key component content prediction (e.g., polyphenols, sugars, and aroma compounds), geographical origin traceability (avoiding origin mislabeling), and storage time identification (determining actual aging time).

[See PDF for image]

Fig. 4

Summary of technologies based on deep learning models for tea quality detection

Tea cultivation

High-quality fresh tea leaves are a core necessity for producing premium tea products. The cultivation and breeding of high-quality tea are influenced by various factors, among which the planting region significantly determines the cultivar-specific characteristics. Planting conditions such as soil types, temperature, humidity, and sunlight intensity are the primary factors influencing the tea growth and quality (Hajiboland 2017). To enhance tea quality, weak-light stress is commonly applied to increase the chlorophyll-a content in tea leaves; however, shading treatments may occasionally result in premature death of tea plants. Sonobe et al. (2020) investigated the potential of hyperspectral remote sensing technology to estimate chlorophyll-a and chlorophyll-b content in shaded tea plants grown under weak-light conditions. The chlorophyll-a/b ratio was used as an indicator to monitor the growth status and health level of tea plants, thus improving the management of shaded tea cultivation. In their study, the DBN was applied to hyperspectral imaging data and demonstrated superior performance in predicting the chlorophyll-a and chlorophyll-b content in tea leaves.

Carotenoids are photoprotective pigments which play a key role in assessing the environmental stress responses of tea plants. Carotenoids also demonstrate antioxidant properties, contributing to the nutritional quality of tea (Yang et al. 2021b). Sonobe and Hirono (2023) proposed a mini-spectrometer-based system to estimate the carotenoid content in tea leaves, and reliable prediction results were achieved using both Cubist and 1D-CNN DL models. Moreover, by introducing Gaussian and spike noise into the reflectance data, it was demonstrated that the 1D-CNN model exhibited greater stability as noise intensity increases. This approach can be used for effective prediction of carotenoid content in tea leaves under real cultivation conditions. Tea buds, the youngest part in the tea plant, are generally selected for producing high-quality teas (Xu et al. 2022). The quality of tea buds is highly associated with the fertilizer being applied. Fertilizer with insufficient nitrogen content can restrict the improvement of tea bud quality, whereas excessive nitrogen also has negative effects. Zhang et al. (2024b) combined NIRS with DL methods to rapidly predict the quality characteristics of various tea bud cultivars, aiming to optimize nitrogen fertilizer use. This study validated the feasibility of using RTA (the ratio of amino acids (AA) to total phenolics (TP) as an indicator for tea bud quality and nitrogen fertilization status. A convolutional neural network model, named “TeabudNet”, was designed to predict TP, AA, and RTA contents in tea buds. Correlation coefficients (r_p) of 0.924, 0.936, and 0.962 were achieved, respectively, which outperformed traditional ML methods. Furthermore, ResNet-18 was applied to classify the nitrogen fertilization status of tea buds and tea powders. This research shows great potential for real-time monitoring of tea quality and precise nitrogen fertilizer management using NIRS and DL.

Pests and diseases are the two main threats to tea plants. Proper pesticide application is necessary for pest and disease control. However, pesticide residues in teas can greatly affect human health under long-time tea consumption. To detect pesticide residues in tea, Zhu et al. (2021) proposed a method that combines surface-enhanced Raman scattering (SERS) with a 1D-CNN for the identification of multiple pesticide types. In their study, SERS spectral data of 20 pesticides and of pure tea samples were collected using a handheld Raman spectrometer. Data augmentation techniques, including spectral shifting, random noise addition, and linear spectral combination, were employed to enrich the dataset. A 1D-CNN model was subsequently constructed, which demonstrated high accuracy and robustness for on-site identification of five representative pesticide residues, confirming its suitability for multi-class pesticide recognition. In addition, Li et al. (2023a) further developed a SERS-based method enhanced by gold-silver octahedral hollow cages (Au–Ag OHCs) for quantitative analysis of pesticide residues. Compared with traditional handheld Raman spectrometers, the self-fabricated Au–Ag OHCs significantly enhanced the SERS signal, thereby improving the sensitivity of pesticide detection. The results showed that, when coupled with a CNN algorithm, the proposed method achieved superior performance in detecting thiram and pymetrozine, with coefficients of determination for prediction (R²_P) achieved at 0.995 and 0.997, respectively. Furthermore, the results obtained were consistent with the HPLC reference result, highlighting the potential of using CNN-based approaches to analyze Raman spectral data for quantitative analysis of pesticide residues.

Table 1 summarizes the application of deep learning methods in tea quality assessment during cultivation and breeding. Research studies demonstrated that DL methods, such as DBN, applied to hyperspectral imaging data can effectively predict chlorophyll content during the tea growth process, which helps improve shading treatments for tea trees and ensures tea quality. CNN can effectively handle a large amount of data with noise incorporated from the environment during real-time tea monitoring, thereby enhancing the detection robustness. The ResNet model has also been used for rapid prediction of tea bud quality characteristics, optimizing the use of nitrogen fertilizer. Additionally, DL models in combination with Raman spectroscopy can detect low-concentration components such as pesticide residues by effectively capturing subtle features from tea data. However, current research on applying DL in tea quality detection during the cultivation and breeding stage is limited, with a primary focus on specific areas such as predicting chlorophyll and carotenoid content, classifying nitrogen fertilization status, and detecting and quantifying pesticide residue. The impact of other environmental factors, such as soil properties and moisture content, on tea quality can be explored in the future.

Tea processing

Tea processing involves multiple stages, with the aim of transforming freshly-picked leaves into drinkable products. Different processing methods can contribute to the development of distinct flavor in tea leaves (Bokuchava And Skobeleva 1980).

Fresh tea leaves classification

The tea processing steps begin with the plucking of tea leaves, young buds and tender leaves are typically selected and picked. However, freshly harvested tea leaves often consist of leaves with different tenderness levels. Applying the same processing methods for tea leaves with different tenderness will impact the final tea product quality. Therefore, it is crucial to classify fresh tea leaves into different tenderness levels before processing. Zhao et al. (2024) proposed an improved YOLOv8x model with spatial pyramid pooling cross stage partial connections (SPPCSPC) and CBAM modules (YOLOv8x-SPPCSPC-CBAM) for the quality grading of fresh tea leaves, using images collected with an industrial camera. This approach can address the issues (such as low grading accuracy) encountered in machine sorting and machine vision methods. The model incorporates a SPPCSPC module to enhance the detection accuracy for small tea leaf targets. Additionally, it employs a CBAM module to enhance the focus on key regions, enabling more precise extraction of fine details from fresh tea leaves. Results demonstrated that the model achieved outstanding performance in recognizing both scattered and stacked fresh tea leaves, with mean average precision (mAP) values of 98.2% and 99.1% obtained, respectively. This approach provides a technical solution for non-destructive grading of fresh tea leaves.

Withering and rolling process

Withering is one of the key steps at the initial stage of tea processing. With the evaporation of moisture from tea leaves, fresh and tender tea leaves are transformed into a pliable texture for subsequent rolling and shaping. Precise control of the withering process is crucial for tea production, as both insufficient and excessive withering can negatively impact the quality of the final product. An et al. (2020) proposed a CNN-based moisture prediction model using confidence estimation based on tea images captured by an industrial camera. The model automatically extracted deep moisture-related features from raw images, without relying on conventional color and texture indicators. The model demonstrated excellent performance on an external validation set, achieving a correlation coefficient (r_p) of 0.9957 and a root mean square error of prediction (RMSEP) of 0.0059 (fractional moisture content), outperforming traditional PLSR and support vector regression (SVR) methods. However, the model accuracy shows a certain dependency on high-resolution images obtained from the computer vision system. Rolling is a crucial step following withering, where external forces are applied to alter the shape of tea leaves while facilitating the release of aromatic compounds. Tea strip-making is a specific step in the rolling process, involving the use of mechanical force at high temperature to mold tea leaves into a distinctive strip-like shape. Wang et al. (2024a) developed a dual-modality detection system that integrates machine vision with NIRS to automatically monitor the degree of strip-making in tea leaves. Among three classification algorithms evaluated, the BPNN developed based on spectral features achieved the best classification performance, with an average classification accuracy of 98.45%, while the CNN model developed based on image features achieved only 80.62% accuracy. Although the system can accurately identify and rapidly classify tea leaves at different strip-making stages, the model performance was significantly reduced when fusing image and spectral features into a single model, primarily due to noise interference and model complexity.

Tea fermentation process

The fermentation process is a crucial step in processing fermented teas such as oolong tea and black tea. The unique tea flavor can be developed during fermentation by controlling biochemical oxidation reactions, which alter both the external characteristics and internal composition of the tea leaves. Depending on the degree of fermentation, fermented tea can be categorized into different types, for example, black tea is classified as a fully fermented tea. Accurately assessing the fermentation degree of black tea is important, as both under-fermentation and over-fermentation can affect tea quality. Huang et al. (2024b) applied hyperspectral imaging (HSI) to acquire images of tea leaves at different fermentation stages and developed a classification model using a CNN-LSTM framework optimized via particle swarm optimization (PSO), achieving a classification accuracy of 96.78% using 2500 regions of interest (ROI)-based spectral blocks. The model outperformed traditional approaches such as SVM and partial least squares discriminant analysis (PLS-DA). However, the input structure still relied on conventional ROI extraction. As a result, the model exhibits certain limitations under practical conditions where tea leaves may stack together. To address this issue, Zhu et al. (2025a) introduced an improved CNN (3D-SwinT-CNN) model that integrates a three-dimensional swin transformer within the CNN architecture. By incorporating dilated convolution and a sliding-window self-attention mechanism, the model enhances global feature extraction while preserving spatial details. The model is trained and tested on 1600 hyperspectral data cubes, resulting in an accuracy of 98.13%, which outperformed the baseline 3D-CNN. This study demonstrated the strong generalizability of DL models in black tea fermentation classification. However, the complex DL architecture requires substantial training effort and computational resources, posing challenges for the deployment of DL in small-scale tea factories or in edge computing environments. To mitigate model complexity, a lightweight CNN-based method combined with knowledge distillation was applied on black tea images collected via machine vision (Ding et al. 2024b). A teacher–student network is constructed using EfficientNet_v2 and ShuffleNet_v2_x1.0. With limited image data augmentation (from 187 to 3740 samples) and the adoption of the mean guided distillation (MGD), the model achieves a classification performance of 92% while maintaining a compact architecture, thereby providing a lightweight modelling solution for classifying the fermentation levels in black teas. However, stable lighting conditions and high-resolution image acquisition equipment are required to ensure the model’s accuracy. On the other hand, monitoring the changes of key chemical components in real-time during fermentation can offer a more accurate result for the determination of fermentation degree. Luo et al. (2022b) combined SERS with chemometric methods to extract Raman fingerprint features from tea samples, and applied k-means clustering to divide fermented teas into distinct stages. A one-dimensional ResNet-18 (1D-ResNet18) model was then constructed for fermentation stages classification, achieving an accuracy of 83.33%. Additionally, a 1D-CNN model was used to quantitatively predict 10 tea quality indicators (e.g., catechin, epigallocatechin gallate (EGCG) and epicatechin (EC), with R²_P values ranging from 0.50 to 0.82. Although this approach offers advantages such as rapid and low-cost analysis for classifying tea fermentation stages, the signal stability of SERS remains a big challenge in dealing with large-scale tea production. In addition, quantitative analysis of black tea quality traits is constrained by the limited dataset available. To address this issue, Zhu et al. (2025b) proposed an improved deep convolutional generative adversarial network (DCGAN-L) for joint generation of hyperspectral data and corresponding labels, which effectively enhanced the performance of RF and broad learning system (BLS) models in quantifying the catechin content in tea. This study offers a potential solution to address the small-sample size issue encountered when using hyperspectral imaging data to perform regression tasks; however, further validation is needed to evaluate its applicability to real-world tea samples.

Oolong tea is classified as a semi-fermented tea, undergoing fermentation during the tossing process followed by fixation to halt fermentation. Zheng et al. (2024) proposed a data fusion system integrating visible and near-infrared (VIS-NIR) spectroscopy and computer vision for assessing the fermentation degree of oolong tea. The study employed SPA to select characteristic variables and fused the spectral and image data at the feature level. A CNN model based on the fused data achieved a prediction accuracy of 95.15%. Compared to single-sensor approaches, the fusion model significantly improved the classification performance. However, results showed that the SVM model outperformed the CNN model, indicating that traditional algorithms may be more effective than DL models when the sample size is small. Oxidation is a key process in the fermentation of oolong tea, which directly influences its flavor profile and the overall quality (Chen et al. 2010). Han et al. (2024) employed 13 gas sensors to collect the sensor signals of volatile organic compounds (VOCs) from tea leaves. PCA in combination with multiple classification models was applied for data analysis. Results showed that the back-propagation artificial neural network (BP-ANN) achieved the highest prediction accuracy of 94.11%, outperforming other DL models such as LeNet5 and AlexNet.

Drying and roasting process

Drying is a crucial step in tea processing, aiming at reducing the moisture content of tea leaves to extend shelf life and ensure product quality (Li et al. 2024a). You et al. (2024) developed a method using computer vision and a 1D-CNN to predict moisture content and visualize moisture distribution during the drying process of Tencha tea. The method extracts color space features from images and then Z-score normalization was applied, achieving a high r_p of 0.9548. Although the model demonstrates high accuracy, the study only consists of 150 samples for model development, which may limit its generalization capability due to the relatively small dataset size. In contrast, Chen et al. (2022) conducted a study on sun-drying of Pu-erh tea by integrating image information and environmental parameters to achieve dual prediction of moisture content and sensory quality traits using a combination of CNN and GRU. This study utilized 546 samples across different batches and time periods, resulting in a representative dataset collected under diverse sampling conditions. Additionally, to balance model interpretability and accuracy, high-level image features were extracted using a trained RexNet model, and feature selection was performed using neighborhood component analysis (NCA). Roasting is another process that can reduce moisture content in tea leaves, appropriate roasting techniques can further enhance tea quality. Huang et al. (2024a) investigated the use of a colorimetric sensor array (based on tetraphenylporphyrin (TPP) dyes modified with porous materials) to monitor the roasting quality of large-leaf yellow tea. Combined with a CNN model, the system achieved 100% accuracy in identifying the roasting grades of tea.

Table 2 summarizes the application of DL in quality assessment during the tea processing stage. Tea processing techniques impart unique flavors to tea, with each type of tea following a specific processing method. The variability in tea quality caused by different processing techniques can be effectively addressed using DL, due to the capability of DL for automatic extraction of high-dimensional features and superior nonlinear modeling characteristics. When integrated with intelligent sensing technologies such as computer vision, electronic noses, and electronic tongues, DL demonstrates significant advantages in extracting and analyzing both the appearance characteristics and internal compositional indicators of tea. After harvesting, tea leaves typically undergo key processes such as withering, rolling, fermentation, and drying. Currently, the application of DL in tea processing is mainly concentrated on the fermentation and drying stages. HSI combined with DL has shown great potential in the detection of tea fermentation degrees. Models such as CNN-LSTM can capture both spatial and spectral features of tea hyperspectral data, enabling high-precision monitoring of fermentation stages. The 3D-SwinT-CNN model enhances global feature extraction while preserving spatial information, thus improving the classification accuracy for black tea with different fermentation stages. Meanwhile, the DCGAN-L model can simultaneously generate tea hyperspectral data and corresponding labels, addressing the small sample size issue for model training. Furthermore, the combination of computer vision technology and CNN also demonstrates high accuracy in predicting moisture content during the tea drying process. Existing studies have gradually shifted towards a combined analysis of appearance features (such as color and morphology) as well as chemical quality indicators (such as catechins and moisture content). With the advancement of data collection technologies in the near future, the application of multi-modal fusion DL is expected to be a key research direction for enhancing the intelligent monitoring capabilities in tea processing.

Tea product evaluation

The quality control and evaluation of final tea products are critical for the development of the tea industry to ensure health benefits and market values. Tea product evaluation includes tea variety classification, quality grade identification, key component prediction, geographical origin identification, and storage time prediction.

Tea variety classification

Based on the tea cultivars and different processing methods, tea is classified into various types. Zhu et al. (2024) combined computer vision techniques with deep CNNs (such as ResNet-50 and MobileNet) to classify five varieties of oolong tea. The study constructed three types of datasets—feature images, cropped images, and augmented images—through manual image cropping, sliding window techniques, and random image transformations. Among them, the augmented image dataset significantly alleviated the overfitting problem caused by limited sample size, with ResNet-50 achieving the best performance and a Top-1 accuracy rate exceeding 93%. However, this method still relies on using standardized lighting conditions to acquire high-quality images. To address the adaptability issues of traditional computer vision methods under varying lighting conditions, Wei et al. (2022) proposed a tea variety classification approach based on LED-induced fluorescence imaging combined with the VGG16 model, achieving an identification accuracy of 97.5% across five tea categories. Compared to conventional RGB imaging, fluorescence imaging offers significant advantages in illumination consistency. It also effectively overcomes the limitations of traditional image processing in complex backgrounds by inducing characteristic feature responses.

In addition, Cui et al. (2022a) employed HSI to capture spectral images of three types of tea (i.e., black, green, and yellow tea). The wavelet feature maps derived from redundant discrete wavelet transform (RDWT), was fed into a lightweight CNN and SVM classifier to construct a fusion model for tea variety classification. By extracting multi-scale spectral information, the model achieved a classification accuracy of 98.7%, in the meantime, enhancing the model’s sensitivity to internal compositional differences among different tea types. However, the study was conducted to cover only three major tea categories. Although the samples were collected from multiple tea companies, they were mainly from the same tea production region. Moreover, the HSI system has a high equipment cost and is slow in data acquisition, making it a challenge to perform routine analysis in large-scale tea production. Furthermore, Zhong et al. (2019) proposed a DL method to differentiate different tea types. This method integrates an electronic tongue system with CNN-based automatic feature extraction (CNN-AFE). During data processing, short-time Fourier transform (STFT) was applied to convert the electronic tongue response into time–frequency spectrograms, which were subsequently processed by a CNN for feature extraction and classification. This method overcomes the instability associated with traditional handcrafted feature extraction and achieves a high classification accuracy of 99.9%. However, the system relies on a complex sensor array, resulting in relatively high maintenance costs. In comparison, Liu et al. (2020) employed a DNN combined with smartphone-based image acquisition to achieve rapid classification of seven commercially available chrysanthemum tea types. Without relying on specialized hardware, the DNN achieved classification accuracies of 96% for flowering period and 89% for cultivar identification, outperforming traditional morphological feature-based methods. This study highlights the flexibility of computer vision systems (CVS) and provides a generalized framework for large-scale tea identification. Additionally, Zhao et al. (2023) proposed an end-to-end model for the rapid classification of black tea, white tea, and green tea by integrating electrochemical fingerprints with a 1D-CNN. The study achieved a classification accuracy of over 98% on 6000 synthesized fingerprint samples. Similarly, Dong (2024) developed a recognition system which combines electrochemical sensors with a CNN, enabling rapid classification of nine different types of tea with an accuracy of 96.3% achieved. The system has a relatively low cost and simple structure, making it suitable to be deployed for tea product evaluation during distribution. However, data acquisition by the electrochemical sensors may be affected by environmental noise.

Tea grade classification

Tea products of different grades are sold at different prices, with tea appearance as a key factor in the grading process. Zhang et al. (2023a) developed a tea grade classification model that integrates lightweight deep CNN with transfer learning, using the tea data collected from a self-designed computer vision system. Specifically, they constructed three lightweight CNN models with different input sizes based on images of Wuyi black tea and Zhuyeqing green tea. The models were then fine-tuned using transfer learning on AlexNet and ResNet50 architectures. Among the models, those with input sizes of 32 × 32 and 64 × 64 achieved a balance between classification accuracy and computational efficiency. However, due to the limited sample size, the models might be overfitting. To address the limitation of small sample sizes and the lack of regional diversity, Zhang et al. (2023b) developed a Longjing tea grade recognition framework based on MobileNet V2 and instance-level transfer learning. A mixed training set consisting of small samples from multiple production regions was used to train the feature extractor, followed by classification using a multi-class TrAdaBoost combined with SVM. The proposed method achieved classification accuracies of 93.6% and 91.5% for Qiantang and Yuezhou Longjing tea datasets, respectively. In addition, Guo et al. (2024a) proposed an approach that integrates image processing with DL to evaluate the appearance of black tea. By introducing an improved Inception network enhanced with MBConv modules, the model achieves lightweight representation and yields a classification accuracy of 95% based on transfer learning. The study demonstrates that color and morphological features can be used to effectively distinguish between different grades of black tea, which has great potential for on-site tea grading applications. However, the model was trained and tested on a single tea variety, and its generalization capability across diverse tea types remains to be validated.

Complementary to appearance-based recognition techniques, Yang et al. (2021a) applied NIRS to collect spectral data of tea samples at different quality grades, and introduced an innovative approach by converting spectral data into pseudo-image formats to construct three DL models—TeaNet, TeaResNet, and TeaMobileNet. These models achieved 100% classification accuracy in tea grade identification. To comprehensively utilize hyperspectral image features for tea quality assessment, Ding et al. (2024a) combined both the spectral and images information derived from hyperspectral imaging for grading Huangshan Maofeng and Dianhong Gongfu tea, by integrating spatial principal component images extracted via PCA into a ResNet-50 model. With transfer learning and hyperparameters tuning using two-strategy particle swarm optimization (TSPSO), the DL model achieved an improved classification accuracy of 92.31%. Furthermore, Zheng et al. (2023) proposed an augmentation strategy for electronic tongue data, namely the computational model of taste pathways with time-channel expansion (CMTP-TCE). This method simulates biological gustatory pathways to transform low-dimensional sensor signals into high-dimensional perceptual information. Combined with a dot-product attention mechanism and a residual network (DPAM-ResNet), the classification of tea into five quality grades was achieved. Based on biomimetic data augmentation, the method enhanced sample diversity and achieved an accuracy of 97.66% on the augmented dataset.

Tea quality evaluation based on a single sensing modality is not sufficient to comprehensively assess the overall multidimensional quality attributes of tea—such as color, aroma, taste, composition, and shape. Data fusion is a strategy to address this issue. Song et al. (2021) introduced a feature-level fusion strategy that integrates NIRS with computer vision data to evaluate the overall quality of Keemun black tea. By using a CNN to extract fused features and employing a softmax classifier, the model achieved a prediction accuracy of 100% for tea grade classification. This approach significantly improved the accuracy of comprehensive quality evaluation for Keemun black tea and demonstrated the advantage of multimodal information fusion in enhancing discriminative capability. Building upon previous work, Liang et al. (2025) further introduced a temporal convolutional network (TCN) to construct a comprehensive tea quality evaluation framework. The improved Inception model was employed for appearance grade recognition, while a PLSR model was used to extract taste-related factors. These were combined with key spectral features processed by the TCN, resulting in a total of nine critical quality indicators. The integrated model achieved a high classification accuracy of 98.2%. By performing cross-modal and multi-stage feature fusion, the study enabled holistic modeling of tea quality, making it one of the most advanced integrated approaches for overall tea quality evaluation including appearance, composition, and taste. Similarly, Ren et al. (2024) developed a DL model for black tea grade classification by integrating data from NIRS, electronic eye, electronic tongue, and electronic nose sensors. The CNN-based model achieved a classification accuracy of up to 99.14%. However, while the fusion model delivers high accuracy, data synchronization and standardization across heterogeneous sensors is a challenge that hinders its practical application.

Tea composition prediction

The concentration of certain chemical components in tea is an important indicator for tea quality evaluation. Li et al. (2022a) employed NIRS combined with a 1D-CNN model to quantitatively analyze the sugar content in Huangshan Maofeng tea. The tea spectra were preprocessed with standard normal variate (SNV), and the model achieved an average prediction accuracy of R²_P = 0.95. In addition, He et al. (2021) applied near-infrared hyperspectral imaging (NIR-HSI) to acquire spectral images of 150 fresh and dried chrysanthemum flower tea samples. Wavelength selection algorithms and a CNN regression model were applied to simultaneously detect five trace components. Although the proposed method outperformed traditional PLSR and SVR models in predicting flavonoid compounds, the prediction accuracy for other components (such as buddleoside and apigenin) is relatively low, which might be due to the lack of sufficient training data and the relevance of the selected wavelengths used in modelling. Accordingly, applying suitable wavelength selection and attention mechanisms can help to effectively capture the spectral features of active components. Zhang et al. (2024a) combined NIRS with DL and developed a FICSS-CNN-CSAM model, which is an integration of wavelength selection (FICSS) and channel–spatial attention mechanisms (CSAM). The model achieved high prediction accuracy for four catechins in black tea, with all components yielding R²_P > 0.92. This approach enhances the model’s capability to capture fine-grained variations in chemical composition. However, the sample data used in the study were collected from one production region, which presented a limitation of the study for tea samples from other regions. In addition, utilizing integrated features from hyperspectral images can provide a more comprehensive component-related sample information. Luo et al. (2022a) employed 1D-CNN and 2D-CNN architectures to extract spectral and spatial features from green tea images, and developed a RF model to predict tea polyphenol content. However, as an ensemble learning method, the RF model exhibits limited computational efficiency, making it less suitable for large-scale, real-time tea quality assessment.

During the tea brewing process, aroma and color are key indicators for determining the tea quality in terms of taste and flavor. Yang et al. (2025) conducted a systematic investigation into the flavor dynamics of two oolong tea varieties—Shuixian and Rougui—during consecutive infusions using the traditional “Gongfu tea” brewing method. Three analytical techniques were employed: GC-MS, gas chromatography-olfactometry-mass spectrometry (GC-O-MS), and atmospheric pressure chemical ionization tandem mass spectrometry (APCI-MS/MS). A total of 48 aroma compounds were identified using GC-MS, with Rougui tea exhibiting a higher overall aroma intensity. Based on GC-O-MS results, an aroma wheel comprising eight categories of olfactory descriptors was constructed to illustrate the sensory distinctions between the teas. Real-time aroma monitoring was achieved through APCI-MS/MS, and predictive modeling of aroma and color changes over seven brewing cycles was performed using multivariate polynomial regression (MPR) and LSTM. This study demonstrates the potential of DL in flavor characterization, while also highlights its dependence on high-quality sensory data.

Tea origin identification

In recent years, mislabeling of tea origin has become a major concern, as the geographical origin significantly influences tea quality. Chang et al. (2022) combined NIRS with DL algorithms to classify 400 Maofeng tea samples from four different origins. The RepSet model achieved the best classification performance with an accuracy of 99.30%, outperforming both the BPNN and the modified AlexNet models. This study demonstrates the advantages of NIRS in non-destructive tea testing and highlights the strong feature recognition capability of DL models. Electronic nose technology is currently playing an important role in tea quality evaluation comparing to human olfactory assessment methods. Chang and Lu (2024) collected gas feature data of spring teas from six different production regions using electronic nose technology and developed a classification model based on a residual lightweight channel-spatial attention network (RLCSA-Net). By integrating a lightweight channel-spatial attention (LCSA) mechanism with residual dense blocks (RDB), RLCSA-Net effectively enhanced feature extraction with a small-sample size, achieving an accuracy of 98.08% in identifying spring teas from different regions. Yu and Gu (2021) proposed a hybrid DL approach to identify green teas from 12 different origins. The study utilized a ResNeXt-based CNN to extract deep features from electronic nose signals, followed by an SVM classifier to enhance discriminative performance for such a small dataset.

Tea storage time identification

For fermented tea, the storage time and storage condition can impact tea quality, as different types of tea undergo distinct quality changes during the storage and aging process (Lv et al. 2023). For example, black tea is a fully fermented tea which experiences a reduced quality over extended storage periods. It is unfortunate that, in some cases, “over-storage” black tea are sold as premium “fresh” black tea through fraudulent sales or mislabeling, as consumers are unable to identify the tea quality themselves. Hong et al. (2021) used NIR-HSI combined with DL methods to distinguish black tea samples with four different storage years. Wavelet transform (WT) was applied for denoising, and PCA was used for feature extraction and waveband selection. Calibration models were developed using full-spectrum data or using key wavelengths extracted from principal component loadings, result showed that CNN, LSTM, and the combined CNN–LSTM models all outperformed traditional ML methods such as logistic regression (LR) and SVM. However, the study also pointed out that the performance of DL models was limited by the insufficient sample size. Although techniques such as PCA and wavelength selection can improve model stability, the results indicated that the employed DL models were heavily dependent on specific preprocessing procedures, thereby increasing the complexity for practical deployment.

Unlike black tea, the flavor and quality of dark tea improve gradually during storage. Yang et al. (2021d) extracted signal features from Pu-erh tea samples with five different storage years using a self-developed voltametric electronic tongue (VE-Tongue) system. Based on a transfer learning-enabled 1D-CNN model, a high recognition accuracy of 98.80% was achieved. Compared to traditional ML methods, the introduction of transfer learning significantly improved model performance even the sample size is small. In a further study, Yang et al. (2021c) combined electronic tongue and electronic eye technologies to identify dark tea with different storage time/year. The 1D-CNN and 2D-CNN were applied to extract liquid and appearance features of dark tea, followed by feature-level fusion for differentiation of the tea storage time. On this basis, a BPNN classifier was optimized using a Bayesian optimization algorithm (BOA). This method achieved an accuracy of 99.07% on the test set, outperforming unimodal models and demonstrating the potential of multi-sensor fusion in tea identification. However, the inconsistent signal features derived from different sensors remain a challenge in the implementation of data fusion strategies. For dark tea being stored under appropriate conditions, the brand and aging time are two main factors determining its market price. To address the issue of counterfeiting high-value dark tea, Tan et al. (2024) combined excitation-emission matrix (EEM) fluorescence spectroscopy with a CNN for dual-task identification of dark tea brand and its aging period. Ultimately, a ResNet-based model was developed which achieved a brand recognition accuracy of 96.9% and an aging period recognition accuracy of 87.5%. The study leveraged transfer learning and multi-task learning to enable feature sharing and knowledge transfer between brand and ageing classification tasks, significantly improving the model’s generalization performance with the use of a small sample dataset.

Other applications

The quality of tea varies significantly across different harvesting periods, with spring tea being particularly valuable economically. Kang et al. (2023) proposed a method combining electronic nose technology with an adaptive pooling attention mechanism (APAM) to identify the quality of spring tea harvested at different times. APAM enhances the feature learning capacity for gas information by integrating adaptive multi-scale pooling with efficient interactive convolution. When coupled with the lightweight MobileNetV1 model, this approach achieved a classification accuracy of 97.67% on a small dataset of 240 samples. These results demonstrated the applicability of APAM method for non-destructive gas recognition tasks using a small sample size. However, compared to traditional attention mechanisms, the computational complexity and integration cost of such deep attention modules may limit their deployment in real-world applications.

In addition, mixing non-authentic materials into high-value tea is an illegal adulteration practice that severely compromises the tea quality, safety, and consumer trust. Xu et al. (2019) integrated electronic nose and VIS–NIR technologies with a CNN model to enable rapid detection of Pu-erh tea types, blending ratios (i.e., proportions of different tea materials), and adulteration levels (i.e., the amount of non-authentic substances added). The study demonstrated that CNNs outperform traditional methods such as LDA and PLSR in extracting local features. However, CNNs also tends to produce redundant information, which may undermine the overall detection performance. This issue is particularly noticeable in multimodal fusion, where feature redundancy and noise amplification significantly increase the difficulty of model training and the complexity of model deployment. Another concern regarding tea adulteration is mixing tea waste into high-quality tea and selling it at the price of high-quality tea. To address this issue, Besharati et al. (2024) combined RGB image processing with DL to evaluate the level of adulteration with tea waste in Iranian black tea. The study employed an image patching approach together with the MobileNet V3 model, achieving an identification accuracy of 95% on patch images. Although this method can increase the dataset size, small image patches may lack representativeness and fail to provide sufficient contextual information for accurate adulterant quantification, ultimately resulting in a reduced classification performance.

An innovative research study was conducted by Huang et al. (2023a) who investigated the relationships between different types of tea and its suitable target groups, using DL approach. An ID_Tea knowledge graph was firstly developed, encompassing 330 tea varieties and 29 consumer suitability categories. The authors proposed an enhanced link prediction model, IntGCN, which integrates graph convolutional networks (GCN) and squeeze-and-excitation networks (SENet) within the InteractE framework, to explore the compatibility between tea types and consumer characteristics. Compared to traditional graph neural network models, IntGCN leverages transfer learning to facilitate knowledge migration from general datasets to domain-specific knowledge graphs, thereby significantly improving the accuracy of suitability relationship predictions. However, it should be noted that high-quality knowledge graph data is key for ensuring model performance and robustness.

Table 3 summarizes the application of DL in the quality assessment of final tea products. By leveraging transfer learning, DL methods effectively mitigate the issue of limited dataset samples while enhancing the ability to distinguish different tea varieties, quality grades, and storage times. In tea variety recognition, the CNN-based automatic feature extraction method, CNN–AFE, overcomes the limitations of manual feature extraction, significantly improving the recognition accuracy. Additionally, by introducing the MBConv module into the improved Inception network, a lightweight DL model can be developed to effectively identify different grades of black tea with the use of computer vision technology. CNN plays a significant role in detecting tea storage time, particularly for processing data from various sensors. Furthermore, the incorporation of attention mechanisms has enhanced the feature extraction capabilities of DL models, thereby improving the accuracy of tea quality identification and component analysis. To differentiate similar tea products, multi-source information such as appearance and compositional features or spectral and spatial data, can be integrated into DL models to enhance the model reliability for quality assessment. However, current research on tea quality evaluation primarily focuses on the tea variety classification and intrinsic quality attributes prediction based on a relatively small sample set. Future research could explore the suitability of DL models for wider applications and the deployment of DL-based methodologies in the tea industry.

Table 1. Application of various deep learning methods for tea quality evaluation during cultivation and breeding

Task	Tea type	Specific techniques	DL model	Purpose	Number of datasets	Model performance	References
Chlorophyll content detection	Green tea	Hyperspectral remote sensing	DBN	Prediction of chlorophyll content in shaded tea leaves	77	Mean prediction accuracy for chlorophyll-a: RPD = 3.00; RMSE = 8.04 µg/cm2. chlorophyll-b: RPD = 2.81; RMSE = 2.17 µg/cm2.	(Sonobe et al. 2020)
Carotenoid content detection	21 Japan teas	Compact spectrometer	1D-CNN	Evaluation of carotenoid content in tea leaves	702	Prediction accuracy of carotenoid content: RPD = 1.50; RMSE = 1.03; R2 = 0.56.	(Sonobe And Hirono 2023)
Nitrogen content detection	Baiye 1, Fuding Dabai, Zhonghuang 1, Zhonghuang 2, Longjing 43, Zijuan	NIRS	CNN (TeabudNet), ResNet-18	Rapid assessment of quality components in tea powder and nitrogen status in tea buds and powder	1120,742	Tea powder composition prediction (Rp): TP:0.924, AA: 0.936, RTA:0.962. Nitrogen status classification accuracy: tea powder: 96.64%, tea bud: 92.62%.	(Zhang et al. 2024a, b)
Pesticide residue detection	Longjing tea	SERS	1D-CNN	Rapid on-site detection of pesticide residues in tea leaves	5000 × 20 classes	Pesticide residue identification accuracy: 100%.	(Zhu et al. 2021)
	Zhenjiang tea	SERS	1D-CNN	Quantitative detection of thiram and pymetrozine pesticide residues in tea	105	Pesticide content prediction for thiram: RPD = 14.68; RMSEP = 0.132 µg/mL; Rp2=0.995. pymetrozine: RPD = 6.66; RMSEP = 0.291 µg/mL; Rp2=0.977.	(Li et al. 2023a)

Table 2. Application of various deep learning methods for tea quality evaluation during tea processing

Task	Tea type	Specific techniques	DL model	Purpose	Number of datasets	Model performance	References
Fresh tea leaves classification	Green tea	Image acquisition	YOLOv8x-SPPCSPC-CBAM	Classification of fresh tea leaves by six quality grades	2800,2800	Quality grade classification for Scattered images: P = 0.958; R = 0.967; mAP = 0.982. Stacked images: P = 0.991; R = 0.977; mAP = 0.991.	(Zhao et al. 2024)
Withering and rolling process	Jin Guanyin black tea	Image acquisition	CNN	Rapid prediction of moisture content in withering black tea	3060	Moisture content prediction: RPD = 9.5781; RMSEP = 0.0059; rp=0.9957.	(An et al. 2020)
	Shucha Zao, Baiye 1, Longjing 43	NIRS, Image acquisition	CNN	Classified evaluation of the strip shaping degree of processed tea products	1073	Accuracy of strip-lining degree classification (Mean ± SD): 80.62% ± 1.49%.	(Wang et al. 2024a)
Tea fermentation process	Black tea	Image acquisition	Shufflenet_v2_x1.0, Efficientnet_v2	Evaluation of black tea fermentation across six degrees	3740	Black tea fermentation recognition: P = 0.9208; R = 0.9190; F1 = 0.9192.	(Ding et al. , b)
	Ying Hong 9	HSI	PSO-CNN-LSTM	Evaluation of black tea fermentation across five degrees	2500	Black tea fermentation recognition: Accuracy = 96.78%.	(Huang et al. , b)
	Jiukeng black tea	HSI	3D-SwinT-CNN	Evaluation of black tea fermentation across four degrees	1600	Black tea fermentation recognition: Accuracy = 98.13%; P = 99.00%; R = 98.30%; F1 = 98.65%.	(Zhu et al. 2025 a)
	Fengqing Daye	SERS	1D-ResNet18, 1D-CNN	Evaluation of black tea fermentation across five degrees and quantification of quality parameter changes during fermentation	360	Black tea fermentation recognition: Accuracy = 83.33%; best prediction for EGCG: Rp2=0.82.	(Luo et al. , b)
	Jiukeng black tea	HSI	DCGAN-L	Quantitative analysis of catechin content in fermented black tea	705,305,405,405	Best prediction result (Rp2): CC = 0.8426, C = 0.8661, ECG = 0.9033, EGC = 0.3598.	(Zhu et al. 2025b)
	Iron Goddess tea	NIRS, Image acquisition	CNN	Evaluation of oolong tea fermentation across three degrees	315	Oolong tea fermentation recognition: Accuracy = 95.15%.	(Zheng et al. 2024)
	Iron Goddess tea	Gas sensor	LeNet5, AlexNet	Monitoring aroma changes during the oxidation process of oolong tea and distinguishing four oxidation stages	256	Recognition of oxidation stages of oolong tea: Accuracy = 91.17%.	(Han et al. 2024)
Drying and roasting process	Tencha	Image acquisition	1D-CNN	Prediction of moisture content during the drying process of Tencha	150	Moisture content prediction: RPD = 3.22; RMSEP = 61.7 g/kg; rp=0.9548.	(You et al. 2024)
	Pu-erh tea	Image acquisition	CNN, GRU	Prediction of moisture content during the sun-drying process of Pu-erh tea and its final product quality	546	Moisture content prediction: RPD = 26.3513; RMSEP = 0.7508; Rp2=0.9986. The RMSE of product quality prediction ranges from 0.13 to 0.31.	(Chen et al. 2022)
	Large-leaf yellow tea	Colorimetric sensor	CNN	Monitoring the roasting quality of large-leaf yellow tea	240	Achieved 100% classification accuracy on both calibration and prediction sets.	(Huang et al. 2024a)

Table 3. Application of various deep learning methods for quality evaluation of tea products

Task	Tea type	Specific techniques	DL model	Purpose	Number of datasets	Model performance	References
Tea variety classification	Oolong tea	Image acquisition	ResNet50	Classification of different oolong tea varieties	1430	Top-1 accuracy:93.2%; Top-5 accuracy:99.5%.	(Zhu et al. 2024)
	Green and oolong teas	Fluorescence imaging	VGG16	Classification of 5 tea varieties	760	Classification accuracy: 97.5%.	(Wei et al. 2022)
	Yellow, green and black teas	NIR-HSI	L-CNN-SVM	Classification of three types of tea	450	Overall classification accuracy: 98.7%.	(Cui et al. 2022a)
	Yuzhu, Show bud, Iron Goddess, Biluochun and Westlake teas	Electronic tongue	CNN-AFE	Classification of five tea varieties	5100	Tea classification accuracy: 99.9%.	(Zhong et al. 2019)
	7 commercial chrysanthemum teas	Image acquisition	DNN	Identification of flowering stages and tea types of different chrysanthemum teas	1581	Flowering stage identification: 96%; Chrysanthemum tea variety classification: 89%.	(Liu et al. 2020)
	Black, white and green teas	Electrochemi-cal fingerprinting	1D-CNN	Rapid identification of 3 tea types	6000	Test set classification accuracy: 98.08%.	(Zhao et al. 2023)
	9 commercial teas	Electrochemi-cal profiles	CNN	Classification of nine tea varieties	270	Classification accuracy: 96.3%. P = 0.96; R = 0.95; F1 = 0.95.	(Dong 2024)
Tea grade classification	Wuyi black tea and Zhuyeqing green tea	Image acquisition	CNN, AlexNet, ResNet50	Identification of three quality grades across two tea types	1217	The simplified CNN achieved accuracy comparable to transfer learning models (Wuyi black tea: 92.33%, Zhuyeqing: 95.47%).	(Zhang et al. 2023a)
	Longjing tea	Image acquisition	MobileNet V2	Quality identification of four grades of Longjing tea from two regions	1612	Classification accuracy: Qiantang Longjing:93.6%; Yuezhou Longjing:91.5%.	(Zhang et al. 2023b)
	Yingde black tea	Image acquisition	Inception	Classification of black tea into three appearance quality grades	1100	Test set classification accuracy:95.00%; Independent validation accuracy:97.22%.	(Guo et al. 2024a)
	Black and Green teas	NIRS	CNN (TeaNet, TeaResnet, TeaMobilenet)	Identification of 50 quality grades from 21 tea brands and 24 grades from 6 Longjing brands	1000,480	The main dataset (50 categories) accuracy:100%. The Longjing subset accuracy: 99.2%.	(Yang et al. 2021a)
	Huangshan Maofeng and Dianhong Gongfu teas	HSI	TSPSO- ResNet50	Identification of 13 quality grades across two types of tea (Huangshan Maofeng: 6 grades; Dianhong Gongfu: 7 grades)	260	Test set classification accuracy: 92.31%.	(Ding et al. 2024a)
	Fujian tea	Electronic tongue	DPAM-ResNet	Classification of tea quality	200	Classification accuracy:97.66%; F1 = 97.68%.	(Zheng et al. 2023)
	Keemun black tea	Image acquisition, NIRS	CNN	Differentiation of seven grades of Keemun black tea	560	Test set classification accuracy: 100.00%.	(Song et al. 2021)
	Yingde black tea	Image acquisition, NIRS	Inception, TCN	Rapid assessment of three quality grades during black tea production	355	Test set classification accuracy: 98.2%.	(Liang et al. 2025)
	Congou black tea	NIRS, electronic eye, electronic nose, electronic tongue	CNN	Evaluation of 7 tea quality grades	700	Test set classification accuracy: 99.14%.	(Ren et al. 2024)
Tea composition prediction	Huangshan Maofeng tea	NIRS	1D-CNN	Estimation of sugar content in Huangshan Maofeng tea	158	Sugar content prediction: RMSEP = 1.16; Rp2=0.95.	(Li et al. 2022a)
	Chrysanthemum tea	NIR-HSI	CNN	Rapid prediction of five chemical component contents in chrysanthemum	150	Relatively accurate predictions for luteolin and quercetin. Luteolin fresh samples: Rp2 = 0.9284; RMSEP = 0.0258 mg/g; RPD = 3.80, dry samples: Rp2 = 0.9265; RMSEP = 0.0261 mg/g; RPD = 3.75. Quercetin fresh samples: Rp2 = 0.8385; RMSEP = 0.0209 mg/g; RPD = 2.53, dry samples: Rp2 = 0.8556; RMSEP = 0.0198 mg/g; RPD = 2.68.	(He et al. 2021)
	Jinjunmei, Maojian, and Yunwu teas	NIRS	FICSS-CNN-CSAM	Prediction of four major catechin contents in black tea	105	Prediction results for EC: R2 = 0.92; RMSE = 0.018. ECG: R2 = 0.96; RMSE = 0.11. EGC: R2 = 0.97; RMSE = 0.14. EGCG: R2 = 0.97; RMSE = 0.32.	(Zhang et al. 2024a)
	Green tea	HSI	1D-CNN, 2D-CNN	Prediction of tea polyphenol content in green tea	140	Polyphenol content prediction: Rp2 = 0.938; RMSEP = 1.043 mg/g.	(Luo et al. 2022a)
	Oolong tea	GC–MS, GC-O-MS and APCI-MS/MS	LSTM	An innovative approach to characterizing the flavor profile of oolong tea infusion	84	Prediction of aroma release and color change in the seventh infusion: RMSE from 0.008 to 0.033. Combined prediction of the sixth and seventh infusions: RMSE from 0.014 to 0.053.	(Yang et al. 2025)
Tea origin identification	Maojian tea	NIRS	RepSet	Identification of Maojian tea from 4 origins	400	Classification accuracy: 99.30%.	(Chang et al. 2022)
	Spring green tea	Electronic nose	RLCSA-Net	Quality identification of spring green tea from 6 origins	180	Origin classification results (mean ± standard deviation): Accuracy= 98.08% ± 1.24; P = 97.80% ± 1.31; R = 98.35% ± 1.06.	(Chang And Lu 2024)
	Maofeng and Maojian teas	Electronic nose	CNN-SVM	Classification of 12 subcategories of Maofeng green tea and Maojian green tea	1440	Origin classification results: Accuracy = 96.11%; P = 96.86%; R = 96.11%; F1 = 96.03%.	(Yu And Gu 2021)
Tea storage time identification	Jinguanyin black tea	NIR-HSI	CNN, LSTM, CNN-LSTM	Rapid detection of 4 storage ages of black tea	60	Test accuracy: 83.60%.	(Hong et al. 2021)
	Pu-erh tea	Electronic tongue	1D-CNN	Identification of 5 storage durations of Pu-erh tea	1595	Test accuracy: 98.53%; P = 99%; R = 99%; F1 = 0.99.	(Yang et al. d )
	Pu-erh tea	Electronic tongue, electronic eye	1D-CNN, 2D-CNN	Identification of 5 storage durations of Pu-erh tea	1500	Test accuracy: 99.07%; P = 99.2%; R = 99.0%; F1 = 0.992.	(Yang et al. c)
	Dark tea	EEM fluorescence spectroscopy	ResNet	Identification of black tea brands and their aging periods	384	Brand identification accuracy: 96.9%; aging period identification accuracy: 87.5%.	(Tan et al. 2024)
Other applications	Spring tea	Electronic nose	MobileNet-V1 + APAM	Identification of tea quality from 6 harvesting periods	240	Harvesting periods classification results: Accuracy = 97.62% ± 0.75; P = 97.57% ± 0.83; R = 97.71% ± 0.77; F1 = 97.64% ± 0.61.	(Kang et al. 2023)
	Pu-erh tea	Electronic nose, VIS-NIR	CNN	Rapid detection of the type, blending ratio, and adulteration ratio of ancient Pu-erh tea	120,100,100	Tea type identification: R² = 0.81; blending ratio prediction: R² = 0.86; adulteration ratio prediction: R² = 0.92.	(Xu et al. 2019)
	Iranian tea	Image acquisition	MobileNet V3, EfficientNet V2	Identification of counterfeit Iranian black tea and quantification of adulteration in tea leaves	800	Recognition accuracy for four levels of tea adulteration: 95.00%.	(Besharati et al. 2024)
	330 tea types	Tea knowledge graph	IntGCN (Improved InteractE)	Prediction of the “suitability for consumers” relationship in different types of tea	6698	Prediction results: MRRa = 61.6%; H@10b = 75.0%.	(Huang et al. 2023a)

^aaverage reciprocal rank of the first correct result.

^bproportion of queries where the correct result appears in the top 10.

Challenges and future prospects

Challenges in deep learning applications for tea quality detection

Compared to traditional ML, DL methods offer significant advantages in tea quality assessment. For example, DL can automatically extract features from raw data, reducing the reliance on predefined preprocessing patterns and enabling end-to-end training. Meanwhile, the multi-level network structure in DL methods helps in extracting deep-level features from high-dimensional tea data. Furthermore, deep learning networks stand out in feature transformation and parameter sharing, making them widely applicable for spatial deep feature extraction. For instance, the features of tea obtained through hyperspectral imaging contain both spectral and spatial characteristics. DL methods can be applied to HSI data by integrating convolutional neural networks into different data dimensions, demonstrating their unique advantages in acquiring deep spectral-spatial features (Luo et al. 2022a). Secondly, the tea data acquired by sensors such as NIRS, electronic noses, and electronic tongues often contains a significant amount of noise and irrelevant information. DL methods, by introducing attention mechanisms, can help the model to effectively extract deep-level features that are most relevant to the target information while reducing unnecessary data interference (Chang And Lu 2024).

Nevertheless, the practical application of DL algorithms also faces numerous challenges. While complex network architectures can enhance model performance, DL models require high computational demands. Reducing the model computational time and enhancing its real-time data processing capabilities relies heavily on deep learning hardware (Wang et al. 2024b). For example, the deployment of models for tea quality detection in embedded devices or mobile applications has increased in recent years. However, these devices typically have limited resources and cannot support the execution of complex, large-scale models. In the meantime, complex models often suffer from low interpretability, with the entire operational process being considered as a black box. Understanding, developing, deploying and maintaining DL models require the personnel to have relevant technical knowledge and skillset. However, there is a lack of skilled and well-trained professionals in the tea industry, which constrains the adoption of DL-based sensing technologies. In real world applications, the collected tea leaf images and sensor data are greatly affected by the detection environment (such as variations in illumination conditions, temperature fluctuation, humidity changes, and vibrational noise), which could negatively impact the model performance. Furthermore, in the actual tea industry, data related to tea quality are collected batch by batch and accumulated year by year. Therefore, existing DL models have to be updated periodically to capture new samples to ensure model robustness and applicability.

DL-based tea quality detection requires large, high-quality, and representative datasets. The diversity of tea types, geographic locations, and harvesting times can lead to significant variations in tea characteristics. To ensure strong model generalization, it is essential to collect more diverse and representative datasets. In addition, reference tea data (including tea quality classification and tea component measurement) requires not only experienced tea technologists but also sophisticated and high-cost analytical instruments (e.g., HPLC, GC-MS) (Tudu et al. 2009).

DL algorithms can address the issue of small dataset by employing GANs. For classification tasks, GANs can generate feature data and corresponding tea quality grades using the generator network. For regression tasks such as predicting specific chemical components, Zhu et al. (2025b) proposed an improved deep convolutional generative adversarial network variant, which enabled hyperspectral data augmentation and label generation for catechin content prediction in black tea. However, traditional GAN methods may introduce noise or distort the original data distribution, which can impact the reliability of the generated data. Moreover, since tea quality is influenced by multiple factors such as varieties, processing technologies, and tea aging time, it is challenging for existing GAN architectures to accurately capture the complex relationships between the influencing factors and their sensor-based response. To enhance the representativeness of the generated tea data, structural optimization of GAN must meet the requirements of the specific dataset. Furthermore, data privacy and data sharing are two main issues arising from profit-driven competition among tea enterprises, which further hinder the tea industry’s access to sufficient and diverse datasets for effective model training. To address this challenge, approaches such as developing a secured framework for data-sharing or adopting federated learning techniques can be applied, thus enabling enterprises to collaboratively train a DL model without the need to share their raw data.

When performing classification and regression tasks on specific tea datasets, DL models often employ highly complex network structures, which may lead to overfitting—where the model performs well on the training set but fails to generalize effectively to the test set. DL leverages transfer learning, where models are pre-trained on large-scale benchmark datasets in related domains or on similar large-scale tea datasets. The model parameters are then reused and fine-tuned, transferring the learned knowledge to the target tea dataset. This approach can mitigate the overfitting issue in DL models and reduce the complexity of model training. For example, Yang et al. (2021d) pre-trained a 1D-CNN model on a speech recognition dataset, followed by fine-tuning, the authors successfully applied the model to electronic tongue data of Pu-erh tea, enabling the effective identification of tea storage time. Similarly, Zhang et al. (2023a) fine-tuned the pre-trained AlexNet and ResNet50 models, which leveraged transfer learning for accurate tea quality classification. However, the current deep transfer learning approach primarily relies on adjusting pre-trained model parameters for knowledge transfer. While commonly used large-scale datasets (such as ImageNet) can provide good generalization, they lack task-specific optimization, which limits the transfer efficiency. Moreover, due to data distribution differences between the source domain (where the model is trained) and the target domain (where the model is applied), some samples in the source domain may not be suitable for transfer learning. Negative transfer results might occur if those unsuitable samples are included in transfer learning (Zhang et al. 2023b). On the other hand, studies have shown that utilizing pre-trained deep learning networks as feature extractors and combining them with traditional algorithms can prevent model overfitting. For instance, Yu and Gu (2021) extracted deep features from electronic nose data using CNN and employed an SVM classifier to enhance classification performance. The proposed CNN-SVM model successfully achieved fine-grained classification of green tea varieties. The CNN-SVM is an integrated model combining DL and traditional ML, it offers strong interpretability even though a significant effort is required during the modeling phase.

Finally, integrating multiple sources of data can provide a comprehensive tea quality evaluation. DL-based tea quality detection must address the challenges of multimodal data fusion. Different types of data (such as image texture and color, near-infrared spectra, electronic tongue, and electronic nose signals) vary in its feature dimensions and may differ in spatial, temporal, and measurement characteristics, making it difficult to integrate these heterogeneous datasets for tea quality evaluation (Li et al. 2024b). Currently, the methods applied to multi-source tea data fusion include early fusion, intermediate fusion, and hybrid fusion. Early fusion integrates data from different sources at the input level. For example, Zheng et al. (2024) combined the preprocessed VIS-NIR spectral data and image features directly for DL model development to detect the fermentation stages of oolong tea. Intermediate fusion involves the integration of different types of data at the feature level after processing by different models. For instance, Yang et al. (2021c) used 1D-CNN and 2D-CNN to extract features from electronic tongue and electronic eye data, respectively. A BPNN was then applied to classify the fused feature data to identify the storage time of Pu-erh tea. Hybrid fusion combines the advantages of intermediate fusion and late fusion (combining the output results of different models) by integrating the extracted features from one data source and the decision results from another data source. For example, Liang et al. (2025) integrated the predicted tea appearance grades (derived from an Inception-based computer vision model), the key component content (derived from NIRS data using PLSR method) and the quality features (extracted from NIRS data using TCN method) to discriminate black tea quality.

Future prospects

Challenges associated with tea quality evaluation in real-world practical applications are driving the further development of DL technologies. Among the challenges mentioned in Sect. 6.1, the model complexity and its large hyperparameters limit the deployment of DL models on resource-constrained devices (i.e., small, hand-held, and low-cost devices). Researchers have proposed model compression techniques such as knowledge distillation to solve the hyperparameter issue, in which the knowledge of complex and high-performing large models (teacher models) can be transferred to lightweight smaller models (student models) through knowledge distillation. This enables “student models” to achieve comparable performance to that of “teacher models” while utilizing significantly fewer parameters (Ding et al. 2024b). New DL models such as Transformer may be more complex, future research could explore the potential of integrating the model with knowledge distillation techniques, using it as a teacher model to further enhance the effectiveness of knowledge distillation and support the intelligent processing of tea-related data.

Model interpretability is another challenge that hinders the application of DL-based non-destructive sensing systems in the tea industry. If the internal decision-making logic of a model cannot be understood and trusted by the operators, the system is unlikely to be widely adopted at the industry level, regardless of its performance. In response to this issue, a variety of interpretation methods for deep learning models have been proposed in recent years. For example, local interpretable model-agnostic explanations (LIME) is a model-agnostic local explanation method, which interprets the predictions of complex models by perturbing input samples and fitting a simple model within the local neighborhood. Meanwhile, shapley additive explanations (SHAP) provides a unified quantification of the importance of all input features, making it suitable to identify key factors that influence tea quality assessment (Aldughayfiq et al. 2023). Selvaraju et al. (2019) developed a gradient-weighted class activation mapping (Grad-CAM), which generates heatmaps using gradient information from the last convolutional layer to locate the most critical feature regions for a given prediction. Methods such as SmoothGrad and Integrated Gradients have been designed to reduce noise and instability in gradient-based visualizations (Li et al. 2022b).

In recent years, layer-wise relevance propagation (LRP) and its variants have become a research focus in model interpretability. LRP propagates relevance scores backward through the network layers to compute the contribution of each input feature to the final prediction. It is particularly suitable to be embedded in automatic detection systems, offering stable and visual interpretability. As an extension of LRP to GNNs, graph layer-wise relevance propagation (GLRP) shows strong interpretability for multi-sensor fusion models in the tea industry. Based on the output of the model, GLRP propagates relevance scores layer by layer to the input nodes, therefore, the contribution of each sensor modality to the final prediction can be quantified (ŞAhiN et al. 2024).

In addition, due to input data distribution shifts (caused by actual environmental conditions) and the growing demand for models to handle dynamic sampling data in the tea industry, incremental learning represents a promising future direction for developing robust and adaptive DL models. In practical scenarios, models need to have capability of “training, deployment, and updating” simultaneously. Incremental learning enables models to continuously acquire new knowledge without retraining the model from scratch, thereby reducing time and computational constraints to a certain extent (Diaz-Chito et al. 2017).

Moreover, DL-based research on tea quality detection requires a large number of training samples. For example, collecting regional datasets from major tea-producing countries (e.g., China, India, Japan, Sri Lanka) would be beneficial for validating the application of DL models across different regions. In addition, existing studies have addressed the overfitting issue in small tea leaf datasets by fine-tuning pre-trained models from other large-scale datasets using transfer learning. However, due to the differences in data distribution, not every sample from the source domain is suitable and usable for transfer learning. Future research can be conducted by dynamically adjusting sample weights and incorporating appropriate distance metrics, in combination with domain adaptation techniques, to automatically identify samples that contribute more significantly to transfer learning. It should be noted that traditional ML still plays a key role in tea quality detection, due to its suitability for small datasets. Future research can focus on integrating traditional ML with DL to yield a complementary result with the increasing scale of the tea dataset, as DL models have the capability to deal with complex big data.

Last but not least, achieving a comprehensive quality assessment in tea quality detection necessitates the integration of data from multiple sources. In this regard, strengthening interdisciplinary collaboration and knowledge transfer is important to enhance the application of DL to multimodal tea quality detection (Wang et al. 2022). For example, tea processors/researchers can collaborate with computer science experts for model development and validation. Sensing companies and tea industries can collaborate to develop multiple sensors for on-site, on-line, and in-process tea quality monitoring, thus, a comprehensive framework for detecting tea quality along the farm-to-market tea supply chain can be developed. Many multimodal network architectures developed in other fields can be adapted for tea quality detection. For example, in the field of meteorology, multimodal precipitation prediction models can be developed by integrating diverse information sources (such as ground observation station data and radar data) through cross-modal feature encoders. Compared with single-source prediction methods, this fusion-model leverages cross-attention mechanisms to combine spatiotemporal and sequential features, significantly improving real-time prediction and its accuracy (Cui et al. 2022b). In addition, in the manufacturing and industrial processes, digital twin (DT) technology based on deep multimodal information fusion (MIF) can effectively combine sensor data with component-level simulation features, supporting real-time fault detection in aero-engines (Huang et al. 2023b). In future studies, a unified model incorporating image encoders, sensor encoders, and cross-modal attention mechanisms can be developed to achieve deep alignment and fusion of multi-sensor features for tea quality detection, thereby enabling more precise and reliable quality assessment.

Conclusion

Advanced sensing technologies (such as NIRS, HSI, electronic noses and electronic tongues) play a crucial role in enhancing the automation and rapid quality detection in the tea industry. Coupled with DL methods for data processing and modeling, it can become a powerful tool to ensure quality monitoring throughout the tea production chain as well as to strengthen the competitiveness of the tea industry. DL methods have demonstrated excellent performance in analyzing high-dimensional tea data and have been widely applied in tasks such as tea quality classification and composition quantification. This review systematically examines a range of DL algorithms applied to tea quality assessment, including CNN, LSTM, ResNet, GAN, and many others. It further explores the application of DL models in three key stages during tea production: tea cultivation, tea processing, and final tea product evaluation. At the end of the review, practical challenges encountered in current research are discussed, and future studies to address those challenges are also proposed.

Compared with conventional chemical analysis methods and traditional ML techniques, DL enables faster and more intelligent quality detection while reducing human intervention for feature extraction. The hierarchical feature extraction capabilities in DL models are particularly suited for handling high-dimensional data. The non-linear mapping ability and strong generalization capacity of deep learning models contribute to their outstanding performance in various tea quality detection tasks. Furthermore, advanced DL techniques such as data augmentation, transfer learning, and knowledge distillation can enhance models’ adaptability to solve complex tea quality evaluation problems.

In recent years, DL technology has been widely applied in the scientific research for tea quality assessment, and promising results have been obtained. Increasingly, tea production enterprises are trying to integrate the DL with the automatic sorting systems to detect tea quality in real-time, thereby improving production efficiency and product quality. However, the successful real-world applications of DL in tea quality evaluation have not been reported, due to the fact that there are still challenging issues to be addressed. In practical applications, the detection equipment deployed by enterprises often has limited resource capacities, which restricts the deployment of complex DL models. Furthermore, DL models are commonly viewed as black-box systems due to the lack of model interpretability, which presents a significant barrier for their widespread adoption in the tea industry. Meanwhile, training DL models requires large and representative data, while data collection and annotation remain a challenge. Additionally, the data acquired by the sensing devices are impacted by various environmental factors such as lighting conditions, temperature and humidity variations, and noise interference. Moreover, when multiple sensors are used to capture data features, the challenge of multimodal data fusion must also be addressed. Thus, there is still a long way to apply DL for real-world applications in tea quality assessment.

Future research can further explore the application of model compression techniques on resource-constrained devices. In terms of knowledge distillation, advanced architectures such as transformer models and GNNs are promising for expanding the options of teacher models. Meanwhile, more samples are needed to meet the requirement of DL modeling. GANs can be employed for data augmentation to increase tea samples dataset for model training. The large public datasets for specific tasks of tea quality assessment can be constructed to develop DL models with high robustness and generalization ability, promoting the real-world applications of DL in the field. When fusion of multi-sensor data for tea quality assessment, cross-modal attention mechanisms and multimodal networks (which are well-established in medical and industrial domains) can be introduced and adopted for tea quality evaluation. With the continuous advancement of non-destructive sensing technologies and DL methodologies, DL-based sensing systems have immense potential for widespread deployment. However, tea variety classification, tea grade classification, and water and color related quality attributes analysis using rapid and non-destructive technologies (such as computer vision, near-infrared spectroscopy and hyperspectral maiming, etc.) might be the most promising directions for industry deployment. At present, the large number of samples can be obtained to form the big datasets for these directions in rapid and low-cost manners, which can fully reveal the advantage of DL methods. Research can focus on these directions to promote the real-world applications.

Acknowledgements

This work was supported by the Fundamental Research Funds for the Zhejiang Provincial [grant number: 2024QZJH41], and Zhejiang province agricultural machinery research, manufacturing and application integration project [grant number: 2023-YT-06].

Author contributions

T.W.: Visualization, Writing - original draft, Writing - review and editing; L.Z.: Writing - review and editing; Y.Z.: Writing - review and editing; H.Q.: Project administration, Writing - review and editing; Y.P.: Writing - review and editing; C.Z.: Conceptualization, Visualization, Supervision, Writing - original draft, Writing - review and editing; Y.L.: Conceptualization, Funding acquisition, Writing - review and editing.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

Ahmad, A; Saraswat, D; El Gamal, A. A survey on using deep learning techniques for plant disease diagnosis and recommendations for development of appropriate tools. Smart Agric Technol; 2023; 3, 100083. [DOI: https://dx.doi.org/10.1016/j.atech.2022.100083]

Aldughayfiq, B; Ashfaq, F; Jhanjhi, NZ; Humayun, M. Explainable AI for retinoblastoma diagnosis: interpreting deep learning models with LIME and SHAP. Diagnostics; 2023; 13, 11 1932. [DOI: https://dx.doi.org/10.3390/diagnostics13111932]

Alzubaidi, L; Zhang, J; Humaidi, AJ; Al-Dujaili, A; Duan, Y; Al-Shamma, O; Santamaria, J; Fadhel, MA; Al-Amidie, M; Farhan, L. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data; 2021; 8, 53. [DOI: https://dx.doi.org/10.1186/s40537-021-00444-8]

Alzubi, OA; Qiqieh, I; Alzubi, JA. Fusion of deep learning based cyberattack detection and classification model for intelligent systems. Clust Comput; 2022; 26, pp. 1363-1374. [DOI: https://dx.doi.org/10.1007/s10586-022-03686-0]

Alzubi, JA; Alzubi, OA; Qiqieh, I; Singh, A. A blended deep learning intrusion detection framework for consumable edge-centric IoMT industry. IEEE Trans Consum Electron; 2024; 70, pp. 2049-2057. [DOI: https://dx.doi.org/10.1109/tce.2024.3350231]

An, T; Yu, H; Yang, C; Liang, G; Chen, J; Hu, Z; Hu, B; Dong, C. Black tea withering moisture detection method based on convolution neural network confidence. J Food Process Eng; 2020; 43, e13428. [DOI: https://dx.doi.org/10.1111/jfpe.13428]

Besharati, MS; Pourdarbani, R; Sabzi, S; Sotoudeh, D; Ahmaditeshnizi, M; García-Mateos, G. Counterfeit detection of Iranian black tea using image processing and deep learning based on patched and unpatched images. Horticulturae; 2024; 10, 665. [DOI: https://dx.doi.org/10.3390/horticulturae10070665]

Bhargava, A; Bansal, A; Goyal, V; Bansal, P. A review on tea quality and safety using emerging parameters. J Food Meas Charact; 2022; 16, pp. 1291-1311. [DOI: https://dx.doi.org/10.1007/s11694-021-01232-x]

Bokuchava, MA; Skobeleva, NI. The biochemistry and technology of tea manufacture. Crit Rev Food Sci Nutr; 1980; 12, pp. 303-370. [DOI: https://dx.doi.org/10.1080/10408398009527280]

Calvini, R; Pigani, L. Toward the development of combined artificial sensing systems for food quality evaluation: A review on the application of data fusion of electronic noses, electronic tongues and electronic eyes. Sensors (Basel); 2022; 22, 577. [DOI: https://dx.doi.org/10.3390/s22020577]

Chang, J; Lu, A. RLCSA-net: a new deep learning method combined with an electronic nose to identify the quality of tea. J Food Meas Charact; 2024; 18, pp. 5222-5231. [DOI: https://dx.doi.org/10.1007/s11694-024-02539-1]

Chang, C; Li, Z; Li, H; Hou, Z; Zuo, E; Zhao, D; Lv, X; Zhong, F; Chen, C; Tian, F. A novel fast method for identifying the origin of Maojian using NIR spectroscopy with deep learning algorithms. Sci Rep; 2022; 12, 21418. [DOI: https://dx.doi.org/10.1038/s41598-022-25671-8]

Chen, YL; Duan, J; Jiang, YM; Shi, J; Peng, L; Xue, S; Kakuda, Y. Production, quality, and biological effects of oolong tea (Camellia sinensis). Food Reviews Int; 2010; 27, pp. 1-15. [DOI: https://dx.doi.org/10.1080/87559129.2010.518294]

Chen, Y; Zhao, X; Jia, X. Spectral–spatial classification of hyperspectral data based on deep belief network. IEEE J Sel Top Appl Earth Observ Remote Sens; 2015; 8, pp. 2381-2392. [DOI: https://dx.doi.org/10.1109/jstars.2015.2388577]

Chen, G; Zhang, X; Wu, Z; Su, J; Cai, G. An efficient tea quality classification algorithm based on near infrared spectroscopy and random forest. J Food Process Eng; 2020; 44, e13604. [DOI: https://dx.doi.org/10.1111/jfpe.13604]

Chen, C; Zhang, W; Shan, Z; Zhang, C; Dong, T; Feng, Z; Wang, C. Moisture contents and product quality prediction of Pu-erh tea in sun-drying process with image information and environmental parameters. Food Sci Nutr; 2022; 10, pp. 1021-1038. [DOI: https://dx.doi.org/10.1002/fsn3.2699]

Choudhary, K; DeCost, B; Chen, C; Jain, A; Tavazza, F; Cohn, R; Park, CW; Choudhary, A; Agrawal, A; Billinge, SJL; Holm, E; Ong, SP; Wolverton, C. Recent advances and applications of deep learning methods in materials science. Npj Comput Mater; 2022; 8, 59. [DOI: https://dx.doi.org/10.1038/s41524-022-00734-6]

Christin, S; Hervet, É; Lecomte, N; Ye, H. Applications for deep learning in ecology. Methods Ecol Evol; 2019; 10, pp. 1632-1644. [DOI: https://dx.doi.org/10.1111/2041-210x.13256]

Cui, Q; Yang, B; Liu, B; Li, Y; Ning, J. Tea category identification using wavelet signal reconstruction of hyperspectral imagery and machine learning. Agriculture (Basel); 2022; 12, 1085. [DOI: https://dx.doi.org/10.3390/agriculture12081085]

Cui, Y; Qiu, Y; Sun, L; Shu, X; Lu, Z. Quantitative short-term precipitation model using multimodal data fusion based on a cross-attention mechanism. Remote Sens; 2022; 14, 5839. [DOI: https://dx.doi.org/10.3390/rs14225839]

Diaz-Chito, K; Georgouli, K; Koidis, A; Martinez del Rincon, J. Incremental model learning for spectroscopy-based food analysis. Chemometr Intell Lab Syst; 2017; 167, pp. 123-131. [DOI: https://dx.doi.org/10.1016/j.chemolab.2017.06.002]

Ding, Y; Zeng, R; Jiang, H; Guan, X; Jiang, Q; Song, Z. Classification of tea quality grades based on hyperspectral imaging spatial information and optimization models. J Food Meas Charact; 2024; 18, pp. 9098-9112. [DOI: https://dx.doi.org/10.1007/s11694-024-02862-7]

Ding, Z; Yang, C; Hu, B; Guo, M; Li, J; Wang, M; Tian, Z; Chen, Z; Dong, C. Lightweight CNN combined with knowledge distillation for the accurate determination of black tea fermentation degree. Food Res Int; 2024; 194, 114929. [DOI: https://dx.doi.org/10.1016/j.foodres.2024.114929]

Dong, Y. Convolutional neural networks for sensitive identification of tea species using electrochemical sensors. J Food Meas Charact; 2024; 18, pp. 5878-5886. [DOI: https://dx.doi.org/10.1007/s11694-024-02615-6]

Dong, F; Bi, Y; Hao, J; Liu, S; Yi, W; Yu, W; Lv, Y; Cui, J; Li, H; Xian, J; Chen, S; Wang, S. A new comprehensive quantitative index for the assessment of essential amino acid quality in beef using Vis-NIR hyperspectral imaging combined with LSTM. Food Chem; 2024; 440, 138040. [DOI: https://dx.doi.org/10.1016/j.foodchem.2023.138040]

Gharibzahedi, SMT; Barba, FJ; Zhou, J; Wang, M; Altintas, Z. Electronic sensor technologies in monitoring quality of tea: a review. Biosensors (Basel); 2022; 12, 356. [DOI: https://dx.doi.org/10.3390/bios12050356]

Goodfellow, I; Pouget-Abadie, J; Mirza, M; Xu, B; Warde-Farley, D; Ozair, S; Courville, A; Bengio, Y. Generative adversarial networks. Commun ACM; 2020; 63, pp. 139-144. [DOI: https://dx.doi.org/10.1145/3422622]

Guo, M-H; Xu, T-X; Liu, J-J; Liu, Z-N; Jiang, P-T; Mu, T-J; Zhang, S-H; Martin, RR; Cheng, M-M; Hu, S-M. Attention mechanisms in computer vision: a survey. Comput Vis Media; 2022; 8, pp. 331-368. [DOI: https://dx.doi.org/10.1007/s41095-022-0271-y]

Guo, J; Liang, J; Xia, H; Ma, C; Lin, J; Qiao, X. An improved inception network to classify black tea appearance quality. J Food Eng; 2024; 369, 111931. [DOI: https://dx.doi.org/10.1016/j.jfoodeng.2023.111931]

Guo, J; Zhang, K; Adade, SYS; Lin, J; Lin, H; Chen, Q. Tea grading, blending, and matching based on computer vision and deep learning. J Sci Food Agric; 2024; 105, pp. 3239-3251. [DOI: https://dx.doi.org/10.1002/jsfa.14088]

Hajiboland, R. Environmental and nutritional requirements for tea cultivation. Folia Hortic; 2017; 29, pp. 199-220. [DOI: https://dx.doi.org/10.1515/fhort-2017-0019]

Han, Z; Ahmad, W; Rong, Y; Chen, X; Zhao, S; Yu, J; Zheng, P; Huang, C; Li, H. A gas sensors detection system for real-time monitoring of changes in volatile organic compounds during oolong tea processing. Foods; 2024; 13, 1721. [DOI: https://dx.doi.org/10.3390/foods13111721]

He, J; Zhang, C; Zhou, L; He, Y. Simultaneous determination of five micro-components in Chrysanthemum morifolium (Hangbaiju) using near-infrared hyperspectral imaging coupled with deep learning with wavelength selection. Infrared Phys Technol; 2021; 116, 103802. [DOI: https://dx.doi.org/10.1016/j.infrared.2021.103802]

Hong, Z; Zhang, C; Kong, D; Qi, Z; He, Y. Identification of storage years of black tea using near-infrared hyperspectral imaging with deep learning methods. Infrared Phys Technol; 2021; 114, 103666. [DOI: https://dx.doi.org/10.1016/j.infrared.2021.103666]

Hu, Y; Sheng, W; Adade, SY-SS; Wang, J; Li, H; Chen, Q. Comparison of machine learning and deep learning models for detecting quality components of vine tea using smartphone-based portable near-infrared device. Food Control; 2025; 174, 111244. [DOI: https://dx.doi.org/10.1016/j.foodcont.2025.111244]

Huang, Q; Wu, Z; Wang, M; Tao, Y; He, Y; Marinello, F. Prediction of tea varieties’ suitable for people relationship: based on the InteractE-SE + GCN model. Agriculture (Basel); 2023; 13, 1732. [DOI: https://dx.doi.org/10.3390/agriculture13091732]

Huang, Y; Tao, J; Sun, G; Wu, T; Yu, L; Zhao, X. A novel digital twin approach based on deep multimodal information fusion for aero-engine fault diagnosis. Energy; 2023; 270, 126894. [DOI: https://dx.doi.org/10.1016/j.energy.2023.126894]

Huang, C; Dong, S; Yang, Q; Zhang, J; Yang, M; Zhang, S; Dai, Q; Ning, J; Li, L. Aroma visualization: a cutting-edge sensor for evaluating the roasting quality of large-leaf yellow tea. LWT; 2024; 208, 116684. [DOI: https://dx.doi.org/10.1016/j.lwt.2024.116684]

Huang, M; Tang, Y; Tan, Z; Ren, J; He, Y; Huang, H. Detection of black tea fermentation quality based on optimized deep neural network and hyperspectral imaging. Infrared Phys Technol; 2024; 143, 105625. [DOI: https://dx.doi.org/10.1016/j.infrared.2024.105625]

Kang, S; Zhang, Q; Li, Z; Yin, C; Feng, N; Shi, Y. Determination of the quality of tea from different picking periods: an adaptive pooling attention mechanism coupled with an electronic nose. Postharvest Biol Technol; 2023; 197, 112214. [DOI: https://dx.doi.org/10.1016/j.postharvbio.2022.112214]

Karwowska, K; Skotnicka, M; Śmiechowska, M. Tea production and its forecasts, and the possibility of tea cultivation in the context of environmental requirements in China. Zeszyty Naukowe SGGW W Warszawie - Problemy Rolnictwa Światowego; 2019; 19, pp. 180-191. [DOI: https://dx.doi.org/10.22630/prs.2019.19.1.17]

Kaushal, S; Nayi, P; Rahadian, D; Chen, H-H. Applications of electronic nose coupled with statistical and intelligent pattern recognition techniques for monitoring tea quality: a review. Agriculture (Basel); 2022; 12, 1359. [DOI: https://dx.doi.org/10.3390/agriculture12091359]

Li, S; Lo, CY; Pan, MH; Lai, CS; Ho, CT. Black tea: chemical analysis and stability. Food Funct; 2013; 4, pp. 10-18. [DOI: https://dx.doi.org/10.1039/c2fo30093a]

Li, M; Pan, T; Bai, Y; Chen, Q. Development of a calibration model for near infrared spectroscopy using a convolutional neural network. J Near Infrared Spectrosc; 2022; 30, pp. 89-96. [DOI: https://dx.doi.org/10.1177/09670335211057234]

Li, X; Xiong, H; Li, X; Wu, X; Zhang, X; Liu, J; Bian, J; Dou, D. Interpretable deep learning: interpretation, interpretability, trustworthiness, and beyond. Knowl Inf Syst; 2022; 64, pp. 3197-3234. [DOI: https://dx.doi.org/10.1007/s10115-022-01756-8]

Li, H; Luo, X; Haruna, SA; Zareef, M; Chen, Q; Ding, Z; Yan, Y. Au-Ag OHCs-based SERS sensor coupled with deep learning CNN algorithm to quantify thiram and pymetrozine in tea. Food Chem; 2023; 428, 136798. [DOI: https://dx.doi.org/10.1016/j.foodchem.2023.136798]

Li, L; Chen, Y; Dong, S; Shen, J; Cao, S; Cui, Q; Song, Y; Ning, J. Rapid and comprehensive grade evaluation of Keemun black tea using efficient multidimensional data fusion. Food Chem X; 2023; [DOI: https://dx.doi.org/10.1016/j.fochx.2023.100924]

Li, N; Yao, Z; Ning, J; Sun, L; Lin, Q; Zhu, X; Li, C; Zheng, X; Jin, J. Comparison of different drying technologies for green tea: changes in color, non-volatile and volatile compounds. Food Chem X; 2024; [DOI: https://dx.doi.org/10.1016/j.fochx.2024.101935]

Li, Z; Wang, D; Zhu, T; Tao, Y; Ni, C. Review of deep learning-based methods for non-destructive evaluation of agricultural products. Biosyst Eng; 2024; 245, pp. 56-83. [DOI: https://dx.doi.org/10.1016/j.biosystemseng.2024.07.002]

Liang, N; Sun, S; Zhang, C; He, Y; Qiu, Z. Advances in infrared spectroscopy combined with artificial neural network for the authentication and traceability of food. Crit Rev Food Sci Nutr; 2020; 62, pp. 2963-2984. [DOI: https://dx.doi.org/10.1080/10408398.2020.1862045]

Liang, J; Guo, J; Xia, H; Ma, C; Qiao, X. A black tea quality testing method for scale production using CV and NIRS with TCN for spectral feature extraction. Food Chem; 2025; 464, 141567. [DOI: https://dx.doi.org/10.1016/j.foodchem.2024.141567]

Lin, Y; Ma, J; Wang, Q; Sun, DW. Applications of machine learning techniques for enhancing nondestructive food quality and safety detection. Crit Rev Food Sci Nutr; 2023; 63, pp. 1649-1669. [DOI: https://dx.doi.org/10.1080/10408398.2022.2131725]

Liu, H; Yu, D; Gu, Y. Classification and evaluation of quality grades of organic green teas using an electronic nose based on machine learning algorithms. IEEE Access; 2019; 7, pp. 172965-172973. [DOI: https://dx.doi.org/10.1109/access.2019.2957112]

Liu, C; Lu, W; Gao, B; Kimura, H; Li, Y; Wang, J. Rapid identification of chrysanthemum teas by computer vision and deep learning. Food Sci Nutr; 2020; 8, pp. 1968-1977. [DOI: https://dx.doi.org/10.1002/fsn3.1484]

Liu, Y; Pu, H; Sun, D-W. Efficient extraction of deep image features using convolutional neural network (CNN) for applications in detecting and analysing complex food matrices. Trends Food Sci Technol; 2021; 113, pp. 193-204. [DOI: https://dx.doi.org/10.1016/j.tifs.2021.04.042]

Liu, H; Guan, H; He, F; Song, Y; Li, F; Sun-Waterhouse, D; Li, D. Therapeutic actions of tea phenolic compounds against oxidative stress and inflammation as central mediators in the development and progression of health problems: a review focusing on MicroRNA regulation. Crit Rev Food Sci Nutr; 2024; 64, pp. 8414-8444. [DOI: https://dx.doi.org/10.1080/10408398.2023.2202762]

Lu, T; Yu, F; Xue, C; Han, B. Identification, classification, and quantification of three physical mechanisms in oil-in-water emulsions using AlexNet with transfer learning. J Food Eng; 2021; 288, 110220. [DOI: https://dx.doi.org/10.1016/j.jfoodeng.2020.110220]

Luo, N; Li, Y; Yang, B; Liu, B; Dai, Q. Prediction model for tea polyphenol content with deep features extracted using 1D and 2D convolutional neural network. Agriculture (Basel); 2022; 12, 1299. [DOI: https://dx.doi.org/10.3390/agriculture12091299]

Luo, X; Gouda, M; Perumal, AB; Huang, Z; Lin, L; Tang, Y; Sanaeifar, A; He, Y; Li, X; Dong, C. Using surface-enhanced Raman spectroscopy combined with chemometrics for black tea quality assessment during its fermentation process. Sens Actuat B Chem; 2022; [DOI: https://dx.doi.org/10.1016/j.snb.2022.132680]

Lv, H; Feng, X; Song, H; Ma, S; Hao, Z; Hu, H; Yang, Y; Pan, Y; Zhou, S; Fan, F; Gong, S; Chu, Q; Chen, P. Tea storage: a not thoroughly recognized and precisely designed process. Trends Food Sci Technol; 2023; 140, 104172. [DOI: https://dx.doi.org/10.1016/j.tifs.2023.104172]

Mahdavifar, S; Ghorbani, AA. Application of deep learning to cybersecurity: a survey. Neurocomputing; 2019; 347, pp. 149-176. [DOI: https://dx.doi.org/10.1016/j.neucom.2019.02.056]

Mamoshina, P; Vieira, A; Putin, E; Zhavoronkov, A. Applications of deep learning in biomedicine. Mol Pharm; 2016; 13, pp. 1445-1454. [DOI: https://dx.doi.org/10.1021/acs.molpharmaceut.5b00982]

Mao, G; Liao, G; Zhu, H; Sun, B. Multibranch attention mechanism based on channel and spatial attention fusion. Mathematics; 2022; 10, 4150. [DOI: https://dx.doi.org/10.3390/math10214150]

Moreira, J; Aryal, J; Guidry, L; Adhikari, A; Chen, Y; Sriwattana, S; Prinyawiwatkul, W. Tea quality: an overview of the analytical methods and sensory analyses used in the most recent studies. Foods; 2024; 13, 3580. [DOI: https://dx.doi.org/10.3390/foods13223580]

Mukhiddinov, M; Muminov, A; Cho, J. Improved classification approach for fruits and vegetables freshness based on deep learning. Sensors (Basel); 2022; 22, 8192. [DOI: https://dx.doi.org/10.3390/s22218192]

Ren, G; Wu, R; Yin, L; Zhang, Z; Ning, J. Description of tea quality using deep learning and multi-sensor feature fusion. J Food Compos Anal; 2024; 126, 105924. [DOI: https://dx.doi.org/10.1016/j.jfca.2023.105924]

ŞAhiN, E; Arslan, NN; Özdemir, D. Unlocking the black box: an in-depth review on interpretability, explainability, and reliability in deep learning. Neural Comput Appl; 2024; 37, pp. 859-965. [DOI: https://dx.doi.org/10.1007/s00521-024-10437-2]

Selvaraju, RR; Cogswell, M; Das, A; Vedantam, R; Parikh, D; Batra, D. Grad-cam: visual explanations from deep networks via gradient-based localization. Int J Comput Vision; 2019; 128, pp. 336-359. [DOI: https://dx.doi.org/10.1007/s11263-019-01228-7]

Shafiq, M; Gu, Z. Deep residual learning for image recognition: a survey. Appl Sci; 2022; 12, 8972. [DOI: https://dx.doi.org/10.3390/app12188972]

Sidhu, D; Vasundhara, M; Dey, P. The intestinal-level metabolic benefits of green tea catechins: mechanistic insights from pre-clinical and clinical studies. Phytomedicine; 2024; 123, 155207. [DOI: https://dx.doi.org/10.1016/j.phymed.2023.155207]

Song, Y; Wang, X; Xie, H; Li, L; Ning, J; Zhang, Z. Quality evaluation of Keemun black tea by fusing data obtained from near-infrared reflectance spectroscopy and computer vision sensors. Spectrochim Acta Mol Biomol Spectrosc; 2021; 252, 119522. [DOI: https://dx.doi.org/10.1016/j.saa.2021.119522]

Sonobe, R; Hirono, Y. Carotenoid content estimation in tea leaves using noisy reflectance data. Remote Sens; 2023; 15, 4303. [DOI: https://dx.doi.org/10.3390/rs15174303]

Sonobe, R; Hirono, Y; Oi, A. Quantifying chlorophyll-a and b content in tea leaves using hyperspectral reflectance and deep learning. Remote Sens Lett; 2020; 11, pp. 933-942. [DOI: https://dx.doi.org/10.1080/2150704x.2020.1795294]

Sun, J; Yang, F; Cheng, J; Wang, S; Fu, L. Nondestructive identification of soybean protein in minced chicken meat based on hyperspectral imaging and VGG16-SVM. J Food Compos Anal; 2024; 125, 105713. [DOI: https://dx.doi.org/10.1016/j.jfca.2023.105713]

Tan, X; He, J; Li, H; Jiang, J; He, Z; Guo, J; Qiu, B. Utilizing excitation-emission matrix fluorescence spectroscopy and convolutional neural networks for dark tea brand and aging period identification. J Food Compos Anal; 2024; 133, 106458. [DOI: https://dx.doi.org/10.1016/j.jfca.2024.106458]

Tang, T; Luo, Q; Yang, L; Gao, C; Ling, C; Wu, W. Research review on quality detection of fresh tea leaves based on spectral technology. Foods; 2023; 13, 25. [DOI: https://dx.doi.org/10.3390/foods13010025]

Tibaduiza, D; Anaya, M; Gomez, J; Sarmiento, J; Perez, M; Lara, C; Ruiz, J; Osorio, N; Rodriguez, K; Hernandez, I; Sanchez, C. Electronic tongues and noses: a general overview. Biosensors (Basel); 2024; 14, 190. [DOI: https://dx.doi.org/10.3390/bios14040190]

Tudu, B; Jana, A; Metla, A; Ghosh, D; Bhattacharyya, N; Bandyopadhyay, R. Electronic nose for black tea quality evaluation by an incremental RBF network. Sens Actuators B Chem; 2009; 138, pp. 90-95. [DOI: https://dx.doi.org/10.1016/j.snb.2009.02.025]

Turgut, SS; Küçüköner, E; Karacabey, E. Teapot: a chemometric tool for tea blend recipe estimation. Appl Food Res; 2021; 1, 1 100006. [DOI: https://dx.doi.org/10.1016/j.afres.2021.100006]

Wang, D; Cao, W; Zhang, F; Li, Z; Xu, S; Wu, X. A review of deep learning in multiscale agricultural sensing. Remote Sens; 2022; 14, 3 559. [DOI: https://dx.doi.org/10.3390/rs14030559]

Wang, H; Gu, J; Wang, M. A review on the application of computer vision and machine learning in the tea industry. Front Sustain Food Syst; 2023; 7, 1172543. [DOI: https://dx.doi.org/10.3389/fsufs.2023.1172543]

Wang, X; Liu, Y; Gong, N; Wang, J; Han, Z; Wu, Z. Design of a tea strip-lining degree discrimination system based on appearance characteristics and near-infrared spectroscopy. J Food Compos Anal; 2024; 136, 106779. [DOI: https://dx.doi.org/10.1016/j.jfca.2024.106779]

Wang, Y; Gu, H-W; Yin, X-L; Geng, T; Long, W; Fu, H; She, Y. Deep leaning in food safety and authenticity detection: an integrative review and future prospects. Trends Food Sci Technol; 2024; 146, 104396. [DOI: https://dx.doi.org/10.1016/j.tifs.2024.104396]

Wei, K; Chen, B; Li, Z; Chen, D; Liu, G; Lin, H; Zhang, B. Classification of tea leaves based on fluorescence imaging and convolutional neural networks. Sensors (Basel); 2022; 22, 7764. [DOI: https://dx.doi.org/10.3390/s22207764]

Wen, J; He, J. Agricultural development driven by the digital economy: improved efficientnet vegetable quality grading. Front Sustain Food Syst; 2024; 8, 1310042. [DOI: https://dx.doi.org/10.3389/fsufs.2024.1310042]

Xu, S; Sun, X; Lu, H; Zhang, Q. Detection of type, blended ratio, and mixed ratio of pu’er tea by using electronic nose and visible/near infrared spectrometer. Sensors (Basel); 2019; 19, 2359. [DOI: https://dx.doi.org/10.3390/s19102359]

Xu, M; Wang, J; Zhu, L. Tea quality evaluation by applying e-nose combined with chemometrics methods. J Food Sci Technol; 2021; 58, pp. 1549-1561. [DOI: https://dx.doi.org/10.1007/s13197-020-04667-0]

Xu, W; Zhao, L; Li, J; Shang, S; Ding, X; Wang, T. Detection and classification of tea buds based on deep learning. Comput Electron Agric; 2022; 192, 106547. [DOI: https://dx.doi.org/10.1016/j.compag.2021.106547]

Yang, J; Wang, J; Lu, G; Fei, S; Yan, T; Zhang, C; Lu, X; Yu, Z; Li, W; Tang, X. Teanet: deep learning on near-infrared spectroscopy (NIR) data for the assurance of tea quality. Comput Electron Agric; 2021; 190, 106431. [DOI: https://dx.doi.org/10.1016/j.compag.2021.106431]

Yang, Y-Z; Li, T; Teng, R-M; Han, M-H; Zhuang, J. Low temperature effects on carotenoids biosynthesis in the leaves of green and albino tea plant ( (L.) O. Kuntze). Sci Hortic; 2021; 285, 110164. [DOI: https://dx.doi.org/10.1016/j.scienta.2021.110164]

Yang, Z; Gao, J; Wang, S; Wang, Z; Li, C; Lan, Y; Sun, X; Li, S. Synergetic application of E-tongue and E-eye based on deep learning to discrimination of Pu-erh tea storage time. Comput Electron Agric; 2021; 187, 106297. [DOI: https://dx.doi.org/10.1016/j.compag.2021.106297]

Yang, Z; Miao, N; Zhang, X; Li, Q; Wang, Z; Li, C; Sun, X; Lan, Y. Employment of an electronic tongue combined with deep learning and transfer learning for discriminating the storage time of Pu-erh tea. Food Control; 2021; 121, 107608. [DOI: https://dx.doi.org/10.1016/j.foodcont.2020.107608]

Yang, Z; Ma, W; Lu, J; Tian, Z; Peng, K. The application status and trends of machine vision in tea production. Appl Sci; 2023; [DOI: https://dx.doi.org/10.3390/app131910744]

Yang, N; Simon, J; Fang, W; Ayed, C; Zhang, WE; Axell, M; Viltoriano, R; Fisk, I. Development of analytical aroma wheels for oolong tea infusions (Shuixian and Rougui) and prediction of dynamic aroma release and colour changes during Chinese tea ceremony with machine learning. Food Chem; 2025; [DOI: https://dx.doi.org/10.1016/j.foodchem.2024.141537]

Ye, Z; Liu, Y; Li, Q. Recent progress in smart electronic nose technologies enabled with machine learning methods. Sensors (Basel); 2021; [DOI: https://dx.doi.org/10.3390/s21227620]

You, J; Li, D; Wang, Z; Chen, Q; Ouyang, Q. Prediction and visualization of moisture content in Tencha drying processes by computer vision and deep learning. J Sci Food Agric; 2024; 104, pp. 5486-5494. [DOI: https://dx.doi.org/10.1002/jsfa.13381]

Yu, D; Gu, Y. A machine learning method for the fine-grained classification of green tea with geographical indication using a MOS-based electronic nose. Foods; 2021; 10, 795. [DOI: https://dx.doi.org/10.3390/foods10040795]

Yu, XL; Sun, DW; He, Y. Emerging techniques for determining the quality and safety of tea products: a review. Compr Rev Food Sci Food Saf; 2020; 19, pp. 2613-2638. [DOI: https://dx.doi.org/10.1111/1541-4337.12611]

Yu, Y; Huang, J; Wang, L; Liang, S. A 1D-inception-ResNet based global detection model for thin-skinned multifruit spectral quantitative analysis. Food Control; 2025; 167, 110823. [DOI: https://dx.doi.org/10.1016/j.foodcont.2024.110823]

Zhang, X; Yang, J; Lin, T; Ying, Y. Food and agro-product quality evaluation based on spectroscopy and deep learning: a review. Trends Food Sci Technol; 2021; 112, pp. 431-441. [DOI: https://dx.doi.org/10.1016/j.tifs.2021.04.008]

Zhang, C; Wang, J; Lu, G; Fei, S; Zheng, T; Huang, B. Automated tea quality identification based on deep convolutional neural networks and transfer learning. J Food Process Eng; 2023; 46, e14303. [DOI: https://dx.doi.org/10.1111/jfpe.14303]

Zhang, C; Wang, J; Yan, T; Lu, X; Lu, G; Tang, X; Huang, B. An instance-based deep transfer learning method for quality identification of Longjing tea from multiple geographical origins. Complex Intell Syst; 2023; 9, pp. 3409-3428. [DOI: https://dx.doi.org/10.1007/s40747-023-01024-4]

Zhang, M; Zhang, T; Wang, Y; Duan, X; Pu, L; Zhang, Y; Li, Q; Liu, Y. Accurate prediction of tea catechin content with near-infrared spectroscopy by deep learning based on channel and spatial attention mechanisms. Chemosensors; 2024; 12, 184. [DOI: https://dx.doi.org/10.3390/chemosensors12090184]

Zhang, W; Sanaeifar, A; Ji, X; Luo, X; Guo, H; He, Q; Luo, Y; Huang, F; Yan, P; Li, X; He, Y. Data-driven optimization of nitrogen fertilization and quality sensing across tea bud varieties using near-infrared spectroscopy and deep learning. Comput Electron Agric; 2024; 222, 109071. [DOI: https://dx.doi.org/10.1016/j.compag.2024.109071]

Zhao, L; Wang, L; Jia, Y; Cui, Y. A lightweight deep neural network with higher accuracy. PLoS One; 2022; 17, e0271225. [DOI: https://dx.doi.org/10.1371/journal.pone.0271225]

Zhao, H; Xue, D; Zhang, L. Electrochemical fingerprints identification of tea based on one-dimensional convolutional neural network. J Food Meas Charact; 2023; 17, pp. 2607-2613. [DOI: https://dx.doi.org/10.1007/s11694-023-01812-z]

Zhao, X; He, Y; Zhang, H; Ding, Z; Zhou, C; Zhang, K. A quality grade classification method for fresh tea leaves based on an improved YOLOv8x-SPPCSPC-CBAM model. Sci Rep; 2024; 14, 4166. [DOI: https://dx.doi.org/10.1038/s41598-024-54389-y]

Zheng, W; Xia, X; Shi, Y; Men, H. CMTP-TCE: a data augmentation method of electronic tongue combined with Dot-product attention mechanism and residual network for food quality classification. IEEE Sens J; 2023; 23, pp. 21652-21660. [DOI: https://dx.doi.org/10.1109/jsen.2023.3304062]

Zheng, P; Solomon Adade, SY; Rong, Y; Zhao, S; Han, Z; Gong, Y; Chen, X; Yu, J; Huang, C; Lin, H. Online system for monitoring the degree of fermentation of oolong tea using integrated visible-near-infrared spectroscopy and image-processing technologies. Foods; 2024; 13, 11 1708. [DOI: https://dx.doi.org/10.3390/foods13111708]

Zhi, R; Zhao, L; Zhang, D. A framework for the multi-level fusion of electronic nose and electronic tongue for tea quality assessment. Sensors (Basel); 2017; 17, 5 1007. [DOI: https://dx.doi.org/10.3390/s17051007]

Zhong, Yh; Zhang, S; He, R; Zhang, J; Zhou, Z; Cheng, X; Huang, G; Zhang, J. A convolutional neural network based auto features extraction method for tea classification with electronic tongue. Appl Sci; 2019; [DOI: https://dx.doi.org/10.3390/app9122518]

Zhu, J; Sharma, AS; Xu, J; Xu, Y; Jiao, T; Ouyang, Q; Li, H; Chen, Q. Rapid on-site identification of pesticide residues in tea by one-dimensional convolutional neural network coupled with surface-enhanced Raman scattering. Spectrochim Acta Mol Biomol Spectrosc; 2021; 246, 118994. [DOI: https://dx.doi.org/10.1016/j.saa.2020.118994]

Zhu, Y; Chen, S; Yin, H; Han, X; Xu, M; Wang, W; Zhang, Y; Feng, X; Liu, Y. Classification of oolong tea varieties based on computer vision and convolutional neural networks. J Sci Food Agric; 2024; 104, pp. 1630-1637. [DOI: https://dx.doi.org/10.1002/jsfa.13049]

Zhu, F; Wang, J; Zhang, Y; Shi, J; He, M; Zhao, Z. An improved 3D-SwinT-CNN network to evaluate the fermentation degree of black tea. Food Control; 2025; 167, 110756. [DOI: https://dx.doi.org/10.1016/j.foodcont.2024.110756]

Zhu, F; Zhang, Y; Wang, J; Luo, X; Liu, D; Jin, K; Peng, J. An improved deep convolutional generative adversarial network for quantification of catechins in fermented black tea. Spectrochim Acta A Mol Biomol Spectrosc; 2025; [DOI: https://dx.doi.org/10.1016/j.saa.2024.125357]