Computer Vision from Tea Cultivation to Quality Evaluation(MDPI Foods 2026-08-17)

Logical Operator Operator

Search Text

Search Type

add_circle_outline

remove_circle_outline

Article Menu

Font Type:

Arial Georgia Verdana

Font Size:

Aa Aa Aa

Line Spacing:

  

Column Width:

  

Background:

Open Access Review

by
Zunren Chen

1,†,
Jinfeng Wang

2,†,
Yilan Sun

2,
Jie Pang

2,
Wei Xin

3,4,
Qinhua Zhang

5,* and
Junling Zhou

6,*

1

Library of Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China

2

School of Food Science, Fujian Agriculture and Forestry University, Fuzhou 350002, China

3

College of Ecology and Resources Engineering, Wuyi University, Wuyishan 354300, China

4

Fujian Key Laboratory of Big Data Application and Intellectualization for Tea Industry, Wuyi University, Nanping 529000, China

5

Jinshan College, Fujian Agriculture and Forestry University, Fuzhou 350002, China

6

Haixia Institute of Science and Technology, Fujian Agriculture and Forestry University, Fuzhou 350002, China

Authors to whom correspondence should be addressed.

†

These authors contributed equally to this work.

Submission received: 23 June 2026 / Revised: 7 August 2026 / Accepted: 11 August 2026 / Published: 17 August 2026

Abstract

Existing reviews on AI in tea production are either agriculture-generic or limited to isolated tasks. This review thoroughly compares vision technologies (RGB, hyperspectral, near-infrared, thermal, Light Detection and Ranging (LiDAR), Unmanned Aerial Vehicle (UAV)) and establishes a task-oriented algorithm selection framework for the tea industry. For small-sample or near-linear problems, traditional machine learning (ML) (support vector machine (SVM); partial least squares regression (PLSR)) remains effective. For unstructured field tasks, deep learning achieves superior performance: pest detection accuracy exceeds 97%, tea bud detection reaches 96.8% with RGB images, and hyperspectral imaging predicts nitrogen content with R 2> 0.90 and tea polyphenols with R 2 up to 0.925. Algorithm choice further differentiates by task granularity: lightweight convolutional neural networks (CNNs) balance speed and accuracy for edge deployment at 16 fps; You Only Look Once (YOLO) series detectors enable real-time localization on mobile platforms at 93.1% accuracy, 24 ms per target. No single algorithm dominates all tea tasks; selection is a trade-off among accuracy, speed, data availability, and computational constraints. These findings outline a structured analysis of the challenges and pathways for transitioning computer vision (CV) from laboratory research toward field-deployable tools.

1. Introduction

Tea, an ancient beverage originating in the East, is now the second most consumed drink globally after water [1]. It constitutes a major global industry at the intersection of culture, health, and economics. According to the Food and Agriculture Organization (FAO), global tea production exceeded 6.5 million tons in 2022, with a trade value of approximately $17 billion USD (FAO, 2022), serving as an economic pillar in many rural and developing regions [2]. On one hand, consumer demand is shifting towards higher, more standardized levels of quality, safety, and traceability [3]. On the other hand, structural issues within the industry are intensifying. Major tea-producing regions worldwide are experiencing an aging rural population and a migration of young labor to urban areas, making it increasingly difficult to secure sufficient labor for harvest and other tasks, while costs continue to rise [4]. Furthermore, traditional production is heavily reliant on empirical judgment from fertilization and pest control to harvesting standards and processing, hindering standardization, resulting in inconsistent product quality and difficulties in achieving scalable, high-grade output [5]. This reliance on subjective, experience-based methods represents a fundamental bottleneck to modernization and sustainable development [6].

The advent of new-generation AI, particularly deep learning, has driven transformative breakthroughs across various fields, including agriculture [7]. Within this, AI-powered CV stands out as a key to enabling technology. Capable of automated recognition, measurement, and analysis from image and video data, CV can match or surpass human visual perception. This offers a powerful toolkit to address the traditional constraints of the tea industry [8]. Applications range from macro-scale monitoring of plantation growth and nutrient status using drone-mounted multispectral cameras [9,10] to real-time pest and disease identification via CNNs and precise detection of harvestable tea buds in complex natural environments using object detection algorithms [11]. Moreover, technologies like hyperspectral imaging (HSI) enable non-destructive, real-time monitoring of biochemical constituents during critical processing stages such as withering and fermentation, facilitating digital and precise process control [12].

To provide a comprehensive yet focused overview, this review adopts a pragmatic definition of CV that extends beyond conventional image-based techniques. Specifically, we include not only RGB, HSI, and thermal imaging (TIC) but also near-infrared spectroscopy (NIRS), LiDAR, and UAV [13]. These technologies can be categorized into imaging modalities (RGB, HSI, TIC), spectroscopic techniques (NIRS), 3D spatial sensors (LiDAR), and aerial platforms (UAV). NIRS is considered because, in tea processing, it is often integrated with imaging systems for real-time quality prediction [14]; LiDAR is included for its ability to provide three-dimensional spatial information that complements 2D images for bud localization and canopy analysis [15], and UAV is treated as a platform that carries vision sensors, enabling large-scale, multi-temporal monitoring [11]. The selection of these technologies was guided by three practical criteria-scalability (RGB cameras and UAVs can cover large areas), robustness under field conditions (LiDAR’s insensitivity to lighting variation), and feasibility for real-world deployment (NIRS portability for inline processing). This broad yet practical scope reflects the multimodal sensing strategies commonly employed in current tea industry research. Given this focus on practical application, we prioritize studies that report quantitative performance metrics, whether conducted in the laboratory, pilot line, or field.

The literature covered in this review was primarily identified through searches in Web of Science, Scopus, and Google Scholar using keywords including “computer vision,” “deep learning,” “tea industry,” “precision agriculture,” “hyperspectral imaging,” “NIRS,” “LiDAR,” and “UAV-based remote sensing.” The search covered publications from 2018 to 2025, with emphasis on studies reporting quantitative performance metrics. References were selected based on their relevance to the tea production chain and their contribution to the comparative analysis of vision technologies and algorithms.

Recent reviews have summarized AI applications in tea management, harvesting, and processing, but a quantitative, full-chain comparison of vision technologies with a dedicated algorithm selection framework has been less explored. Unlike existing reviews that focus on isolated tasks or general agriculture, this review provides a quantitative comparison of vision technologies alongside a task-oriented algorithm selection framework across the entire tea production chain. It offers critical comparative analyses of technological approaches, identifies practical constraints, and examines translational challenges from lab to field, aiming to bridge the gap between research and industrial adoption.

2. Key Enabling Technologies and Data Acquisition Systems in the Tea Industry

2.1. AI-Powered CV Technology: Core Concepts for Tea Applications

AI-powered CV integrates CV and deep learning to automatically extract information from images. Its pipeline typically includes data acquisition using RGB, HSI, TIC, or LiDAR sensors, data processing (preprocessing, segmentation), feature learning (via CNNs, autoencoders), and task execution (classification, detection, segmentation). In the tea industry, deep learning, especially CNNs, has become dominant for handling complex, unstructured scenes such as pest detection and bud recognition, while traditional ML (SVM, PLSR) remains useful for small-sample or near-linear problems. The operational framework of a CV system typically mirrors human visual processing, encompassing four key stages: data acquisition, data processing, data learning, and task execution (see Figure 1).

Figure 1. Typical workflow of CV systems for tea applications. The red circles and lines indicate the propagation paths selected by the proposed strategy for the convolutional neural network.

2.2. Multidimensional Visual Perception Technology

The efficacy of CV systems in the tea industry is fundamentally dependent on the underlying perception technologies for information acquisition. The detailed comparison of these technologies, including their core benefits, key applications, performance examples, validation settings, cost, maturity levels, and spatial resolution, is presented in Table 1. Unlike human vision, which is confined to the visible spectrum, modern perception technologies construct a multidimensional visual perception system that vastly exceeds human sensory capabilities by expanding the spectral range and integrating spatial, temporal, and chemical dimensions. This system provides holistic insights from macro to micro scales and from external morphology to internal chemistry, enabling intelligent monitoring and decision-making across the entire tea production chain, from cultivation to processing.

Table 1. Comparison of multidimensional vision technologies in the tea industry.

In terms of spatial resolution (SR) and acquisition speed, the six technology categories differ substantially. RGB cameras typically achieve 0.1–1 mm/pixel at 30–60 fps. HSI offers similar spatial resolution but much slower acquisition (0.1–10 fps) due to scanning. NIRS provides no spatial image (point/line scan) but is fast (10–100 spectra/s). TIC gives lower spatial resolution (1–5 mm/pixel) at 1–30 fps. LiDAR generates 3D point clouds with cm-to-dm resolution at 10–100 k points/s. UAV-based remote sensing covers large areas at 0.5–5 cm/pixel per flight (10–30 min per 50 ha). A rough cost ranking (low to high) is: RGB < NIRS < TIC ≈ UAV < HSI ≈ LiDAR. At the lower end, RGB cameras and consumer-grade NIRS devices are relatively affordable and widely available. TIC and UAV systems fall into the medium range, though costs vary considerably depending on sensor resolution and payload capacity. HSI and LiDAR remain the most expensive, due to their high-precision optics, complex calibration procedures, and substantial data storage and computational demands. To facilitate the comparison of technological maturity, each technology is assessed using the Technology Readiness Level (TRL), where levels 1–3 indicate basic research and proof-of-concept, 4–6 indicate validation in laboratory or relevant environments, and 7–9 indicate system prototype demonstration and full operational deployment.

2.2.1. RGB Imaging

RGB imaging, which simulates human trichromatic vision, is the most widely adopted visual perception technology due to its low cost and simple deployment. In the tea industry, RGB cameras are primarily used to capture morphological, color, and textural features for tasks such as identifying bud maturity, preliminary screening of pronounced foliar disease lesions, and grading the appearance and color of dry tea. However, RGB imaging is inherently limited to surface reflectance in the visible spectrum. It cannot penetrate surfaces, detect chemical constituents, or effectively identify early-stage, latent physiological changes. Its performance is also highly susceptible to variations in ambient lighting conditions [13]. To mitigate these lighting-induced issues, common preprocessing strategies include histogram equalization, white balance correction, and conversion to HSV or Lab color spaces to separate luminance from chrominance [25]. Adaptive thresholding and illumination normalization can further improve robustness in non-uniform lighting [26].

2.2.2. HSI

HSI, which integrates spectroscopy with imaging, is a core enabling technology for intelligent perception in the tea industry [27]. Unlike an RGB pixel containing only three broad bands, a hyperspectral pixel captures reflectance across hundreds of contiguous narrow bands (typically with spectral resolution < 10 nm), forming a rich data cube with combined spatial and spectral dimensions [28].

This capability stems from the absorption of specific near-infrared photons by vibrational overtones of chemical bonds involving hydrogen (O-H, N-H, C-H). Consequently, subtle changes in key internal chemical components produce distinctive responses in the spectral signature [18]. HSI thus enables non-destructive, spatially resolved chemical mapping—for instance, visualizing moisture distribution during withering, quantifying the dynamics of polyphenol transformation during fermentation, or identifying pre-visual plant stress [19]. In contrast, multispectral imaging (MSI) acquires data at only a few discrete bands, trading spectral resolution for lower cost and faster acquisition. While HSI typically provides higher prediction accuracy for internal chemical constituents (R 2> 0.91 for nitrogen and polyphenols [19]), MSI is more suitable for large-scale, real-time field monitoring due to its reduced data volume [29].

The primary challenges associated with HSI are data redundancy and high computational cost [30]. Effective implementation often requires dimensionality reduction via wavelength selection algorithms or leveraging deep learning models like CNNs to automatically extract critical spectral-spatial features [31].

2.2.3. NIRS

NIRS operates on a similar spectroscopic principle to HSI but typically acquires spectral data in a point or line-scan mode, emphasizing rapid, portable, and online analysis [6]. Its high sensitivity to hydrogen-containing groups makes it indispensable for real-time quality control during tea processing [20]. The miniaturization of NIRS with fiber-optic probes facilitates direct integration into production lines for non-destructive, inline monitoring. Key applications include real-time moisture monitoring during fixation and drying for precise thermal control, and objective assessment of fermentation degree by tracking specific spectral features [32].

A current research frontier involves fusing NIR spectral data with RGB image data to leverage both chemical and morphological information for building more robust quality prediction models.

2.2.4. TIC and LiDAR

TIC captures temperature distribution by detecting an object’s emitted infrared radiation. In tea garden management, it can identify canopy temperature anomalies indicative of water stress, providing a direct guide for precision irrigation [33]. However, TIC accuracy is highly sensitive to environmental conditions (ambient temperature, wind, humidity, solar radiation) and leaf emissivity (typically 0.95–0.98 for green vegetation). To obtain reliable data, calibration is essential; common practices include using a reference blackbody, correcting for reflected temperature, applying emissivity adjustment, and measuring under stable weather conditions (midday, clear sky). Without such steps, temperature errors can exceed 2–3 °C, leading to incorrect stress diagnosis [21].

LiDAR, in contrast, provides precise depth information to generate 3D point clouds. This technology is critical for intelligent harvesting, enabling accurate 3D localization of tea buds within complex, overlapping foliage and guiding robotic arms for obstacle avoidance and grasping [34]. Furthermore, LiDAR mounted on UAVs can create detailed 3D digital maps of tea gardens for canopy structure analysis and biomass estimation. However, dense canopy occlusion limits LiDAR performance: laser pulses are often intercepted by upper leaves, resulting in incomplete point clouds of lower buds. Multi-return or full-waveform LiDAR partially mitigates this but still struggles with deep internal structures [22].

2.2.5. UAV

Unmanned Aerial Systems equipped with multispectral, HSI, or RGB cameras form an aerial platform for macro-scale monitoring [24]. They enable rapid, large-area acquisition of high-resolution imagery over difficult terrain. These systems are used to monitor field vigor and uniformity, derive vegetation indices like Leaf Area Index, assess nutrient status, and detect early-stage pest or disease outbreak foci [23]. By analyzing multi-temporal r