Recent Advances and Applications of Nondestructive Testing in Agricultural Products: A Review

Abstract

With the rapid development of agricultural intelligence, nondestructive testing (NDT) has shown considerable promise for agricultural product inspection. Compared with traditional methods—which often suffer from subjectivity, low efficiency, and sample damage—NDT offers rapid, accurate, and non-invasive solutions that enable precise inspection without harming the products. These inherent advantages have promoted the increasing adoption of NDT technologies in agriculture. Meanwhile, rising quality standards for agricultural products have intensified the demand for more efficient and reliable detection methods, accelerating the replacement of conventional techniques by advanced NDT approaches. Nevertheless, selecting the most appropriate NDT method for a given agricultural inspection task remains challenging, due to the wide diversity in product structures, compositions, and inspection requirements. To address this challenge, this paper presents a review of recent advancements and applications of several widely adopted NDT techniques, including computer vision, near-infrared spectroscopy, hyperspectral imaging, computed tomography, and electronic noses, focusing specifically on their application in agricultural product evaluation. Furthermore, the strengths and limitations of each technology are discussed comprehensively, quantitative performance indicators and adoption trends are summarized, and practical recommendations are provided for selecting suitable NDT techniques according to various agricultural inspection tasks. By highlighting both technical progress and persisting challenges, this review provides actionable theoretical and technical guidance, aiming to support researchers and practitioners in advancing the effective and sustainable application of cutting-edge NDT methods in agriculture.

1. Introduction

Nondestructive testing (NDT) refers to methods designed to evaluate the conditions of objects or structures without causing any damage, enabling the detection of internal or surface defects efficiently and non-invasively [1]. Over recent decades, NDT has been extensively utilized for quality assessment across various industrial applications [2], such as pipeline weld inspections in the petroleum industry [3], defect detection in ceramics [4], and integrity evaluation of reinforced concrete structures [5].

In recent years, with continuous advancements in NDT technologies, the application scope of NDT has gradually expanded from traditional industrial sectors to broader fields, showing great promise in agricultural and food industries. Driven by increasing consumer concerns regarding the safety and quality of agricultural products, there is an urgent need for efficient, objective, and non-invasive methods to assess product quality during production processes [6].

NDT techniques have become crucial due to their nondestructive nature, rapid response, and scalability. They have been widely employed in various agricultural quality assessment tasks, including fruit ripeness evaluation, internal defect detection, and product grading. For instance, Wen et al. [7] recently discussed the principles, advantages, and limitations of multiple NDT methods used for the rapid, nondestructive assessment of grape (Vitis sp.) quality. Similarly, Mahanti et al. [8] reviewed advanced NDT approaches applied to fruit damage detection, providing a review of typical practical applications. Furthermore, Zhang et al. [9] successfully applied multispectral imaging techniques in the laboratory for nondestructively evaluating seed vigor in alfalfa (Medicago sativa L.).

Currently, a diverse range of technologies is available for performing nondestructive evaluations. To provide a comprehensive overview of widely adopted NDT methods,

Table 1. Widely adopted nondestructive testing methods and their principles.

Method	Principle	References
Radiographic Testing (RT)	Detects defects by measuring attenuation patterns of radiation passing through objects	[10]
Computed Tomography (CT)	Uses X-ray scanning from multiple angles to reconstruct cross-sectional images, revealing internal structures and density distributions	[11,12]
Magnetic Particle Detection (MPD)	Identifies defects through magnetic particles attracted to leakage magnetic fields	[13]
Acoustic Emission (AE)	Monitors stress waves generated during material deformation or damage to detect internal defects or structural changes	[14]
Computer Vision (CV)	Acquires images through sensors and analyzes them computationally after converting images into numerical matrices	[15]
Eddy Current Testing (ECT)	Detects defects in conductive materials using electromagnetic induction phenomena	[16]
Near-Infrared Spectroscopy (NIRS)	Analyzes how materials respond to near-infrared light to determine internal composition or quality attributes	[17]
Infrared Thermography (IRT)	Detects temperature distributions and thermal conditions of objects through infrared radiation	[18]
Hyperspectral Imaging (HSI)	Creates hypercubes containing spatial and spectral features from multiple wavelength images	[19]
Electronic Nose (EN)	Simulates human olfaction to distinguish complex gaseous components through sensor arrays	[20]

summarizes the fundamental principles of these methods.

Despite the success of these techniques in industrial NDT applications, their direct transferability to agricultural product testing remains limited due to significant differences in product structure, composition, and inspection requirements. Many traditional NDT methods encounter practical constraints in agricultural scenarios. For example, radiographic techniques involving gamma rays, though highly penetrative, may cause biochemical reactions in agricultural products [21]. Magnetic particle and eddy current tests rely on material magnetic or conductive properties, limiting their applicability primarily to detecting ferromagnetic contaminants [22,23].

To better meet agricultural inspection requirements, current widely adopted NDT techniques frequently employed in agriculture include computer vision, near-infrared spectroscopy, hyperspectral imaging, computed tomography, and electronic nose technology. Specifically, computer vision methods primarily rely on visual feature recognition and are highly suited to rapid external quality assessments [24]. Near-infrared spectroscopy can effectively monitor internal composition and quality attributes by quantitatively analyzing spectral data obtained from agricultural products [25]. Hyperspectral imaging integrates spectral and spatial information simultaneously, overcoming limitations associated with near-infrared spectroscopy, such as a lack of spatial data or nonlinear spectral-material relationships, thereby providing more comprehensive and efficient agricultural assessments [26]. Computed tomography, initially developed for medical imaging, reconstructs internal cross-sectional images based on X-ray attenuation coefficients, and its technological advances have progressively broadened its use into agricultural inspections [27,28]. The electronic nose mimics human olfactory systems, using gas sensor arrays to identify complex gaseous components, offering a practical method for rapid, non-invasive evaluation of agricultural products based on volatile organic compounds [29].

Each widely adopted NDT method possesses unique strengths and limitations. Given the significant variation in structure, composition, and specific inspection needs among different agricultural products, selecting appropriate NDT technologies tailored to specific tasks is critical to achieving accurate and efficient assessments. Therefore, this paper presents a review of the technological principles, characteristics, strengths, and limitations of widely adopted NDT methods in agricultural applications. Additionally, practical recommendations are provided for selecting suitable NDT methods according to different agricultural evaluation tasks, including external quality inspection, internal quality and composition analysis, internal defect identification, ripeness assessment, species identification, pesticide or contaminant detection, and freshness determination. Finally, this review discusses existing challenges and proposes future directions for advancing NDT technologies in agricultural product evaluation, aiming to guide and inspire researchers and practitioners in this rapidly evolving field.

2. NDT Techniques for Agricultural Products

2.1. Computer Vision

Compared to traditional manual inspection methods, computer vision techniques significantly improve both the speed and accuracy of agricultural product inspection. Computer vision generally involves acquiring images through imaging sensors (e.g., digital cameras), transforming these images into numerical matrices, and performing further analytical tasks through computational analysis [15]. As illustrated in Figure 1, the typical workflow of a computer vision system begins with image acquisition, followed by preprocessing steps—including noise reduction, image sharpening, and contrast enhancement—to optimize image quality. Subsequently, algorithms or models extract relevant features (such as object location, size, and shape) from the preprocessed images for further analysis. Finally, the identified targets undergo classification and detection based on these extracted features [30].

Computer vision systems often integrate various technologies, including artificial intelligence (AI), pattern recognition, image processing, and machine learning, to accomplish specific detection or classification tasks in diverse application scenarios [31,32]. Traditional computer vision approaches typically rely on classical image processing techniques combined with conventional machine learning algorithms, such as support vector machines (SVMs), k-nearest neighbor (KNN), and shallow neural networks [33,34]. For instance, Ji et al. [35] developed a method for classifying potatoes (Solanum tuberosum L.) by integrating hyperspectral imaging (HSI) with SVMs. They employed K-means clustering to segment potato images into regions, extracted spectral data from each region, and then classified these segments using an SVM model. Experimental results indicated a classification accuracy of up to 90%. Despite their high accuracy for specific tasks, such traditional methods typically depend heavily on manually engineered features, limiting their adaptability across diverse agricultural applications.

In recent years, deep learning methods—particularly convolutional neural networks (CNNs)—have emerged as powerful tools for automatically extracting relevant image features without manual feature engineering [36,37]. CNNs generally comprise convolutional layers, pooling layers, and fully connected layers. In convolutional layers, a set of learnable filters systematically convolve with input data to generate feature maps, thereby eliminating the need for manual feature design. Pooling layers subsequently reduce spatial dimensions, lowering computational complexity. Finally, fully connected layers aggregate the features learned in previous layers to predict classification or detection outcomes [38].

Both traditional computer vision approaches and deep learning-based CNN methods have been extensively applied in nondestructive testing of agricultural products [39,40,41,42].

Table 2. Practical applications of computer vision in nondestructive agricultural product inspection.

Detection Task	Target or Application Example	References
Sorting and grading	Automatic grading and sorting based on external quality, such as sorting fruits and vegetables	[43]
Foreign object detection	Detection of foreign materials within agricultural products, such as cotton (Gossypium spp.) and walnut (Juglans regia L.) contamination	[24,44]
Defect detection and quality assessment	Detection of external defects and quality evaluation during agricultural production processes, such as fruit and vegetable quality control	[45,46,47]
Variety identification	Accurate classification and differentiation of agricultural product varieties from images, such as barley (Hordeum vulgare vulgare L.) and tea (Camellia sinensis) leaves varieties	[48,49]

summarizes representative practical applications of computer vision techniques in agricultural NDT tasks.

2.2. Near-Infrared Spectroscopy

Near-infrared spectroscopy (NIRS) plays a significant role in the field of nondestructive testing, with widespread applications across agriculture, food processing, and other sectors [50,51,52]. Typically, NIRS operates by illuminating samples with near-infrared light in the spectral range of approximately 780–2526 nm [53]. This irradiation stimulates stretching and bending vibrations of chemical bonds such as C–H, N–H, and O–H within the sample, resulting in distinct absorption peaks that allow measurement and analysis of the internal composition and quality attributes of the product [17].

As illustrated in Figure 2, modern NIRS is fundamentally supported by three critical pillars: (1) the principles of vibrational spectroscopy, (2) advancements in instrumentation, and (3) the development and application of chemometric techniques [54]. A typical spectroscopic system comprises a stable light source, photodetectors, a sampling apparatus, and optical isolators. Spectral data collected from the sample are subsequently analyzed using computational approaches combined with chemometric methods [7]. Chemometric methods are particularly essential because they facilitate building accurate and robust mathematical models from complex spectral datasets, thereby enabling both qualitative and quantitative analyses of sample constituents [55].

Within chemometric modeling, variable selection techniques are especially crucial. These methods are designed to identify and extract informative spectral features or wavelengths from complex NIRS datasets [56]. To clarify the most commonly used variable selection methods in chemometrics,

Table 3. Common variable selection methods in chemometrics.

Method	Basic Principles	Advantages	Disadvantages
Manual selection	Removing variables with poor-quality information manually	Simple and easy to perform	Risk of discarding informative variables inadvertently
Univariate linear regression	Selects variables based on individual linear relationships	Simplicity	Selected variables may lack robustness and acceptance
Multiple linear regression	Uses multiple wavelengths to isolate individual absorbers and normalize baselines	Combines multiple wavelength data efficiently	Inconsistent performance under varying noise conditions
Successive projections algorithm (SPA)	Uses vector-space projections to obtain subsets with minimal collinearity	Simplicity of computation and implementation	May select variables with low signal-to-noise ratios
Uninformative variable elimination (UVE)	Variable selection based on the stability of regression coefficients	Removes irrelevant variables and prevents overfitting	Typically selects numerous variables; latent variables often required
Simulated annealing (SA)	Probabilistic global optimization inspired by physical annealing processes [58]	Escapes local minima, seeking global optimum	Computationally intensive; may not guarantee optimal subset
Artificial neural networks (ANN)	Mimics human brain learning; adapts through training data	High flexibility; suitable for complex relationships	Difficult interpretation; prone to overfitting; data-intensive
Genetic algorithms (GA)	Probabilistic heuristic optimization inspired by natural selection theory	Explores multiple subsets thoroughly; finds near-optimal solutions	Slow convergence; sensitivity to initial parameters
Interval selection methods	Based on continuity characteristics of molecular spectra bands	Maintains spectral wavelength continuity	High complexity in optimization

summarizes their basic principles, advantages, and disadvantages [57].

NIRS-based nondestructive techniques have demonstrated considerable potential within agricultural applications. For example, Carlomagno et al. [59] developed a near-infrared (NIR) transmission spectroscopy system using wavelengths between 730 and 900 nm to classify peach (Prunus persica) maturity. They designed multiple acquisition stations and enhanced signal quality through wavelet packet denoising and outlier detection, achieving an accuracy of 82.5% using a minimum-distance classifier. Additionally, Stella et al. [60] conducted a review summarizing the application of NIRS for detecting external quality defects in olives (Olea europaea)—such as mechanical damage, bruising, and pest infestations—as well as real-time field monitoring of key chemical indicators like moisture content, oil content, and phenolic compounds through portable devices. This approach allowed for effective monitoring of quality changes during olive development. More recently, Tan et al. [61] introduced an intelligent NIRS reflectance detection scheme (INIS), employing NIR reflectance spectroscopy combined with a back-propagation (BP) neural network model to measure sugar content in cherry tomatoes (Solanum lycopersicum var. cerasiforme) in the laboratory.

The principal advantages of NIRS include rapid, nondestructive, non-invasive, and chemical-free analysis capabilities. However, the reliance of NIRS techniques on established reference methods and the necessity for chemometric modeling for feature extraction limit their generalizability and broad applicability across diverse agricultural scenarios [50].

2.3. Hyperspectral Imaging

Hyperspectral imaging (HSI) is an advanced technique that captures spatial and spectral information simultaneously by acquiring images across numerous wavelengths. The resulting data structure, often referred to as a “hypercube,” consists of two spatial dimensions (x and y) and one spectral dimension (λ), allowing comprehensive spatial and compositional analysis of the target objects [19]. The typical components of an HSI system include a light source, imaging spectrometer, camera, sample stage, and associated computational processing units [62].

Due to the substantial data volume generated by hyperspectral imaging, advanced data-processing techniques—such as Principal Component Analysis (PCA), support vector machines (SVMs), and convolutional neural networks (CNNs)—are frequently employed to effectively interpret and analyze hyperspectral data cubes [63,64,65]. Despite the inherent complexity involved in data handling and analysis, HSI enables detailed analysis down to minute regions or even individual pixels. This makes it particularly suitable for identifying components and analyzing spatial distribution in complex mixtures.

In agricultural applications, hyperspectral imaging has demonstrated considerable potential for nondestructive testing tasks. For instance, Wang et al. [66] reviewed the application of HSI for evaluating core chemical constituents of fruits—such as carbohydrates, organic acids, water content, and polyphenols—discussing the underlying chemical response mechanisms, existing technical challenges, and future prospects for HSI-based fruit-quality assessment. Similarly, Benelli et al. [67] successfully implemented HSI for nondestructive, in-field monitoring of grape (Vitis sp.) ripeness, highlighting the practical applicability of this technology for real-time agricultural decision-making.

In another example, Vigneau et al. [68] evaluated the feasibility of using hyperspectral reflectance (400–1000 nm) for nondestructive estimation of plant nitrogen concentration under field conditions. They developed chemometric models (R² = 0.903, SEP = 0.327% DM) correlated the leaf nitrogen concentration with hyperspectral reflectance of flattened leaves, validating the sensor performance and reflectance correction process. Additionally, further calibration steps were conducted using isolated leaves from potted plants in greenhouse conditions (R² = 0.889, SEP = 0.481% DM) and under realistic field conditions (R² = 0.881, SEP = 0.366% DM), resulting in a model (R² = 0.875, SEP = 0.496% DM) for predicting nitrogen content under both cultivation conditions. These findings underline the robustness and reliability of HSI for precision agriculture applications.

2.4. Computed Tomography

Computed tomography (CT) was originally developed by Godfrey Hounsfield in 1967 for medical imaging applications [69]. With continuous technological advancements, CT has evolved into a powerful nondestructive inspection tool in agriculture, enabling non-invasive evaluation of the internal structures of agricultural products. In practice, CT systems generate images by rotating an X-ray source around the sample. X-rays pass through the object from various angles, and detectors capture the attenuated signals emerging from the sample [12]. Because different tissues and materials exhibit distinct linear attenuation coefficients (LACs), they absorb X-rays differently, leading to varying intensities detected by the sensors [70,71,72]. Moreover, the attenuation of X-rays in materials is influenced by interactions such as Compton scattering, Rayleigh scattering, and the photoelectric effect. Among these interactions, Compton scattering, which relates closely to the electron density of materials, provides the fundamental physical basis for density imaging [28].

CT technology capitalizes on these variations in X-ray attenuation to generate two-dimensional cross-sectional images, effectively revealing internal structural and density distributions within samples [11,70]. Additionally, multiple sequential cross-sectional images can be reconstructed into three-dimensional images, providing more comprehensive and intuitive spatial structural information [73,74].

Due to its high resolution and nondestructive nature, CT has become essential for internal quality assessment of agricultural products. For instance, Donis-Gonzalez et al. [75] utilized CT technology in the laboratory to nondestructively detect internal decay in chestnuts (Castanea spp.), internal defects in pickling cucumbers (Cucumis sativus), water translucency disorder in pineapples (Ananas comosus), the presence of pits in sour cherries (Prunus cerasus var. Montmorency), and damage caused by plum curculio (Conotrachelus nenuphar) infestation. Their results confirmed the viability of CT imaging for internal quality evaluation of fresh agricultural commodities. Moreover, Hughes et al. [76] applied X-ray micro-computed tomography in the laboratory to extract morphological parameters of cereal spikes and grains. Their findings revealed that high-temperature stress negatively affected both spike length and grain number, particularly in the central spike region. Additionally, under mild stress conditions, grain volume increased while grain number decreased, highlighting the potential of CT technology for detailed grain-quality analysis.

2.5. Electronic Nose

An electronic nose (EN) is an intelligent analytical instrument typically comprising an array of gas sensors, an analog-to-digital converter (ADC), and computational units integrated with pattern-recognition algorithms [77]. In practice, the sensor array interacts with volatile organic compounds (VOCs) emitted by the sample, generating characteristic signal patterns. These patterns are subsequently analyzed using various pattern-recognition algorithms (

Table 4. Common pattern-recognition algorithms used in electronic noses.

Method	Technical Characteristics	Practical Examples
Principal Component Analysis (PCA)	Extracts principal features rapidly by ranking components in descending order	[80,81]
Linear Discriminant Analysis (LDA)	Requires continuous independent variables; suitable for linear classification	[82,83]
K-Nearest Neighbor (KNN)	Simple and intuitive; sensitive to outliers	[84,85]
Support Vector Machines (SVMs)	Identifies optimal hyperplanes maximizing margins between classes; effective with small datasets	[86,87]
Artificial Neural Networks (ANNs)	Mimics human brain functions through iterative learning; suitable for nonlinear data	[88,89]

) [78,79]. By mimicking human olfactory functions, electronic noses can effectively identify and differentiate complex gas mixtures, linking these gas profiles directly to the physicochemical characteristics of the sample and facilitating the distinction between natural and synthetic organic sources [20].

Electronic noses provide a rapid, straightforward, and nondestructive analytical solution for characterizing and distinguishing complex gaseous compounds [90]. Due to their versatility and analytical convenience, EN technology has been widely applied across various sectors, including the food industry [91], pharmaceuticals [92], and cosmetics [93]. Considering the unique chemical compositions typically associated with agricultural products, electronic noses are equally well-suited for nondestructive inspection tasks in agriculture [94]. For example, Singh and Gaur [95] developed an Arduino Uno R3-based electronic nose equipped with MQ4, MQ5, MQ9, and MQ135 sensors, successfully estimating the shelf-life of different edible seeds through VOC profiling.

2.6. Other Techniques

In addition to the aforementioned techniques, several other NDT methods have also been applied to agricultural quality assessment, offering unique sensing principles and complementary advantages.

One notable example is laser light backscattering imaging (LLBI), which measures the spatial distribution of light scattered back from the surface and subsurface layers of a sample when illuminated by a laser source. The backscattering pattern contains information about tissue microstructure, firmness, and internal defects, making it particularly suitable for assessing the maturity, texture, and internal damage of fruits and vegetables [96,97]. For instance, Sanchez, et al. [98] employed LLBI in the laboratory to evaluate sweet potatoes (Ipomoea batatas [L.] Lam). Similarly, Mozaffari, et al. [99] demonstrated that LLBI could effectively detect apricot (Prunus armeniaca L.) under laboratory conditions.

Another emerging approach is multispectral imaging (MSI), which acquires images at a limited number of discrete wavelengths across the visible and near-infrared spectrum. While offering lower spectral resolution than hyperspectral imaging, MSI systems benefit from simpler configurations, faster data acquisition, and reduced computational demands, making them more suitable for real-time, in-field applications [100,101]. MSI has been successfully implemented in a variety of agricultural inspection tasks, such as classifying tomato (Solanum lycopersicum var. cerasiforme) ripeness stages [102], detecting fungal contamination in wheat (Triticum aestivum) kernels [103], and evaluating surface defects in citrus fruits [104].

Overall, these additional sensing modalities enrich the toolkit of nondestructive inspection technologies for agricultural products. By integrating LLBI, MSI, and other emerging optical sensing methods with advanced data analysis approaches, future systems may achieve higher robustness, lower cost, and greater adaptability to diverse production environments.

3. Analysis of Agricultural Nondestructive Testing Tasks

Given the extensive diversity of agricultural products in terms of structural characteristics, internal composition, physical properties, and inspection requirements, selecting an appropriate nondestructive testing (NDT) method tailored to specific tasks is essential to ensure accurate and efficient evaluations. Building upon the overview provided in previous sections,

Table 5. Advantages and disadvantages of widely adopted NDT methods for agricultural product inspection.

Method	Accuracy	Adoption Trend	Advantages	Disadvantages	References
Computer Vision (CV)	Up to 96.7% and 93.8% classification accuracy were achieved during real-time testing on actual samples of apples (Malus domestica) and bananas (Musa spp.) [43]	Widely used in commercial sorting and grading lines	Nondestructive, rapid, high accuracy, low cost	Requires large-scale datasets for generalization; lower accuracy for objects difficult to distinguish from backgrounds; sensitive to lighting conditions	[105,106]
Near-Infrared Spectroscopy (NIRS)	82.5% classification accuracy for peach (Prunus persica) maturity [59]	Portable devices increasingly adopted in field and industry	Minimal or no sample preparation required; rapid; suitable for online quality control systems	Dependent on chemometric modeling for information extraction; challenging to establish universally robust models; requires careful parameter tuning	[107]
Hyperspectral Imaging (HSI)	Qualitative defect detection reported; accuracy not quantified [19]	Primarily research; gradual adoption in high-value crops	Non-invasive, no sample preparation needed; high accuracy even for visually similar samples	High equipment cost; large data volumes; longer processing times; environmental conditions can affect results	[108]
Computed Tomography (CT)	Qualitative defect detection reported; accuracy not quantified [75]	Industrial adoption limited by cost/throughput	High resolution; 3D imaging capability; objective and non-invasive analysis	High equipment cost; relatively slow processing; technically demanding operation	[28]
Electronic Nose (EN)	Shelf-life estimation of edible seeds; accuracy not quantified	Portable systems in niche commercial use	Rapid and convenient; effective for distinguishing complex gas compositions	Difficulty in accurately identifying single aromatic compounds	[95,109]

summarizes the representative accuracy, adoption trends, key advantages, and limitations of the primary NDT techniques applied in agricultural contexts, providing practical guidance for method selection.

Given the diversity of inspection tasks for agricultural products, such tasks generally include external quality inspection, internal quality inspection, internal defect identification, ripeness assessment, variety identification, pesticide or contaminant detection, and freshness evaluation. To provide clarity regarding suitable NDT technologies for specific detection objectives,

Table 6. Commonly used NDT methods according to different agricultural inspection tasks.

Detection Task	Detect the Target	Commonly Used Methods	Product Type	Application Examples
External Quality Inspection	Shape, color, size, and surface defects; suitable for grading and sorting of fruits and vegetables	Computer vision, hyperspectral imaging	Fruits, Root vegetables, Cereals, Legumes	[47,110,111]
Internal Quality Inspection	Evaluate the quality indicators of agricultural products such as internal components and defects	Hyperspectral imaging, near-infrared spectroscopy	Fruits, Root vegetables	[112,113]
Internal Defect Detection	Internal structural defects such as decay, cracks, or voids	Computed tomography, hyperspectral imaging	Root vegetables, Fruits	[114,115]
Ripeness Evaluation	Ripeness classification of fruits	Hyperspectral imaging, near-infrared spectroscopy, electronic nose (Fruits with volatile organic compounds)	Fruits, Cereals	[116,117,118]
Variety Identification	Identification and classification of different varieties and origins	Hyperspectral imaging, computer vision, near-infrared spectroscopy	Fruits, Root vegetables, Cereals, Legumes	[48,119,120]
Pesticide or Contaminant Detection	Rapid detection of pesticide residues or contaminants (e.g., heavy metals) on or within products	Electronic nose, near-infrared spectroscopy	Fruits, Root vegetables, Cereals	[121,122]
Freshness Detection	Freshness evaluation of agricultural products including meat, fruits, vegetables, eggs, etc.	Near-infrared spectroscopy, hyperspectral imaging, electronic nose, computer vision	Fruits, Root vegetables, Eggs, Meat	[123,124,125,126]

categorizes major inspection tasks and highlights their primary targets and commonly used NDT methods, supported by representative examples from the existing literature.

In addition to differences among inspection tasks, the suitability of each NDT technique also varies across different categories of agricultural products. For example, near-infrared spectroscopy is particularly effective for cereals and legumes due to its high sensitivity to moisture and protein content, whereas hyperspectral imaging is often preferred for fruits and vegetables where both external appearance and internal quality are critical. Computer vision techniques are widely adopted for uniform-sized products such as graded fruits, eggs, and nuts, while computed tomography is more commonly applied to root vegetables and other products where internal defects significantly affect market value. Electronic noses, by contrast, are highly effective for products with characteristic volatile profiles, such as coffee beans, tea leaves, and certain herbs. Recognizing these affinities can further improve the efficiency and accuracy of method selection.

Building on the comparative insights provided in Table 5 and Table 6, several guiding principles can be distilled for selecting suitable NDT techniques in different agricultural inspection tasks. For instance, surface-related attributes such as color, size, and external defects are most efficiently assessed using computer vision or multispectral/hyperspectral imaging, whereas internal composition and quality attributes (e.g., moisture, sugar content, oil content) are better-suited to near-infrared spectroscopy or hyperspectral imaging. Internal structural defects (e.g., decay, voids, cracks) require penetrating techniques such as computed tomography or, in certain cases, laser light backscattering imaging, while aroma- or gas-related freshness indicators can be effectively monitored with electronic noses. Real-time, high-throughput scenarios benefit from fast-acquisition systems (e.g., CV, MSI, NIRS), whereas laboratory-based in-depth analyses can utilize high-precision methods such as CT or high-end HSI. Cost, portability, and scalability should also be considered—low-cost approaches are preferable for widespread deployment, while high-cost systems may be reserved for specialized tasks where their capabilities are essential. In many cases, integrating complementary techniques (e.g., CV with NIRS, or EN with HSI) can enhance overall accuracy and robustness, particularly when dealing with diverse agricultural products and variable environmental conditions.

4. Challenges and Future Directions

Although NDT technologies have demonstrated significant potential and have been successfully applied to various agricultural products, several key challenges remain unresolved. Addressing these challenges requires not only technical innovation but also strategic thinking from multiple perspectives, including methodological improvements, technology integration, data management, and practical implementation.

First, one of the critical challenges lies in the complexity and heterogeneity of agricultural products themselves. The enormous biological variability within even single product categories—due to differences in varieties, growth conditions, and post-harvest handling—poses considerable difficulties in achieving universal detection models with robust and generalized performance. Future studies must therefore focus on developing adaptive algorithms and generalized modeling strategies that leverage advanced deep learning techniques, such as transfer learning and meta-learning. Such approaches can enhance the flexibility of models, allowing them to perform consistently across diverse scenarios and product categories, thereby overcoming limitations associated with product variability.

Second, despite their individual advantages, single-modality NDT techniques are inherently limited in their capabilities to comprehensively characterize agricultural products. For example, computer vision excels in rapid external quality assessment but lacks the capability to detect internal compositional variations effectively. Conversely, hyperspectral imaging and computed tomography (CT) provide accurate internal analysis yet require costly equipment and sophisticated data processing. Therefore, integrating complementary NDT methods through multimodal data fusion is crucial. By strategically combining techniques—such as computer vision, near-infrared spectroscopy (NIRS), hyperspectral imaging (HSI), electronic nose (EN), and CT—researchers can leverage the synergistic benefits, enhancing overall detection accuracy, speed, and robustness. However, achieving effective multimodal fusion requires resolving critical issues such as synchronization, dimensionality reduction, and feature extraction from heterogeneous datasets, which remain significant research areas.

Third, the proliferation of deep learning and artificial intelligence (AI) techniques presents promising opportunities to tackle current limitations in agricultural NDT. However, AI-driven approaches rely heavily on the availability and quality of annotated datasets, which are often limited, costly, or challenging to acquire in agricultural contexts. The future direction should therefore emphasize developing advanced data augmentation methods and unsupervised or semi-supervised learning techniques to minimize dependency on manually annotated datasets. Furthermore, exploring explainable AI (XAI) is essential to improve interpretability and foster greater confidence and acceptance among stakeholders, facilitating practical implementation in the agriculture industry.

Fourth, considering practical deployment, there remains a significant gap between laboratory-developed methods and their integration into actual agricultural production lines and farm-level operations. Although various innovative techniques exhibit promising laboratory results, their adoption in real-world scenarios is often hindered by cost-effectiveness issues, ease of use, portability, and the capacity for continuous on-site monitoring. Future research must prioritize the design of compact, affordable, and user-friendly portable NDT devices, coupled with real-time analysis capability, enabling rapid decision-making directly at the farm or production line. Achieving such advancements will accelerate widespread adoption of NDT, thus significantly benefiting agricultural quality control and productivity.

Lastly, to maximize the value of NDT in agriculture, it is crucial to establish an integrated cloud-based analytical platform that consolidates multimodal data across diverse geographical regions and production systems. Such a platform, leveraging advancements in cloud computing, big data analytics, and the Internet of Things (IoT), would facilitate seamless data sharing, real-time analysis, predictive modeling, and benchmarking of product quality standards. Developing robust cloud-based systems can not only promote large-scale application of NDT technologies but also enable intelligent agricultural management by providing stakeholders with data-driven decision-making capabilities and comprehensive insights.

In conclusion, future research in agricultural NDT should be strategically oriented toward overcoming biological variability, enhancing multimodal integration, leveraging AI and data-driven solutions, improving practical applicability, and establishing comprehensive data-sharing platforms. Such concerted efforts will undoubtedly enhance agricultural quality evaluation and profoundly reshape the future of intelligent agriculture.

5. Conclusions

With the rapid development of agricultural intelligence and the increasing demand for food safety and quality assurance, NDT technologies have become indispensable tools for monitoring product quality throughout the agricultural production chain. This review comprehensively summarizes several widely adopted NDT techniques currently applied to agricultural products, including computer vision, near-infrared spectroscopy, hyperspectral imaging, computed tomography (CT), and electronic noses. Furthermore, the adaptability and specific advantages of each method are analyzed in the context of different agricultural inspection tasks.

Computer vision is highly effective for rapid external quality grading. Near-infrared spectroscopy excels in analyzing internal chemical compositions. Hyperspectral imaging simultaneously captures spectral and spatial distribution information, providing comprehensive insights into product quality. CT imaging effectively detects internal structural defects, while electronic nose technology uniquely specializes in identifying complex gaseous and volatile components. This paper provides theoretical guidance and practical recommendations for selecting suitable NDT methods and combining them according to specific detection objectives, thereby achieving efficient, accurate, and environmentally friendly inspections.

Moreover, future research directions are identified, emphasizing the importance of advancing multimodal data-fusion technologies and developing portable smart devices alongside cloud-based big-data platforms. These advancements are expected to facilitate a shift toward more intelligent, cost-effective, and portable solutions for agricultural NDT, providing robust technical support for modern agriculture.