Nondestructive Quality Detection of Characteristic Fruits Based on Vis/NIR Spectroscopy: Principles, Systems, and Applications

Abstract

Nondestructive quality detection of characteristic fruits is essential for ensuring nutritional value, economic viability, and consumer safety in global supply chains, yet traditional destructive methods compromise sample integrity and scalability. Visible and near-infrared (Vis/NIR) spectroscopy offers a transformative solution by enabling rapid, non-invasive multi-attribute quantification through molecular overtone vibrations. This review examines recent advancements in Vis/NIR-based fruit quality detection, encompassing fundamental principles, system configurations, and detection strategies calibrated to fruit biophysical properties. Firstly, optical mechanisms and system architectures (portable, online, vehicle-mounted) are compared, emphasizing their compatibility with fruit structural complexity. Then, critical challenges arising from fruit-specific characteristics—such as rind thickness, pit interference, and spatial heterogeneity—are analyzed, highlighting their impact on spectral accuracy. Applications across diverse fruit categories (pitted, thin-rinded, and thick-rinded) are systematically reviewed, with case studies demonstrating the robust prediction of key quality indices. Subsequently, considerations in model development and validation are presented. Finally, persistent limitations in model transferability and environmental adaptability are discussed, proposing future research directions centered on integrating hyperspectral imaging, AI-driven calibration transfer, standardized spectral databases, and miniaturized, field-deployable sensors. Collectively, these methodological breakthroughs will pave the way for autonomous, next-generation quality assessment platforms, revolutionizing postharvest management for characteristic fruits.

1. Introduction

The characteristic fruits play a pivotal role in global nutrition, economic stability, and cultural practices, serving as essential sources of vitamins, antioxidants, and dietary fiber while supporting agricultural livelihoods worldwide [1,2,3]. The imperative for precise quality evaluation arises from consumer demand for consistent sensory attributes and food safety, particularly as international trade expands. Traditional destructive methods (e.g., refractometry, chromatography) compromise sample integrity and lack scalability for industrial applications. Consequently, non-destructive technologies like visible and near-infrared (Vis/NIR) spectroscopy (400–2500 nm) have emerged as transformative solutions [4,5,6]. This technique leverages molecular overtone and combination vibrations to quantify internal attributes, such as soluble solids content (SSC), acidity, and bioactive compounds, without physical alteration [7,8,9]. Validation studies on apples, peaches, and mangoes demonstrate their efficacy, with prediction accuracies for SSC and defect detection [10]. These capabilities enable real-time sorting, reducing postharvest losses in commercial supply chains [11].

The effectiveness of Vis/NIR spectroscopy, however, is intrinsically linked to fruit-specific characteristics. Variations in size, rind thickness, internal structures, and spatial distribution of target compounds significantly influence light scattering, absorption depth, and spectral reproducibility [12]. For instance, thick-rinded fruits (e.g., pomelo) attenuate light penetration, necessitating transmission-mode optimization, while delicate berries (e.g., blueberries) require reflectance adaptations to mitigate surface scattering [13]. As visually summarized in Figure 1, these morphological and compositional traits form the basis for a practical classification of fruits into three main categories: fruits with pits/kernels, delicate fruits with thin rind, and large-sized fruits with thick rind, each requiring specific optical configurations and detection priorities. Such morphological and compositional diversity underscores the need to systematize technology-fruit compatibility frameworks to guide hardware design and algorithm selection.

To enhance clarity and reproducibility for the review process, a PRISMA flow diagram detailing the systematic identification, screening, and inclusion process of literature is provided as Supplementary Figure S1. This review addresses critical gaps by systematizing advancements in Vis/NIR-based fruit quality detection. Unlike existing literature, we establish structured correlations between fruit morpho-physiological traits and spectral methodologies, providing actionable insights for developing tailored detection protocols for fruits with distinct characteristics. Furthermore, we evaluate multiple technological advances, including hyperspectral imaging, portable sensors, and AI-driven calibration transfer for industrial scalability. This review systematically progresses through fundamental principles and system configurations, followed by an analysis of detection challenges inherent to fruit morphological and physiological characteristics. Subsequent sections catalog diverse applications across fruit categories and subsequently present critical considerations in model development and validation, culminating in proposed future research directions for next-generation non-destructive platforms.

2. Vis/NIR Spectroscopy for Quality Detection of Fruits

Vis/NIR spectroscopy has emerged as an effective non-destructive technique for the quality detection of fruits [14]. This method is grounded in the interaction between light and matter, allowing for the evaluation of both the external and internal characteristics of fruits, which are directly related to their quality attributes. The following discussion elaborates on the fundamental principles underlying Vis/NIR spectroscopy, detailing the mechanisms of light interaction with fruit tissues, identifying critical factors that influence spectral data acquisition, and describing the application of chemometric models for interpreting spectral information to assess fruit quality.

2.1. Fundamental Principles

The fundamental principle of Vis/NIR spectroscopy is based on the interaction of electromagnetic radiation, typically in the range of 400–2500 nm, with the molecular structures within fruit tissues. When incident light interacts with fruit surfaces, it may be absorbed, reflected, or transmitted depending on the fruit’s optical properties and tissue composition. These interactions provide spectral data that are correlated with the physicochemical attributes of fruits, such as SSC, titratable acidity (TA), internal browning, and watercore [15,16,17].

In the visible region (380–760 nm), absorption is predominantly influenced by pigments such as chlorophyll, carotenoids, and anthocyanins, which contribute to the coloration of the fruit surface and provide insight into external ripeness and cultivar characteristics [18]. In the near-infrared region (760–2500 nm), absorption arises primarily from the overtone and combination vibrations of chemical bonds, including O–H, C–H, and N–H, found in water, sugars, organic acids, and other quality-related compounds within the fruit tissue [19]. These absorption features create a spectral fingerprint that can be quantitatively linked to the concentration of internal components, especially SSC and water content.

Vis/NIR detection systems generally operate in three modes: transmission, reflection, and diffuse reflectance [20,21,22,23]. In full-transmission mode, the light passes entirely through the fruit, with the detector capturing the transmitted spectrum. This method is particularly suitable for evaluating large or thick-skinned fruits like pomelo and citrus, where internal quality information, including SSC, can be robustly extracted due to deeper light penetration [24]. For instance, transmission-mode Vis/NIR spectroscopy (400–1100 nm) combined with Savitzky–Golay (SG) filtering, standard normal variate (SNV), and mean normalization preprocessing effectively extracted features for detecting internal fungal infection (Alternaria alternata), achieving classification accuracies of 90–97% using back-propagation neural networks (BPNN), demonstrating the capability of transmission spectroscopy in capturing internal biochemical changes caused by disease [25]. Reflectance mode measures the light reflected off the fruit surface and is often used for assessing surface characteristics or shallow internal traits such as the skin integrity or pigment composition [26]. Diffuse reflectance, measured using an integrating sphere system, captures both surface and subsurface scattering signals, enabling the extraction of bulk optical properties such as the absorption coefficient and reduced scattering coefficient. This approach has been effectively applied to assess internal quality attributes like SSC and firmness in apples [27].

Several factors influence the accuracy and reliability of Vis/NIR measurements. The optical properties of the fruit, including skin thickness, internal structure, and tissue homogeneity, significantly affect the penetration depth and scattering behavior of light. In apples, both watercore and ripeness levels have been effectively predicted by fine-tuning the light path configuration and using chemometric algorithms, such as convolutional neural networks (CNN) combined with competitive adaptive reweighted sampling (CARS) [16]. Moreover, fruit variety, ripening stage, and growing conditions contribute to spectral variability. This variability is particularly evident in fruits like strawberries, where pigment content and surface reflectance vary significantly across ripening stages. Accurate prediction of SSC in such cases requires robust calibration models capable of compensating for these variances. In white strawberries, for example, different modeling approaches using Vis/NIR spectroscopy revealed variability in performance due to structural and compositional differences in fruit tissue [28].

2.2. Basic Procedure

Applications of Vis/NIR spectroscopy for nondestructive quality detection of fruits generally follow a structured procedure to ensure accuracy, robustness, and practical utility. This process typically includes sample selection, spectral acquisition, reference measurement of quality indices, data preprocessing, model development, validation, and potential model updating or transfer.

Sample selection is the foundational step. Fruits should be selected to represent the natural variability in terms of cultivar, maturity, size, origin, and, if applicable, health or defect status [29,30,31]. For instance, studies often involve fruits from different geographical regions or at various maturation stages to build models capable of handling real-world variability [32]. The sample set must be of sufficient size; it is commonly divided into a calibration set for model training and an independent prediction set for validation, often in a ratio around 3:1 or using cross-validation techniques [33]. Spectral acquisition is performed using a Vis/NIR spectrometer, which can be a portable handheld device, a benchtop instrument, or an online sorting system [34]. Measurements are usually taken at multiple points on the fruit’s equator to account for potential heterogeneity [30]. Key instrument parameters, such as integration time and scanning range, must be optimized and kept consistent. Prior to scanning, instruments are calibrated using a white reference and a dark current measurement to minimize errors. Environmental conditions, particularly temperature, should be controlled or samples equilibrated to room temperature to prevent spectral drift [35].

Following spectral acquisition, reference physicochemical measurements are destructively performed on the same fruit samples to obtain accurate ground truth values for targeted quality attributes. These commonly include internal quality parameters such as SSC, TA [36], firmness [37], dry matter, and moisture content; chemical constituents like ascorbic acid, total phenolics, and total flavonoids [13]; sensory attributes including taste, texture, and color [37]; as well as defects and disorders such as bitter pit [29] or moldy core [38]. Standard laboratory methods are employed for these measurements, for instance, refractometry for SSC, texture analyzers for firmness, titration for TA and vitamin C, and chromatographic or spectrophotometric techniques for specific compounds. Subsequently, data preprocessing is applied to improve the signal-to-noise ratio and remove unwanted spectral variations unrelated to chemical properties [39,40]. Common techniques involve smoothing using SG filters to reduce high-frequency noise [16], scatter correction methods such as SNV or multiplicative scatter correction (MSC) to mitigate effects from sample morphology [41,42], and derivatization (first or second derivative via SG) to resolve baseline shifts and enhance spectral features. Additional methods like baseline correction, area normalization, and noise reduction algorithms may also be applied. The selection of preprocessing methods is often guided by comparing subsequent model performance [43]. To enhance model efficiency and interpretability, feature wavelength selection is frequently employed to reduce data dimensionality and address multicollinearity. Algorithms such as recursive feature elimination (RFE) [44], genetic algorithm (GA) [45], and Boruta algorithm (BA) [32] are used to identify wavelengths most informative for the target attributes.

Model development establishes a mathematical relationship between preprocessed spectral data (X-matrix) and reference values (Y-matrix). Both linear and nonlinear techniques are utilized. Partial least squares regression (PLSR) is widely used for regression tasks [46], and partial least squares discriminant analysis (PLS-DA) for classification [47]. Nonlinear machine learning methods, such as support vector machine (SVM) [46], artificial neural networks (ANN) [17], extreme gradient boosting (XGBoost), and feedforward neural networks (FFNN) [36] are increasingly adopted to capture complex relationships. Model complexity is carefully optimized to prevent overfitting. Model evaluation is conducted using the calibration set for training and an independent prediction set for validation. For regression models, common metrics include the coefficient of determination (R²), root mean square error of calibration (RMSEC), root mean square error of prediction (RMSEP), and ratio of performance to deviation (RPD). Classification models are assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic (ROC) curve. For sustained practical applications, especially in online systems, model maintenance through updating or transfer is essential. Calibration models may degrade due to changes in the instrument, variety, season, or growing conditions. Techniques such as incremental updating with new samples or calibration transfer algorithms help maintain prediction accuracy without rebuilding the model entirely [48].

Applications of Vis/NIR spectroscopy for nondestructive fruit quality detection are a systematic process encompassing representative sampling, consistent spectral measurement, rigorous reference analysis, tailored data preprocessing, chemometric modeling, and validation. Adherence to this workflow supports the development of reliable tools suitable for integration into postharvest sorting and quality control systems.

2.3. Typical Vis/NIR System

2.3.1. Portable Detection Equipment

Portable Vis/NIR detection devices have been widely developed for in-field or on-site fruit quality assessment due to their flexibility, low cost, and ease of operation [49]. These systems typically integrate light sources, spectrometers, embedded controllers, batteries, and displays in a compact housing. A self-developed portable Vis/NIR transmittance system was designed for internal quality detection in apples, incorporating two 12 V 100 W halogen lamps with filters (500–1100 nm), a micro-spectrometer (SE2050, 700–1100 nm), a voltage-adjustable power supply, and a specialized 3D-printed food-grade silicone tray to secure samples. The angle of illumination is adjustable via a universal joint to accommodate fruits of different sizes, and the entire system is enclosed in a protective casing to avoid external interference (Figure 2A) [16].

In terms of light source type, halogen lamps are commonly adopted for providing broadband Vis/NIR illumination with high stability, as seen in detectors developed for SSC measurement in kiwifruit, apricot, and nectarine. Alternatively, LED arrays with specific peak wavelengths have been used in low-cost sweetness and firmness grade detectors for kiwifruit to reduce power consumption and hardware size [50]. Some systems also employ miniaturized halogen lamps coupled with optical fiber probes to guide light onto the sample surface. The optical arrangement between the light source and detector is mainly configured in reflection mode for portable applications. This diffuse reflection mode enhances the acquisition of chemical information from fruit tissues. For transmission mode, although it can achieve higher detection accuracy in some cases, it requires higher light intensity and is more sensitive to fruit size and shape, limiting its use in portable settings [51]. System integration and miniaturization are critical for portability. Many systems employ single-board computers or microcontrollers as the core processing unit. These are coupled with micro-spectrometers such as the STS-NIR (650–1100 nm) or C12880MA (340–850 nm), enabling lightweight and compact designs. Power is usually supplied by lithium batteries to ensure cordless operation. To improve signal quality and avoid interference from ambient light, light shielding strategies are incorporated. Some devices are equipped with handmade dark chambers or probe hoods that cover the fruit surface during measurement. Additionally, some systems implement software-based corrections such as white and dark reference calibration to enhance spectral accuracy. These portable devices have been validated for detecting SSC, acidity, firmness, and internal defects in fruits such as kiwifruit, apple, pear, and lime, providing rapid results (within 2–3 s per sample) with satisfactory accuracy [52,53]. Their practicality makes them suitable for field use by farmers, distributors, and consumers.

2.3.2. Online Detection System

Online Vis/NIR detection systems are essential for high-throughput sorting and grading of fruits in industrial applications. A key component of such systems is the fruit handling and transport mechanism, which ensures stable and precise positioning of samples under the optical sensing unit. As shown in Figure 2B, an online full-transmittance spectrum measurement system was developed for moldy core detection in apples [38]. This system comprises a conveyor platform with adjustable speed (0.5 m/s), a black chamber to minimize stray light, an illumination unit with a 150 W halogen lamp and a condensing lens, a high-sensitivity spectrograph (615–1044 nm), a position monitoring unit with optical switches and encoders, and a computer for control. The system captures approximately 30 full-transmittance spectra per apple using a short integration time of 5 ms, enabling rapid online assessment with a throughput of about 5 fruits per second. Commonly, conveyor belts or motorized roller systems are employed to transport individual fruits through the detection zone at constant speeds. For instance, in systems designed for apples, fruit cups or custom-designed trays are often used to hold and orient the samples [54]. These trays may have a central aperture to allow transmittance-based measurements, ensuring that the optical path remains consistent during movement [55]. To mitigate the influence of fruit size and orientation on spectral acquisition, some advanced systems incorporate adaptive fruit support mechanisms. These may include adjustable trays or rotating platforms that allow multiple measurements from different angles, improving the representativeness of the spectral data [56].

The speed of the conveyor system is critical for balancing throughput and detection accuracy. Typical speeds range from 0.14 m/s to 0.5 m/s, enabling the processing of 3–5 fruits per second [55]. Higher speeds may reduce dwell time under the sensor, potentially affecting spectral quality, while lower speeds limit throughput. Therefore, optimizing transport speed is essential for efficient online operation. Additionally, synchronization between the fruit’s movement and the spectral acquisition trigger is achieved using sensors, such as optical encoders or photoelectric switches, which signal the spectrometer to capture data when the fruit is precisely aligned with the optical path [57].

Another important aspect is the modularity of online systems. Some designs incorporate multi-mode detection, allowing switching between reflectance and transmittance measurements by reconfiguring light sources and detectors. For example, in systems targeting internal defects like moldy core or watercore, transmittance-mode layouts are preferred due to their ability to probe deeper into the fruit tissue. In such cases, high-power halogen lamps (e.g., 100–150 W) are often used to ensure sufficient light penetration [56]. Furthermore, to maintain consistency across varying fruit sizes and shapes, some systems employ dynamic adjustment mechanisms, such as movable light sources or automated height adaptors for optical components [54].

2.3.3. Vehicle-Mounted Detection System

Vehicle-mounted Vis/NIR spectroscopy systems represent an advanced platform for large-scale, mobile nondestructive detection of fruit quality attributes and diseases in agricultural settings. These systems are typically integrated onto ground-based vehicles, such as tractors or all-terrain cars, enabling efficient scanning of orchards and plantations with minimal manual intervention. The primary advantage of vehicle-mounted systems lies in their mobility and ability to cover extensive areas quickly, while maintaining consistent spectral acquisition conditions under real-field environments. For instance, Sankaran et al. developed a vehicle-mounted Vis/NIR system operating in the 440–900 nm range to detect Huanglongbing (HLB) in citrus trees [58]. This system successfully distinguished between healthy and infected trees with an accuracy of approximately 87%, demonstrating the feasibility of vehicle-based platforms for practical disease monitoring. Similarly, a compact in-field sorting system integrated onto a self-propelled harvest assist platform utilized Vis/NIR transmittance spectroscopy to assess apple quality attributes non-destructively (Figure 2C) [59]. The system employed a three-lane rotating mechanism to achieve a throughput of up to 9 apples per second, with spectral data acquired in the 615–1044 nm range using a short-integration-time mode (5 ms) to enable rapid online detection under field conditions. The integration of GPS technology further allows for precise geo-referencing of spectral data, facilitating the creation of spatial maps of disease prevalence or fruit quality parameters across plantations.

Despite their advantages, vehicle-mounted systems face several challenges that can impact performance. Environmental factors, such as varying sunlight intensity, shadows, and canopy occlusion, may introduce noise into spectral measurements. Additionally, vehicle movement can cause vibrations that affect the stability of optical components, potentially leading to reduced spectral consistency. To mitigate these issues, advanced preprocessing algorithms and robust mounting designs are often employed. For example, some systems incorporate inertial measurement units (IMUs) to compensate for motion artifacts, while machine learning models are used to enhance classification accuracy by filtering out environmental variability. Future developments in vehicle-mounted systems will likely focus on integrating multi-sensor approaches, such as combining Vis/NIR with thermal or hyperspectral imaging, to improve detection capabilities for early-stage diseases and subtle quality defects [60,61,62,63].

Figure 2. Schematic diagrams of three typical Vis/NIR spectroscopy system configurations for fruit quality detection. (A) Portable device [16]. (B) Online system [38]. (C) Vehicle-mounted system [59].

Figure 2. Schematic diagrams of three typical Vis/NIR spectroscopy system configurations for fruit quality detection. (A) Portable device [16]. (B) Online system [38]. (C) Vehicle-mounted system [59].

3. Influence of Fruit Characteristics on Quality Detection by Vis/NIR Spectroscopy

The effectiveness of Vis/NIR spectroscopy for fruit quality detection is significantly influenced by the intrinsic morphological and physiological characteristics of fruits themselves. As illustrated in Figure 3, these influencing factors can be systematically categorized into four primary aspects: internal structures (pits, kernels, and cavities), rind properties, fruit size and shape, and the spatial distribution of internal components. Understanding these characteristic-dependent challenges is crucial for developing robust optical configurations, adaptive sampling methods, and calibrated chemometric models to improve the accuracy and universality of nondestructive detection techniques.

3.1. Fruit Pits, Kernel, and Cavity

The presence of fruit pits, kernels, and internal cavities significantly influences Vis/NIR spectroscopy-based nondestructive quality detection by altering light propagation patterns, introducing variability in spectral data, and complicating the measurement of internal physicochemical properties. In transmission mode, these internal structures can cause light scattering and absorption variations, leading to distorted spectral signatures. For instance, cavities or air-filled spaces may enhance light transmission due to reduced density, as observed in watercore-affected apples, where water-soaked tissues exhibit higher spectral intensity [64]. Conversely, dense pits or kernels can absorb or block light, reducing overall spectral intensity and mimicking the effects of internal defects like a moldy core, which shows lower reflectance in infected regions [54]. This variability necessitates careful selection of measurement orientations and optical configurations to ensure accurate signal acquisition, such as using multiple light sources or specific fruit positioning to mitigate shadowing effects from pits [57].

Regarding physicochemical measurement, pits and kernels can skew assessments of SSC or dry matter, as their composition differs markedly from fleshy tissues. For example, in citrus fruits, the presence of seeds or internal cavities may lead to underestimation of SSC, as these structures do not contribute to sugar content similarly to pulp [65]. To address this, calibration models must be developed using representative samples that include variations in internal structures, often requiring destructive validation to correlate spectral data with actual compositional changes [66].

Data processing methods are adapted to compensate for these influences, employing preprocessing techniques such as MSC or SNV to reduce scattering effects caused by internal heterogeneities. Wavelength selection algorithms, like those based on variational mode decomposition (VMD), help isolate relevant spectral features from noise introduced by pits or cavities [54]. Furthermore, multivariate analyses like PLSR are enhanced with ensemble approaches to improve robustness against structural variability, ensuring reliable prediction of quality attributes despite the presence of internal obstacles. Overall, understanding and mitigating the effects of fruit pits, kernels, and cavities are crucial for advancing Vis/NIR spectroscopy toward accurate and universal application in fruit quality monitoring.

3.2. Fruit Rind

The physicochemical and structural properties of fruit rinds significantly influence the acquisition, interpretation, and modeling of Vis/NIR spectral data [67]. The rind’s surface characteristics, such as texture, thickness, wax composition, and pigment concentration, affect light penetration and scattering, thereby shaping the spectral signatures acquired in reflectance mode. For instance, in citrus fruit, the presence of high carotenoid and chlorophyll content in sun-exposed rinds alters absorption in visible regions (500–700 nm), while variations in rind carbohydrates and water content dominate NIR absorption bands [68]. These compositional differences directly impact the effectiveness of spectral preprocessing methods. Techniques such as MSC and SG derivatives are often employed to mitigate scattering effects caused by surface irregularities and enhance spectral features related to biochemical constituents.

Moreover, the rind’s role in physiological disorders, such as rind breakdown in mandarins or sunscald in pears, adds complexity to quality detection [69]. For example, disorders manifesting as surface browning or pitting introduce distinct reflectance patterns, which can be identified using interval partial least squares (iPLS) to select informative wavelength regions. In such cases, spectral regions between 500–600 nm and 650–750 nm have proven effective in classifying damage severity and predicting disorder susceptibility. Additionally, discriminant models like PLS-DA and linear discriminant analysis (LDA) leverage these spectral differences to categorize fruit based on rind condition or preharvest exposure. However, the predictive accuracy of models may be compromised by natural rind heterogeneity, emphasizing the need for representative sampling and robust calibration strategies that account for biological variability.

3.3. Fruit Size and Shape

The size and shape of fruits significantly influence the acquisition and interpretation of Vis/NIR spectral data, primarily due to their effects on light penetration, scattering, and path length variability. Larger fruits, such as tomatoes with masses ranging from 8.6 g to 212.0 g, exhibit greater light attenuation and scattering within internal tissues, leading to reduced spectral intensity and increased noise in reflectance measurements [70]. Irregular shapes cause inconsistent probe contact during spectral acquisition, resulting in heterogeneous reflectance patterns and potential misalignment of optical sampling points. These physical variations introduce biases in spectral data, particularly in the NIR region (800–2500 nm), where light interaction with biochemical components is depth-dependent [71].

For physicochemical indicator measurement, size and shape affect the representativeness of spectral sampling. In larger or asymmetrical fruits, uneven distribution of compounds like SSC or carotenoids may lead to suboptimal spectral averaging, as localized measurements might not capture overall fruit quality [72]. For instance, in tomatoes, shape variations (e.g., round vs. pear-shaped) alter the proximity of vascular tissues and seed cavities to the surface, influencing the correlation between spectral features and internal attributes [70]. This heterogeneity necessitates multiple sampling points per fruit to improve accuracy. Data processing methods must account for these physical variations. Scatter correction techniques, such as MSC and SG derivatives, are essential to mitigate size- and shape-induced scattering effects. Additionally, the development of separate calibration models for distinct size or shape categories is often required to enhance prediction robustness. Borba et al. emphasized the need for size-specific models to accurately predict SSC in tomatoes [53]. Furthermore, algorithms like PLSR benefit from incorporating size-related predictors or using stratified cross-validation to manage variability.

3.4. Spatial Distribution of the Component Within the Fruit

The spatial heterogeneity of internal components within fruit also influences the effectiveness of Vis/NIR spectroscopy for non-destructive quality detection. This heterogeneity necessitates careful consideration of signal acquisition methods to ensure representative sampling [73,74]. For instance, in Gros Michel bananas, the SSC is not uniformly distributed, leading researchers to standardize the measurement location at the end of the banana finger to minimize variability and enhance reproducibility [75]. Similarly, for detecting cork spot disorder in ‘Akizuki’ pears, a five-point sampling method within a delineated elliptical area on the fruit equator was employed to capture spectral variations caused by the uneven distribution of mineral elements like calcium and boron [76]. This approach mitigates the risk of misclassification arising from localized component concentrations. In apples susceptible to bitter pit, which often originates in the calyx end, spectral data are collected from multiple fixed areas in the calyx half to monitor the development of symptoms related to calcium deficiency [77]. The inherent spatial variation impacts physicochemical measurement methods, often requiring destructive validation techniques to be performed on tissue from the same region scanned by the spectrometer to establish robust calibration models. Furthermore, data processing methods must account for this distribution. Techniques like MSC and SNV are frequently applied to spectral data to reduce scatter caused by physical differences and to highlight chemical information, thereby improving the model’s ability to correlate spectral features with component levels despite spatial inconsistencies. Consequently, understanding and addressing the spatial distribution of internal components is crucial for optimizing light interaction, ensuring accurate model calibration, and ultimately achieving reliable non-destructive quality assessment.

Based on the applicability of Vis/NIR spectroscopy to fruits of different morphological characteristics, we recommend corresponding spectral strategies, including optical modes, preprocessing methods, and modeling methods, as summarized in

Table 1. Recommended Vis-NIR spectroscopy strategies based on fruit morphological characteristics.

Fruit Morphological Feature	Typical Examples	Recommended Optical Mode	Common Preprocessing Methods	Suitable Modeling Strategies	References
Thin rind (<1 mm), small size	Grape, blueberry, cherry	Transmittance or reflectance	SG smoothing, standard normalization, SNV, or MSC to reduce scatter	PLSR, MLR	[78,79,80,81]
Thin-to-medium rind (1–3 mm), medium size	Apple, pear, orange	Transmittance or reflectance	SNV or MSC to reduce scatter, SG + 1st derivative for baseline shift removal	PLSR, SVR	[30,54,73,82]
Thick rind (>3 mm), large size	Watermelon, pomelo, pineapple	Transmittance or interactance	MSC + detrend to reduce scatter and baseline drift, SG smoothing + 2nd derivative to enhance weak absorbance bands	CNN, PLSR, SVR	[24,83,84,85]
Presence of pits/kernels/cavities	Peach, mango, cherry	Interactance or reflectance	SNV to reduce scatter, SG smoothing + 2nd derivative for feature enhancement	PLSR, ANN, RF	[86,87]
Significant non-uniform distribution of quality attribute	Strawberry, melon	Multi-point reflectance or transmittance	Model the local areas, respectively, and perform weighted fusion, MSC, SNV	Local PLSR, CNN	[28,88]

.

4. Applications of Vis/NIR Spectroscopy for Quality Evaluation

The applications of Vis/NIR spectroscopy in fruit quality assessment span a diverse range of fruits and detection objectives, as summarized in Table 2. This table exemplifies its widespread use for evaluating key attributes such as SSC, firmness, internal disorders, and compositional traits across fruits, including citrus, berries, stone fruits, and thick-rind species. To systematically organize these applications, this section adopts a structure based on fruit morphological characteristics—namely, fruits with pits or kernels, delicate fruits with thin rind, and large-sized fruits with thick rind—to discuss specific detection configurations, challenges, and performance outcomes tailored to each category.

4.1. Fruits with Pits or Kernels

4.1.1. Peach

The application of Vis/NIR spectroscopy in peaches focuses primarily on evaluating internal quality attributes such as SSC, firmness, pH, and dry matter content (DMC), all of which are critical for determining maturity, sweetness, and storability. However, the presence of pits introduces specific challenges in spectral measurement due to their interference with light penetration and scattering properties. Studies have demonstrated that spectral acquisition modes, reflectance, interactance, or transmission, must be selected carefully to minimize the impact of the pit on signal quality. Shao et al. utilized the interactance mode of Vis/NIR spectroscopy to predict SSC and pH in peaches, effectively reducing interference from internal structures such as pits through spectral preprocessing techniques including SG smoothing and MSC [89]. Moreover, Sharabiani et al. employed Vis/NIR spectroscopy (350–1150 nm) with PLSR and multiple preprocessing methods (e.g., SG smoothing, MSC, and first derivative) to non-destructively predict quality parameters in Javadi peaches, achieving high accuracy for pH (RMSEP = 0.15, Rp = 0.94) and moderate accuracy for TA (RMSEP = 0.07%, Rp = 0.86) and SSC (RMSEP = 1.17°Brix, Rp = 0.85), demonstrating the effectiveness of chemometric approaches in mitigating pit-related spectral variations [90]. Similarly, Uwadaira et al. emphasized the relevance of Vis/NIR in detecting flesh firmness and water-soluble pectin content, attributes indirectly influenced by pit size and density across varieties [91].

Furthermore, the pit’s composition and size can affect the light path and absorption characteristics, particularly in the NIR region associated with water absorption and chemical bonds (O–H, C–H). Researchers have employed multivariate methods such as PLS and machine learning models to compensate for these structural variations. Yang et al. integrated Vis/NIR with fluorescence spectroscopy and image analysis to classify peach origins, showing that despite pit-related spectral variations, effective wavelength selection and feature fusion could maintain high accuracy [92]. Additionally, Chen et al. combined acoustic sensors with Vis/NIR to improve firmness prediction in yellow flesh peaches, leveraging multi-modal data fusion to mitigate pit-induced inconsistencies [93].

Recent advances also include the use of deep learning approaches, such as multi-task CNN (MT-CNN), which enhances robustness against pit-related noise by extracting common features from multiple quality parameters [86]. These developments highlight that, while the pit poses a challenge, its effects can be systematically addressed through sophisticated instrumentation, algorithmic processing, and data fusion strategies, thereby allowing Vis/NIR spectroscopy to remain a powerful tool for non-destructive quality evaluation in peaches.

4.1.2. Apple

In the application of Vis/NIR spectroscopy for quality evaluation of apples, the presence of pits or kernels introduces specific challenges and considerations, particularly regarding internal quality attributes such as moldy core, watercore, starch content, and SSC [94]. The core and seed cavity regions can influence light penetration and scattering properties, thereby affecting spectral acquisition and interpretation. Studies have demonstrated that the orientation of spectral measurement, such as aligning the stem-calyx axis horizontally and perpendicular to the transmission belt, significantly improves the detection accuracy of moldy core by enhancing signal capture from affected tissues near the core [38]. Additionally, the effectiveness of multivariate models (e.g., PLS, ANN-CA) has been highlighted for estimating internal qualities like firmness, acidity, and starch content using optimally selected wavelengths in the NIR region, minimizing interference from healthy tissues [95,96]. Moreover, the integration of spectral data with acoustic vibration techniques has shown promise in improving classification performance for internal disorders, leveraging complementary physical and biochemical information [34]. For quantitative assessment of watercore and SSC, transmittance spectroscopy combined with variable selection algorithms (e.g., CARS) has proven effective, achieving high predictive accuracy despite structural variations caused by the core [97] (Figure 4A). These approaches underline the importance of optimizing spectral protocols and algorithmic processing to mitigate core-related interferences, thereby enhancing the reliability of non-destructive quality evaluation in apples.

4.1.3. Pear

Vis/NIR spectroscopy has been extensively applied for evaluating internal quality attributes of pear fruit, primarily focusing on SSC, firmness, juiciness, and stone cell content (SCC). SSC and firmness are among the most frequently determined quality parameters [98]. Studies have demonstrated the use of full-spectrum Vis/NIR data combined with PLSR to predict SSC and firmness in varieties such as ‘Fengshui’ pear, yielding promising correlation coefficients of 0.912 and 0.854 for SSC and firmness, respectively [99]. In addition, other nonlinear modeling methods, including least squares SVM (LSSVM) and extreme learning machine (ELM), have also been employed to improve prediction robustness [100].

In addition to SSC and firmness, juiciness, a critical sensory attribute affecting consumer preference, has also been successfully predicted. By combining Vis/NIR with preprocessing techniques such as linear regression correction and spectral ratio (LRC-SR), followed by CARS for feature selection, high-accuracy prediction models were developed with an external validation determination coefficient of 0.93 and root mean square error (RMSE) of 0.97% [101]. Furthermore, the SCC in the Korla fragrant pear, which greatly influences texture, was non-destructively evaluated using effective wavelength selection algorithms such as the successive projections algorithm (SPA) and Monte Carlo uninformative variable elimination (MCUVE), coupled with support vector regression (SVR), achieving high correlation coefficients up to 0.966 for calibration [82]. Recent advances also involve the development of multi-cultivar universal models to enhance generalizability. By integrating interpretable deep learning methods such as Grad-CAM for visualizing significant wavelengths, robust predictive models were established across both green-skinned and brown-skinned pear varieties, significantly reducing the need for building individual models for each cultivar [102]. These studies confirm that Vis/NIR spectroscopy serves as a reliable and efficient tool for non-destructive quality assessment in pears, providing essential technical support for intelligent grading and quality control in the fruit industry.

4.1.4. Mango

The application of Vis/NIR spectroscopy in mango quality evaluation encompasses the prediction of internal attributes, detection of disorders, and adaptability to practical scenarios such as packaged fruits [103]. For internal quality assessment, parameters like firmness, SSC, DMC, and TA have been successfully predicted using Vis/NIR spectroscopy [104]. PLSR models combined with genetic algorithms (GA) selected optimal spectral ranges (e.g., 736–878 nm and 955–1022 nm) for stiffness factor prediction, achieving an R² of 0.82 and RMSEP of 3.28 [104]. Similarly, short-wave NIR (900–1650 nm) paired with variable selection algorithms like SPA improved SSC prediction [87,105]. Beyond quantitative trait prediction, Vis/NIR spectroscopy effectively detects internal physiological disorders such as jelly seed and black flesh. Studies demonstrated that spectral differences in the 550–650 nm range allow discrimination between healthy and disordered fruits, with logistic and LDA models achieving accuracies of 65–76% for disorder prediction at harvest and detection post-storage [106]. Additionally, the technology adapts to challenges like packaging interference; spectral filtering methods (e.g., Gaussian spatial filter) mitigated effects of materials like PVC and PE, maintaining robust prediction performance for firmness, SSC, DMC, and TA (RPD > 2.2) [107]. The integration of multi-omics data further enhances understanding of spectral responses to biochemical changes, such as metabolite variations in anthracnose-infected mangoes [108] (Figure 4B). These applications underscore Vis/NIR spectroscopy’s versatility for non-destructive mango quality monitoring, from harvest to postharvest stages, ensuring accuracy even in complex real-world conditions.

4.1.5. Fresh Jujube

The application of Vis/NIR spectroscopy in the quality evaluation of fresh jujube has been extensively explored, covering internal quality attributes, external defects, varieties, and disease detection [52,109,110]. For internal quality, studies have demonstrated the effectiveness of NIR spectroscopy in predicting vitamin C content and SSC. For instance, a kinetic model based on NIR spectra successfully predicted the degradation of vitamin C in fresh jujube during storage, revealing a zero-order reaction pattern and estimating a shelf life of approximately 15 days at room temperature [111]. Additionally, the detection of SSC in jujubes grown under different cultivation modes (open-field and rain-shelter) was achieved by employing variable selection algorithms such as iteratively retained informative variables (IRIV) and SPA, combined with model updating strategies to enhance robustness and generalizability across cultivation conditions [112]. In a complementary study on Barhi dates, Vis/NIR spectroscopy successfully generated a quality index (Qi) model incorporating both sensory and objective quality parameters. The ANN model demonstrated superior performance over PLSR, achieving higher correlation coefficients (R² = 0.912 vs. 0.793) and lower errors (RMSEC = 0.308 vs. 0.110) in calibration [113].

In terms of external defect identification, hyperspectral imaging within the Vis/NIR range has proven highly effective [114,115]. Research has shown that combining Vis/NIR hyperspectral imaging with image processing algorithms can accurately identify crack features on fresh jujube surfaces. Characteristic wavelengths, such as 467 nm, 544 nm, 639 nm, 673 nm, and 682 nm, were identified as crucial for distinguishing cracked from sound jujubes, achieving discrimination accuracies as high as 100% in optimized models [116]. This approach not only facilitates qualitative discrimination but also enables quantitative analysis of crack location and area. Moreover, Vis/NIR HSI has been applied to monitor disease progression in jujubes, such as black spot disease caused by Alternaria alternata. The spectral variations in the Vis/NIR region (400–1000 nm) were more sensitive to disease-related changes compared to the short-wave infrared region (1000–2000 nm), allowing for effective discrimination of infected tissues and visualization of disease development over time [117]. These studies collectively highlight the versatility and reliability of Vis/NIR spectroscopy and imaging techniques for comprehensive quality assessment of fresh jujube, providing valuable tools for non-destructive monitoring in postharvest management.

4.2. Delicate Fruits with Thin Rind

4.2.1. Blueberry

Visible/near-infrared (Vis/NIR) spectroscopy has proven particularly valuable for assessing blueberries, fruits highly vulnerable to mechanical damage and spoilage owing to their delicate skin and soft texture. This technology enables rapid and accurate assessment of both internal quality attributes and external defects without compromising fruit integrity [118]. Vis/NIR techniques have been successfully employed to predict key maturity parameters such as total soluble solids (TSS) and acidity, as well as bioactive compounds including anthocyanins, total phenolics, and ascorbic acid [119]. Moreover, spectral models developed using PLSR have demonstrated high accuracy in quantifying these parameters, with correlation coefficients often exceeding 0.90 in both calibration and prediction sets [120]. Beyond compositional analysis, Vis/NIR hyperspectral imaging has shown great potential in detecting early disease symptoms that are invisible to the human eye, such as fungal infections, by capturing subtle spectral changes associated with cellular damage and pathogen presence [80]. Additionally, recent studies have explored the use of portable Vis/NIR devices like the cherry-meter, which calculates an index of absorbance difference (IAD) to monitor anthocyanin accumulation and fruit maturity in real-time [120]. Such tools are particularly valuable for determining optimal harvest timing and ensuring postharvest quality. Furthermore, Vis/NIR spectroscopy has been integrated into fermentation processes to monitor microbial metabolites such as exopolysaccharides in blueberry juice, providing a non-destructive means to optimize fermentation conditions and enhance product quality [118]. These applications highlight the versatility and effectiveness of Vis/NIR spectroscopy as a non-destructive tool for comprehensive quality evaluation throughout the blueberry supply chain from harvest and postharvest storage to processing.

4.2.2. Strawberry

In the case of strawberries, a classic example of thin-rinded, delicate produce, Vis/NIR spectroscopy has emerged as a key method for non-destructive assessment of SSC, textural properties, and nutritional components. Guo et al. developed an online detection prototype and found that transmittance spectra combined with the CARS algorithm yielded optimal prediction performance for SSC (Rp = 0.928, RMSEP = 0.412°Brix, RPD = 2.670) [121]. Similarly, Seki et al. confirmed that both Vis/NIR and NIR could effectively predict SSC in white and red strawberries with comparable accuracy (R²p = 0.89 and 0.85, respectively), highlighting the technique’s robustness across cultivars [28]. Rabbani et al. utilized a non-contact silicon-based Vis-NIR system with PLS regression and SG filtering for optical modeling, demonstrating high prediction accuracy for texture qualities like firmness (R = 0.81, RMSEP = 0.41 N in transmittance mode) and brittleness, validated through chemometric analysis. Besides SSC, texture qualities such as firmness and brittleness have also been reliably predicted using silicon-based Vis/NIR sensors, with transmittance mode again showing superior results [122]. Furthermore, Rabasco-Vilchez et al. introduced a reflectance quality index based on portable Vis-NIR spectroscopy, employing PLS models for shelf-life prediction, with validation metrics showing R² up to 0.95 for quality attributes, underscoring the effectiveness of chemometrics in optical modeling [123]. NIR spectroscopy has been employed in classifying strawberries based on nutritional quality. Mancini et al. utilized Fourier transform -NIR combined with PLS-DA to distinguish different cultivars according to vitamin C, anthocyanins, and phenolic acids content, demonstrating the potential of spectral data fusion for rapid phenotyping in breeding programs [124]. As shown in Figure 4C, Qiao et al. developed a hanging transportation system with Vis/NIR spectroscopy for online SSC detection in strawberries, achieving high accuracy using a 1D-CNN-LSTM model (R²p = 0.963, RMSEP = 0.209°Brix, RPD = 5.332) [88]. These applications underscore the versatility of Vis/NIR spectroscopy for nondestructive, real-time quality monitoring in strawberries, contributing to postharvest management and cultivar selection.

4.2.3. Cherry Tomato

Cherry tomato, being a representative delicate fruit with a thin rind, has been extensively studied using Vis/NIR spectroscopy for non-destructive quality assessment. This technique enables rapid and accurate evaluation of key internal quality attributes such as SSC, TA, firmness, and lycopene content [125]. Studies have demonstrated that combining spectral information from multiple bands, such as Vis/NIR and NIR, through data fusion strategies significantly enhances prediction accuracy compared to using a single spectral range [126]. For instance, fusion of Vis/NIR (400–1004 nm) and NIR (1002–1652 nm) spectra improved the predictive performance for SSC and TA with R² values above 0.9 in optimal models. Similarly, Egei et al. (2022) employed Vis/NIR and short-wave IR spectroscopy with PLSR modeling and preprocessing methods like MSC and first derivatives, achieving robust prediction of SSC (R² = 0.72, RMSECV = 0.59°Brix for intact fruits) and lycopene content (R² = 0.68, RMSECV = 15.07 mg kg⁻¹ for homogenates), validating the efficacy of multi-band data fusion in chemometrics [127]. Furthermore, the use of machine learning algorithms, including PLSR, SVR, and BPNN, has been widely adopted for modeling spectral data, with BPNN and SVR often yielding superior results for nonlinear relationships. Brito et al. utilized a portable Vis/NIR spectrometer (F-750) combined with PCA and PLSR for optical modeling, demonstrating accurate non-destructive determination of color parameter a (R²p = 0.94, RMSEP = 2.89) and dry matter (R²p = 0.59, RMSEP = 0.46%), though TA prediction was less reliable, highlighting the practical application of chemometrics in field-based quality assessment [128].

Besides multi-parameter detection, recent research has also focused on specific bioactive components such as lycopene. By integrating spectral data with image features, such as color and texture indicators, through low-level data fusion, prediction models for lycopene achieved an R² value of 0.95 and an RPD of 4.25, demonstrating the effectiveness of supplementing spectral information with visual characteristics [129]. In a study on tomato lycopene detection, multi-point full transmission Vis/NIR spectroscopy combined with weighted average processing and LARS-L1 wavelength selection yielded excellent prediction results [130] (Figure 4D). The PLSR models achieved Rp of 0.96 and RMSEP of 13.44 mg kg⁻¹ for ‘Provence’ tomatoes, and Rp of 0.95 and RMSEP of 7.43 mg kg⁻¹ for ‘Jingcai No. 8’ tomatoes, demonstrating the robustness of the approach for different cultivars. Moreover, the selection of light sources plays a critical role in capturing characteristic absorption bands; xenon lamps, which offer broader spectral coverage including ultraviolet regions, outperform traditional halogen lamps in detecting lycopene and SSC due to better matching with molecular overtone and combination vibrations [81]. In terms of spectral acquisition modes, transmittance spectroscopy generally outperforms reflectance for estimating internal qualities like SSC, as it captures more representative internal information [121]. However, combining both modes can compensate for their individual limitations, providing a more comprehensive analysis. Overall, Vis/NIR spectroscopy has proven to be a powerful tool for non-destructive quality monitoring in cherry tomatoes, showing great potential for application in automated grading and sorting systems.

Table 2. Applications of Vis/NIR spectroscopy in nondestructive quality detection of fruits.

Fruits	Detection Indexes	Data Acquisition Modes	Spectral Range	Models	Validation Approach	Performance	References
Orange	Freezing damage	Online transmission	644–900 nm	DCM-1D-CNN	Internal validation, cross-validation	Accuracy = 91.96%	[74]
Orange	Sweetness classification	Interactance	600–1050 nm	Ensemble classifier	Internal validation, cross-validation	Accuracy = 81.03%	[105]
Fig	Firmness, SSC	Diffuse reflectance	545–1175 nm	SPA-RF	Internal validation, cross-validation	Firmness R2p = 0.9173, RMSEP = 19.9027, RPD = 2.24	[37]
Strawberry	Cultivar classification	Integrating sphere	1000–2500 nm	PLS-DA	Internal validation, cross-validation	Successful discrimination	[124]
Strawberry	SSC	Online transmission	650–980 nm	1D-CNN-LSTM	Internal validation, cross-validation	R2p = 0.963, RMSEP = 0.209°Brix, RPD = 5.332	[88]
White strawberry, Red strawberry	SSC	Reflectance	500–978 nm, 908–1676 nm	PLSR	Internal validation, cross-validation	White: R2p = 0.85–0.89, RMSEP = 0.40–0.43%, RPD = 2.64–2.98 Red: R2p = 0.89, RMSEP = 0.36%, RPD = 3.04–3.05	[28]
Lime	TA	Integrating sphere	833–2500 nm, 898–1720 nm	PLS, DT, XGBoost, FFNN	Internal validation, cross-validation	R2p = 0.66, RMSEP = 0.3896, RPD = 1.33	[36]
Cherry tomato	Lycopene	Reflectance, transmittance	200–1100 nm	PLSR	Internal validation, cross-validation	R2p = 0.91, RMSEP = 11.60 mg/kg, RPD = 3.28	[129]
Navel orange	SSC	Online transmittance	650–950 nm	PLSR	Internal validation, cross-validation	R2p = 0.9406, RMSEP = 0.442%, RPD = 2.77	[55]
Nanguo pears	SSC	Portable reflectance	900–1700 nm	Si-GA-PLS	Internal validation, cross-validation	R2p = 0.9406, RMSEP = 0.1655°Brix	[98]
Pears	Cork spot disorder	Portable reflectance	900–1700 nm	SVM	Internal validation, cross-validation	Accuracy = 84.65%	[76]
Sugar orange	Granulation	Online transmittance	400–1200 nm	PLS-DA	Internal validation, cross-validation	Accuracy = 94.00%, Class error = 5.84%	[73]
Pear	Sunburn severity	Reflectance	400–1100 nm	iPLS, LDA	Internal validation, cross-validation	Accuracy = 83%	[69]
Mango	SSC	Reflectance	900–1650 nm	SPA-PLSR	Internal validation, cross-validation	Rp = 0.78, SEP = 0.67°Brix, RPD = 2.12	[87]
Mango	Anthracnose disease	Reflectance	450–980 nm	LDA, QDA, PLS-DA	Internal validation	Accuracy = 90.9%, sensitivity = 0.929, specificity = 0.989	[103]
Kiwifruit	Sweetness, firmness	Reflectance	800, 810, 850, 880, 900, 940, 970, 1000, 1100 nm	SVM	Cross-validation, independent external validation	Sweetness accuracy = 82.0%, firmness accuracy = 74.0%	[50]
Apple	Moldy core	Online transmittance	650–1000 nm	PLS-DA	Internal validation, independent external validation	Accuracy = 94.44%, recall = 92.59%, precision = 96.15%	[54]
Apple	SSC, firmness, pH, watercore degree	Transmittance	700–1100 nm	CARS-CNN	Internal validation, independent external validation	Rp: 0.951 (SSC), 0.824 (firmness), 0.828 (pH), 0.943 (watercore)	[16]
Lemon	TSS, TA	Reflectance	950–1700 nm	PLSR	Internal validation, independent external validation	R2p: 0.84 (TSS), 0.72 (TA); RMSEP: 0.42% (TSS), 0.45 g/100 mL (TA)	[66]
Avocado	DMC	Diffuse reflectance, interaction	350–2500 nm 310–1135 nm	PLSR	Internal validation, cross-validation	RMSECV: 1.02% dw/fw (non-dehydrated), 1.49% dw/fw (dehydrated)	[19]
Pomegranate	Juice percentage, TSS, TA, Taste, pH, vitamin C	Interactance	400–1000 nm	PLSR	Internal validation	Rp = 0.95–0.98, RMSEP = 0.036–0.583	[33]
Tomato	Lycopene	Transmission	560–1072 nm	PLSR	Internal validation, cross-validation	Rp = 0.95–0.96, RMSEP = 7.43–13.44 mg kg−1	[130]
Banana	SSC, ripeness	Reflectance	610, 680, 730, 760, 810, 860 nm	MLR	Internal validation	SSC: R2p = 0.9915, RMSEP = 0.38%, Ripeness: Avg. accuracy 97%	[75]
Apricot	TSS, TA, DMC	Reflectance	310–1100 nm	ANN-MLP	Internal validation, independent external validation	R2: 0.855 (TSS), 0.681 (TA), 0.857 (DMC)	[17]
Pomelo	SSC	Transmission	400–1100 nm	SNV-CARS-PLSR	Internal validation	R2c = 0.98, RMSEC = 0.46, R2v = 0.89, RMSEV = 0.87	[15]
Barhi dates	TSS, hardness	Reflectance	285–1200 nm	ANN	Internal validation, cross-validation	R2c = 0.912, RMSEC = 0.308, RMSECV = 0.308	[113]

DCM-1D-CNN: Diameter Correction Method and one-dimensional Convolutional Neural Network; SPA-RF: Successive Projections Algorithm and Random Forest.

Table 2. Applications of Vis/NIR spectroscopy in nondestructive quality detection of fruits.

Fruits	Detection Indexes	Data Acquisition Modes	Spectral Range	Models	Validation Approach	Performance	References
Orange	Freezing damage	Online transmission	644–900 nm	DCM-1D-CNN	Internal validation, cross-validation	Accuracy = 91.96%	[74]
Orange	Sweetness classification	Interactance	600–1050 nm	Ensemble classifier	Internal validation, cross-validation	Accuracy = 81.03%	[105]
Fig	Firmness, SSC	Diffuse reflectance	545–1175 nm	SPA-RF	Internal validation, cross-validation	Firmness R2p = 0.9173, RMSEP = 19.9027, RPD = 2.24	[37]
Strawberry	Cultivar classification	Integrating sphere	1000–2500 nm	PLS-DA	Internal validation, cross-validation	Successful discrimination	[124]
Strawberry	SSC	Online transmission	650–980 nm	1D-CNN-LSTM	Internal validation, cross-validation	R2p = 0.963, RMSEP = 0.209°Brix, RPD = 5.332	[88]
White strawberry, Red strawberry	SSC	Reflectance	500–978 nm, 908–1676 nm	PLSR	Internal validation, cross-validation	White: R2p = 0.85–0.89, RMSEP = 0.40–0.43%, RPD = 2.64–2.98 Red: R2p = 0.89, RMSEP = 0.36%, RPD = 3.04–3.05	[28]
Lime	TA	Integrating sphere	833–2500 nm, 898–1720 nm	PLS, DT, XGBoost, FFNN	Internal validation, cross-validation	R2p = 0.66, RMSEP = 0.3896, RPD = 1.33	[36]
Cherry tomato	Lycopene	Reflectance, transmittance	200–1100 nm	PLSR	Internal validation, cross-validation	R2p = 0.91, RMSEP = 11.60 mg/kg, RPD = 3.28	[129]
Navel orange	SSC	Online transmittance	650–950 nm	PLSR	Internal validation, cross-validation	R2p = 0.9406, RMSEP = 0.442%, RPD = 2.77	[55]
Nanguo pears	SSC	Portable reflectance	900–1700 nm	Si-GA-PLS	Internal validation, cross-validation	R2p = 0.9406, RMSEP = 0.1655°Brix	[98]
Pears	Cork spot disorder	Portable reflectance	900–1700 nm	SVM	Internal validation, cross-validation	Accuracy = 84.65%	[76]
Sugar orange	Granulation	Online transmittance	400–1200 nm	PLS-DA	Internal validation, cross-validation	Accuracy = 94.00%, Class error = 5.84%	[73]
Pear	Sunburn severity	Reflectance	400–1100 nm	iPLS, LDA	Internal validation, cross-validation	Accuracy = 83%	[69]
Mango	SSC	Reflectance	900–1650 nm	SPA-PLSR	Internal validation, cross-validation	Rp = 0.78, SEP = 0.67°Brix, RPD = 2.12	[87]
Mango	Anthracnose disease	Reflectance	450–980 nm	LDA, QDA, PLS-DA	Internal validation	Accuracy = 90.9%, sensitivity = 0.929, specificity = 0.989	[103]
Kiwifruit	Sweetness, firmness	Reflectance	800, 810, 850, 880, 900, 940, 970, 1000, 1100 nm	SVM	Cross-validation, independent external validation	Sweetness accuracy = 82.0%, firmness accuracy = 74.0%	[50]
Apple	Moldy core	Online transmittance	650–1000 nm	PLS-DA	Internal validation, independent external validation	Accuracy = 94.44%, recall = 92.59%, precision = 96.15%	[54]
Apple	SSC, firmness, pH, watercore degree	Transmittance	700–1100 nm	CARS-CNN	Internal validation, independent external validation	Rp: 0.951 (SSC), 0.824 (firmness), 0.828 (pH), 0.943 (watercore)	[16]
Lemon	TSS, TA	Reflectance	950–1700 nm	PLSR	Internal validation, independent external validation	R2p: 0.84 (TSS), 0.72 (TA); RMSEP: 0.42% (TSS), 0.45 g/100 mL (TA)	[66]
Avocado	DMC	Diffuse reflectance, interaction	350–2500 nm 310–1135 nm	PLSR	Internal validation, cross-validation	RMSECV: 1.02% dw/fw (non-dehydrated), 1.49% dw/fw (dehydrated)	[19]
Pomegranate	Juice percentage, TSS, TA, Taste, pH, vitamin C	Interactance	400–1000 nm	PLSR	Internal validation	Rp = 0.95–0.98, RMSEP = 0.036–0.583	[33]
Tomato	Lycopene	Transmission	560–1072 nm	PLSR	Internal validation, cross-validation	Rp = 0.95–0.96, RMSEP = 7.43–13.44 mg kg−1	[130]
Banana	SSC, ripeness	Reflectance	610, 680, 730, 760, 810, 860 nm	MLR	Internal validation	SSC: R2p = 0.9915, RMSEP = 0.38%, Ripeness: Avg. accuracy 97%	[75]
Apricot	TSS, TA, DMC	Reflectance	310–1100 nm	ANN-MLP	Internal validation, independent external validation	R2: 0.855 (TSS), 0.681 (TA), 0.857 (DMC)	[17]
Pomelo	SSC	Transmission	400–1100 nm	SNV-CARS-PLSR	Internal validation	R2c = 0.98, RMSEC = 0.46, R2v = 0.89, RMSEV = 0.87	[15]
Barhi dates	TSS, hardness	Reflectance	285–1200 nm	ANN	Internal validation, cross-validation	R2c = 0.912, RMSEC = 0.308, RMSECV = 0.308	[113]

DCM-1D-CNN: Diameter Correction Method and one-dimensional Convolutional Neural Network; SPA-RF: Successive Projections Algorithm and Random Forest.

4.2.4. Grape

Owing to its thin rind and high moisture content, the grape is prone to mechanical injury and quality loss during post-harvest handling, making it a prime candidate for quality monitoring via Vis/NIR spectroscopy. Vis/NIR spectroscopy has emerged as an effective non-destructive tool for evaluating both internal and external quality attributes of grapes. This technique is particularly advantageous for grapes due to their delicate skin, which makes traditional destructive methods less feasible for continuous quality monitoring. One significant application of Vis/NIR spectroscopy in grapes is the detection of fungal infections and phytosanitary status. Studies have demonstrated the use of Vis/NIR spectroscopy coupled with immersion probes to assess must infection levels at grape receiving areas, achieving classification accuracies of up to 90.4% for healthy and infected samples using PLS-DA [131]. This approach provides wineries with a rapid and objective method to quantify infection levels, thereby supporting decisions regarding grape destination and wine quality management. Beyond disease detection, Vis/NIR spectroscopy is also employed for grading and sorting grapes based on multiple internal quality parameters, such as SSC and total phenolic compounds. By combining spectral data with multivariate analysis techniques like PLS-DA, researchers have successfully classified stored grape berries into different quality categories with accuracies exceeding 77% [132]. This non-destructive method allows for efficient quality control without compromising the integrity of the berries, making it suitable for industrial applications.

Moreover, Vis/NIR spectroscopy has been utilized to monitor grape ripening stages. By analyzing spectral data in the range of 400–1100 nm, studies have developed models to discriminate between different maturity stages based on SSC, TA, and phenolic content [79]. The integration of advanced algorithms, such as stacked autoencoders (SAE) and one-dimensional CNN (1D-CNN), has further improved the accuracy of ripeness prediction, with some models achieving up to 94% accuracy in prediction sets. This capability is crucial for determining the optimal harvest time and ensuring consistent grape quality. Additionally, Vis/NIR spectroscopy enables the prediction of key quality attributes, such as SSC and TA, during grape ripening. Predictive models based on PLSR have shown high performance, with R² reaching 0.97 for calibration and 0.94 for prediction [78]. Furthermore, a recent study incorporating Vis/NIR spectroscopy (400–1100 nm) with newly developed spectral reflectance indices and machine learning models demonstrated robust non-destructive prediction of anthocyanin, TSS, TA, and TSS/TA ratio in grapes. The decision tree (DT) model achieved high accuracy for anthocyanin (cross-validation R² = 0.87, RMSE = 87.81 mg L⁻¹) and TSS/TA ratio (R² = 0.74, RMSE = 3.12), while the gradient boosting regression model excelled in predicting TSS (R² = 0.82, RMSE = 0.92%) and TA (R² = 0.70, RMSE = 0.05%). The three-band spectral reflectance indices outperformed traditional indices, with R² values up to 0.88 for anthocyanin, highlighting the effectiveness of chemometric approaches in comprehensive grape quality assessment [133]. These models facilitate real-time quality assessment, providing a non-destructive alternative to conventional chemical analyses. Vis/NIR spectroscopy offers a comprehensive solution for the non-destructive evaluation of grapes, encompassing infection detection, quality grading, ripeness monitoring, and quantitative prediction of internal attributes. Its ability to deliver rapid and accurate results makes it an invaluable tool for enhancing grape quality management throughout the supply chain.

4.2.5. Mulberry

The mulberry (Morus sp.) is recognized for its high productivity during the fruiting season, alongside a long-standing tradition of consumption as both a nutritious food and a source of medicinal compounds [134]. The application of Vis/NIR spectroscopy for mulberry quality evaluation effectively addresses challenges posed by its delicate structure and bumpy surface morphology. Studies demonstrate its capability for non-destructive measurement of internal physicochemical properties. Huang et al. pioneered the use of Vis/NIR spectroscopy (325–1075 nm) combined with chemometrics (PLS, LS-SVM, Multiple linear regression (MLR)) to predict SSC and pH in intact mulberries. By employing the SPA, optimal wavelengths (431 nm and 976 nm for SSC; 627 nm and 696 nm for pH) were identified, achieving R²p of 0.70 and 0.90, respectively. This validated the technique’s robustness despite surface irregularities [135].

Subsequent research expanded to bioactive compounds. Huang et al. utilized hyperspectral imaging (380–1734 nm) with Si and InGaAs detectors to quantify total anthocyanin content (TAC) and antioxidant activity. CARS selected feature wavelengths, and LS-SVM modeling yielded high accuracy (R²v = 0.995 for TAC, RPD = 14.255), highlighting the superiority of spectral imaging for spatial compound distribution analysis [136]. For field applications, Yan et al. validated handheld NIR devices (908–1676 nm) for rapid assessment of SSC, dry matter, polyphenols, and flavonoids. GA-optimized PLS models achieved R²p > 0.93 for SSC and dry matter, demonstrating portability without compromising precision [137]. Additionally, Soltanikazemi et al. applied UV-IR spectroscopy (300–1100 nm) to mulberry juice, where GA-PLS significantly enhanced the prediction of ascorbic acid (Rv = 0.98) and TA (Rv = 0.96) by selecting critical spectral intervals [138].

4.3. Large-Sized Fruits with Thick Rind

4.3.1. Pomelo

The application of Vis/NIR spectroscopy in pomelo quality detection faces unique challenges due to its large size and thick rind, which attenuate light penetration and introduce significant spectral noise. Nevertheless, studies have demonstrated the potential of Vis/NIR techniques for evaluating both external and internal quality attributes. For instance, maturity can be assessed by quantifying surface color indices (L, a, b*) using diffuse reflectance spectroscopy combined with moving window PLS (MWPLS) and modified optical path length estimation and correction (MOPLEC), achieving high correlation coefficients (R > 0.91) between predicted and measured values [139].

For internal quality, water content and granulation—a physiological disorder affecting texture—have been successfully quantified using transmission-mode Vis/NIR spectroscopy. Preprocessing techniques such as SG smoothing and MSC, coupled with genetic algorithm-based wavelength selection and PLS modeling, yielded an R² of 0.71 for water content and 100% classification accuracy for granulation degree [140]. Furthermore, SSC and acidity, key flavor indicators, have been non-destructively predicted using optimized spectral systems and CNN, though with varying accuracy (R² = 0.72 for SSC and 0.55 for acidity) [141]. Notably, the thick rind significantly affects detection performance. Studies comparing intact and peeled pomelos revealed that the peel introduces optical interference, reducing the predictive accuracy of PLS models for SSC. By contrast, models developed using peeled samples achieved higher robustness (RPD = 2.57) and lower error [24] (Figure 4E). These findings highlight the importance of developing specialized optical configurations and advanced algorithms to mitigate scattering and absorption effects caused by the rind.

4.3.2. Watermelon

Vis/NIR spectroscopy has been extensively applied to evaluate watermelon quality non-destructively, focusing primarily on SSC as a key maturity indicator. The technology leverages spectral signatures between 400 and 1700 nm to correlate with internal attributes, though environmental factors like temperature variations significantly impact accuracy. Yao et al. demonstrated that temperature fluctuations (0–40 °C) induce spectral shifts in watermelon juice, particularly in the 1200–1698 nm range, degrading PLSR model performance [142]. Their global PLSR model achieved an RMSEP of 0.480°Brix, outperforming temperature-specific local models, highlighting the necessity for robust temperature compensation. Spatial heterogeneity of SSC within large watermelons presents another challenge. Francis et al. addressed this using hyperspectral imaging (400–1000 nm) and a low-rank deep learning framework [143]. By capturing spectral data from multiple viewpoints (top, bottom, laterals), they mapped SSC variability across flesh regions, achieving superior accuracy (RMSEP = 0.195°Brix) compared to traditional PLSR (RMSEP = 0.730°Brix). This approach underscores hyperspectral imaging’s capability to resolve internal quality gradients often missed by single-point spectroscopy.

For cultivar comparisons, Ibrahim et al. evaluated three watermelon types using Vis/NIR (475–1075 nm) and NIR (950–1650 nm) systems [144]. NIR outperformed Vis/NIR in predicting SSC (R²p = 0.85), lycopene (R²p = 0.92), and vitamin C (R²p = 0.90), attributed to stronger absorption bands for O-H (water/sugars) and C-H (carotenoids) bonds. This emphasizes the role of wavelength range selection in optimizing detection for specific compounds. Recent advancements focus on AI-driven temperature correction. Sun et al. integrated 1D-CNN with gradient-weighted activation mapping (Grad-CAM) to identify temperature-sensitive spectral bands [84]. By suppressing these bands via arithmetic mapping, their knowledge-guided PLSR model reduced RMSEP to 0.324°Brix, which was 32.5% lower than conventional global models. This method surpassed traditional slope/bias correction and external parameter orthogonalization in robustness.

4.3.3. Hami Melon

The application of Vis/NIR spectroscopy for quality evaluation in Hami melon focuses primarily on two critical parameters: SSC and pesticide residue detection, addressing challenges posed by its thick rind (up to 20 mm), large size, and nonuniform internal quality distribution [145,146]. For the SSC assessment, the spectral acquisition position significantly influences prediction accuracy due to spatial heterogeneity in sugar accumulation. Studies optimized detection at the calyx region, where peel thickness is minimal, yielding superior results (Rp = 0.93, RMSEP = 0.84°Brix) compared to equatorial or stem positions. This optimization leverages thinner rind penetration depth (~4–20 mm), enhancing signal-to-noise ratios for flesh SSC correlations [145]. Variable selection algorithms like MC-UVE-SPA further refine models by identifying 18 key wavelengths, enabling multispectral PLS models (RMSEP = 0.95°Brix) with minimal computational burden [146].

Pesticide residue detection employs 1D-CNN to overcome spectral interference from thick rinds. Asymmetric multiscale architectures achieve 93.68% accuracy for lambda-cyhalothrin and 95.79% for beta-cypermethrin in four-level residue classification, utilizing derivative preprocessing to isolate residue-specific features [147]. Deep feature fusion strategies further enhance type discrimination to 95.83% accuracy [148]. Commercial systems like IAS-FT6000 integrate these advances for online sorting, combining Vis/NIR (650–950 nm) with conveyor-based spectral averaging to achieve 95% SSC and 97% defect detection accuracy at speeds of 8–10 fruits/s [83]. Future efforts should prioritize multi-sensor fusion and miniaturized deep learning models to enhance robustness across diverse cultivars and field conditions.

4.3.4. Pineapple

Vis/NIR spectroscopy has emerged as a powerful and versatile tool for the non-destructive evaluation of key quality attributes in pineapple, a large-sized tropical fruit crucial to global trade. Its thick, rough rind poses challenges for light penetration and signal acquisition, yet optimized techniques have shown significant promise across several application domains. One primary application is the detection of internal physiological disorders. Flesh translucency (PFT), a major postharvest issue causing water-soaked flesh and reduced quality, has been successfully identified using Vis/NIR transmission spectroscopy. Key wavelengths identified for translucency detection were concentrated in the visible to shortwave NIR region (400–900 nm), particularly around 710–870 nm and 930–1050 nm, correlating with changes in water content, flesh density, and color associated with the disorder. Spectral preprocessing techniques like SG smoothing and SNV transformation are essential to mitigate noise from the rough surface and enhance classification accuracy. Probabilistic neural network (PNN) models outperformed PLSR in classifying translucency severity (no, slight, heavy) due to their ability to handle nonlinear relationships, achieving high validation accuracies, especially when employing data supplementation for model updating across different harvest batches [149,150].

Quantitative assessment of critical internal chemical constituents is another major focus. SSC, a direct indicator of sweetness and consumer preference, has been predicted non-destructively in intact pineapples using both reflectance and transmittance modes within the Vis/NIR range. PLSR is commonly employed, often coupled with spectral preprocessing. Prediction performance varies based on cultivar, spectral range, and model complexity, typically yielding R²p ranging from 0.72 to 0.88 and RMSEP values around 0.79 to 1.04°Brix [81,151,152]. Nitrate levels, critical for canned pineapple quality as high concentrations cause black stains, have also been quantitatively predicted using Vis/NIR interactance spectroscopy. Optimal models using first-derivative pretreated average spectra in the 600–1200 nm range achieved high correlation (R = 0.95) and low prediction error (RMSEP = 1.77 ppm) with PLSR, enabling rough screening [153].

Maturity assessment, vital for determining optimal harvest and processing timing, has been effectively tackled. Studies utilize Vis/NIR spectra to classify maturity stages (e.g., immature, mature, overmature), often linked to translucency development or SSC levels. PLS-DA models on transmittance spectra achieved high classification accuracies (>90%) for maturity grades [85]. In a comprehensive study utilizing Vis/NIR transmittance spectroscopy coupled with machine learning methodologies, researchers developed a robust maturity detection system for pineapples. The PLSDA model achieved a high classification accuracy of 90.8% for discriminating between immature, mature, and overmature grades. Additionally, the ANN-PLS quantitative model for determining SSC demonstrated strong performance with an R² of 0.7596 and RMSEP of 0.7879°Brix [149] (Figure 4F). Furthermore, the maturity index (TSS/TA ratio) itself has been predicted directly. Both transmittance short-wave NIR spectroscopy (665–955 nm) and reflectance NIR hyperspectral imaging (935–1720 nm) demonstrated reliable performance (R²c = 0.70–0.72, RMSECV = 1.68–2.16) using PLSR models optimized with specific preprocessing [151].

Beyond the fruit itself, Vis/NIR spectroscopy facilitates differentiation between organic and inorganic pineapple juices. Dual handheld NIR spectrometers (SCiO: 740–1070 nm; Tellspec: 900–1700 nm) combined with PLS-DA analysis successfully classified juice types. Notably, fusion of data from both spectrometers without preprocessing yielded perfect classification (100% accuracy). Key discriminatory wavelengths identified via variable importance in projection (VIP) scores corresponded to regions associated with sugars, organic acids, and potentially polyphenols or other compositional markers influenced by cultivation practices [152]. Vis/NIR spectroscopy offers a comprehensive suite of non-destructive solutions for pineapple quality evaluation. Overcoming challenges posed by its size and rind texture, this technology enables reliable detection of internal defects like translucency, quantification of essential chemical components (SSC, nitrate), accurate maturity assessment, and even authentication of processed products like juice. Continued refinement in spectral acquisition modes (transmittance, reflectance, interactance), wavelength range selection, preprocessing algorithms, and robust modeling techniques (PLSR, PNN, PLS-DA, SVM) ensures its growing applicability in pineapple postharvest management and processing for enhanced quality control and reduced waste.

Figure 4. Applications of Vis/NIR spectroscopy for fruit quality evaluation. (A) Multi-algorithm enhanced SSC and watercore detection in apples employing variable selection methods (SI, GA, CARS, SPA) for optimized PLS modeling [97]. (B) Mechanism investigation of mango anthracnose via Vis/NIR combined with multi-omics under natural and inoculated conditions [108]. (C) Online SSC monitoring in strawberries using a Vis/NIR system integrated with a non-destructive hanging grasper [88]. (D) Multipoint transmission spectroscopy for lycopene quantification in tomato cultivars using Vis/NIR spectroscopy [130]. (E) Peel interference compensation in pomelo SSC evaluation through Vis/NIR spectroscopy and chemometric correction [24]. (F) Machine learning-driven maturity prediction in pineapples based on transmittance Vis/NIR spectral analysis [149].

Figure 4. Applications of Vis/NIR spectroscopy for fruit quality evaluation. (A) Multi-algorithm enhanced SSC and watercore detection in apples employing variable selection methods (SI, GA, CARS, SPA) for optimized PLS modeling [97]. (B) Mechanism investigation of mango anthracnose via Vis/NIR combined with multi-omics under natural and inoculated conditions [108]. (C) Online SSC monitoring in strawberries using a Vis/NIR system integrated with a non-destructive hanging grasper [88]. (D) Multipoint transmission spectroscopy for lycopene quantification in tomato cultivars using Vis/NIR spectroscopy [130]. (E) Peel interference compensation in pomelo SSC evaluation through Vis/NIR spectroscopy and chemometric correction [24]. (F) Machine learning-driven maturity prediction in pineapples based on transmittance Vis/NIR spectral analysis [149].

5. Critical Considerations in Model Development and Validation

While the preceding applications demonstrate the considerable potential of Vis/NIR spectroscopy, a critical gap often lies in the methodological rigor of model validation. Translating these research findings into reliable, deployable systems necessitates a hierarchical framework for model development and assessment.

The reliability and real-world applicability of Vis/NIR spectroscopy models are contingent upon a hierarchical and critical validation protocol. The foundation of this protocol is the clear distinction between internal and external validation. While internal validation (e.g., cross-validation) is indispensable for model tuning and diagnosing overfitting, it often yields optimistically biased performance estimates. The gold standard for assessing true predictive power is a strictly independent external validation set—sourced from different growing seasons, geographical origins, or harvest batches [54]. This provides an unbiased assessment of the model’s generalizability to new, unseen data.

Ensuring the integrity of the validation process requires vigilant safeguards against data leakage and overfitting. A strict data-splitting workflow must be followed, whereby all preprocessing steps are calibrated exclusively on the training data before being applied to the validation and test sets. Concurrently, overfitting must be proactively managed. In methods like PLSR, this involves justifying the number of Latent Variables (LVs), typically by selecting the point where the cross-validation error is minimized. Furthermore, to statistically validate the significance of the model and rule out the possibility of a spurious correlation, permutation tests are highly recommended. This procedure involves repeatedly breaking the relationship between the spectra and the reference values by random shuffling, thereby establishing a null distribution of model performance against which the actual model’s performance can be compared.

Beyond the validation strategy, the interpretation and reporting of performance metrics require a nuanced understanding. Common statistics like R² and RMSEP should always be contextualized. Special attention is warranted for the RPD, whose thresholds are not absolute; their diagnostic power is highly dependent on the distribution and heterogeneity of the reference data population. Finally, the model stability is fundamentally linked to the sample size. Generally, the number of samples must be sufficient to account for the complexity of the measured attribute in order to build robust calibrations. To further quantify the uncertainty in performance estimates, reporting confidence intervals for key metrics is necessary, thus providing a more statistically sound representation of the model’s predictive capability.

6. Conclusions and Perspectives

This review has consolidated key evidence demonstrating that Vis/NIR spectroscopy is a robust non-destructive tool for predicting critical biochemical parameters in diverse characteristic fruits. The technology effectively quantifies SSC, firmness, TA, and defect indicators of fruit types, including thin-rinded berries and thick-rinded fruits with internal pits. Deployments in portable, online, and vehicle-mounted configurations have proven adaptable to various operational scenarios, underscoring their versatility in real-world applications such as harvest timing, quality grading, and supply chain monitoring.

The non-destructive nature of Vis/NIR spectroscopy offers significant economic advantages over traditional destructive methods by reducing fruit waste and labor costs. According to industry market reports and publicly available data from equipment suppliers, portable spectrometers require an initial investment of USD 5000–$30,000, while inline industrial systems exceed USD 100,000 but enable high-throughput sorting (5–10 fruits/second) with minimal operational expenses. Current technology readiness level ranges from 7 to 8 for portable devices (field-validated prototypes) to 8–9 for inline systems (commercially deployed in apple/packaging lines). However, deployment barriers persist, including model transferability across fruit cultivars, calibration maintenance costs, and hardware standardization. For small-scale producers, cooperative sharing of spectral libraries and cloud-based calibration tools could mitigate costs.

Despite significant advancements, the accurate assessment of fruit biochemical parameters faces significant challenges rooted in biological variability and technical limitations. Fruit heterogeneity, including variations in size, shape, rind thickness, and internal structures, induces spectral noise and complicates model calibration. Calibration models often lack robustness across cultivars, seasons, and regions, requiring frequent recalibration that increases costs. The transition from laboratory models to industrial online systems is particularly hindered by environmental fluctuations, inconsistent sample presentation, and real-time processing demands, which collectively impede widespread adoption.

Moreover, the advancement of Vis/NIR spectroscopy for fruit quality detection faces a significant constraint due to the scarcity of open, standardized spectral datasets. Most studies fail to provide sufficient data for independent verification, hindering algorithm validation and equitable benchmarking across studies. To address this, we urgently call for: (1) adoption of a minimal metadata standard covering instrument specifications, acquisition parameters, reference measurements, sample geometry, environmental conditions, and preprocessing steps; and (2) establishment of a centralized, open-access spectral database for fruit quality analysis. While no comprehensive public database currently exists covering the diverse range of fruits, we regard its creation as an immediate priority to transition from isolated studies to a collaborative, transparent research environment.

Future research should prioritize the development of integrated technological strategies to overcome these limitations. (1) Establish open-access, standardized spectral databases (timeline: 1–2 years): This involves multi-institutional efforts to collect and share spectral data across diverse fruit cultivars, growing conditions, and seasons, adhering to a minimal metadata standard that includes instrument specifications, acquisition parameters, and environmental conditions. (2) Develop robust calibration transfer protocols (timeline: 2–3 years): Research should focus on conducting multi-site studies to validate calibration models across different spectrometers, environments, and fruit batches. Key tasks include designing experiments to assess model degradation in operational settings and creating adaptive algorithms that minimize recalibration costs. (3) Develop adaptive optical mechanisms for online detection equipment (timeline: 3–4 years): This entails designing and testing optical systems that automatically adjust for variations in fruit size, shape, and rind thickness, such as dynamic focus lenses or multi-angle probes, and integrating these into industrial platforms. The fusion of Vis/NIR with hyperspectral imaging will be particularly critical, as it enhances spatial resolution capabilities, enabling the simultaneous quantification of chemical constituents and the visualization of defect distribution patterns in real time. (4) Miniaturize and enhance portable sensors (timeline: 2–3 years): Efforts should leverage low-power microcontrollers and improved detector sensitivity to develop next-generation handheld devices. (5) Construct AI-driven data fusion architectures: Research should integrate multi-sensor inputs (e.g., Vis/NIR, acoustic, and hyperspectral imaging) with advanced deep learning models to decipher complex spectral-fruit property relationships, which will pave the way for autonomous, next-generation non-destructive quality assessment platforms.