Multimodal Sensing and Machine Learning for Nondestructive Quality and Safety Assessment of Agricultural Products

Abstract

The demand for non-destructive evaluation of agricultural products has stimulated the emergence of multimodal sensing technologies capable of integrating diverse quality- and safety-related information. Recent advances in multimodal sensing and machine learning (ML) have enabled intelligent platforms capable of integrating diverse physicochemical, structural, and biochemical information for comprehensive product characterization. This review summarizes the principles, data-processing strategies, and applications of multimodal sensing systems combining machine vision, visible-near-infrared spectroscopy, hyperspectral imaging, Raman and fluorescence techniques, electronic noses, and auxiliary information with ML algorithms. The complementary capabilities of different sensing modalities for evaluating external appearance, internal quality, chemical composition, freshness, spoilage, and safety hazards are discussed. Meanwhile, the roles of conventional ML and deep learning in feature extraction, pattern recognition, and multimodal data fusion are highlighted. Although advanced algorithms and increased data dimensions improve analytical performance, reliable deployment requires robust datasets, effective calibration transfer, model interpretability, uncertainty assessment, and independent validation. Future advances should prioritize self-supervised and few-shot learning, domain adaptation, physics-informed models, lightweight edge intelligence, and standardized benchmarking to improve model reliability, transferability, and scalability. These innovations will facilitate the deployment of intelligent sensing systems from laboratory environments to real-world agricultural quality and safety management.

1. Introduction

Agricultural quality and food safety are governed by a complex combination of physicochemical, sensory, and biological attributes that span multiple spatial and temporal scales [1,2,3,4,5]. Consumers and processors typically evaluate visible characteristics, including color, texture, maturity, and external defects, whereas regulators and supply-chain stakeholders must additionally monitor intrinsic quality indicators and safety-related risks, such as pesticide residues, mycotoxins, adulterants, and other contaminants. These attributes differ substantially in their expression mechanisms and detection requirements [6,7,8,9,10]. Some properties are directly accessible from external surfaces, while others are associated with internal tissues, molecular composition, or microbial activities. Moreover, many safety hazards occur at trace levels and exhibit heterogeneous distribution within agricultural products. Consequently, a single analytical technique is often insufficient for achieving comprehensive, rapid, and reliable assessment of product quality and safety [11].

Conventional reference methods remain indispensable because they provide accurate and traceable measurements of product quality attributes, microbial contamination, and chemical hazards through well-established laboratory-based techniques [12,13,14], including chromatographic, spectrometric, immunological, and culture-based approaches [15,16,17,18,19,20,21,22]. However, these methods generally involve destructive sampling, complicated sample preparation, specialized reagents and instruments, intensive labor requirements, and relatively long analysis cycles. Their limited sampling capacity also restricts their application for large-scale inspection and real-time process monitoring. In heterogeneous agricultural commodities, destructive sampling strategies may further overlook localized defects, contamination, or quality variations [23,24,25,26]. These limitations have driven the development of rapid, non-destructive sensing technologies capable of enabling high-throughput and continuous evaluation throughout production, storage, and distribution chains.

Recent advances in sensor technology, artificial intelligence, and data analytics have promoted a transition from conventional analytical testing toward intelligent sensing systems. Different sensing modalities provide complementary information associated with diverse product characteristics. Machine vision converts surface appearance into measurable color, shape, and texture descriptors [27,28,29,30,31]. Visible-near-infrared spectroscopy (Vis-NIR) enables rapid characterization of agricultural products by capturing spectral signatures associated with their chemical composition and internal structures. Multispectral and hyperspectral imaging (HSI) further integrate spectral and spatial information, allowing visualization of chemical and physical heterogeneity within samples. Raman and surface-enhanced Raman (SERS) scattering provide vibrational fingerprints with higher molecular specificity [32,33,34,35,36,37,38]. Fluorescence sensing offers sensitive responses toward endogenous fluorophores and biochemical interactions, whereas electronic noses (E-noses) characterize volatile signatures associated with ripening, spoilage, or contamination. However, the high-dimensional and heterogeneous data generated by these sensing platforms require advanced analytical strategies. Machine learning (ML) and deep learning (DL) approaches have therefore become essential for extracting informative patterns, establishing nonlinear relationships, and converting complex sensor signals into actionable outputs, including classification, quantitative prediction, and anomaly detection [39,40,41,42].

The integration of multiple sensing modalities represents an emerging strategy to overcome the limitations of individual sensors by combining complementary information from different physical domains. Compared with single-modal sensing, multimodal systems have the potential to provide more comprehensive characterization, improve robustness against environmental variations, and enhance prediction reliability. Nevertheless, the successful implementation of multimodal sensing requires not only effective data fusion algorithms but also careful consideration of sensor compatibility, information redundancy, model complexity, and practical deployment conditions.

Four key questions guide this review: Q1. Which sensing technologies are appropriate for specific quality and safety targets? Q2. How do ML and DL approaches improve information extraction beyond traditional chemometric methods? Q3. Under what conditions can multimodal fusion enhance analytical robustness, generalization, and reliability? Q4. What are the major barriers preventing laboratory-developed models from being translated into real-world intelligent inspection systems?

Although numerous reviews have summarized individual sensing technologies, specific agricultural commodities, or particular machine learning algorithms, several critical challenges remain insufficiently addressed. Many previous studies emphasize reported prediction accuracy while providing limited discussion of experimental design factors, including data partition strategies, sample independence, preprocessing leakage, external validation, reference-method uncertainty, and industrial applicability [43,44]. Furthermore, DL is frequently regarded as a universal replacement for traditional chemometric approaches, despite the fact that many agricultural datasets contain limited sample numbers, strong correlations, and insufficient diversity to support reliable model generalization [45,46]. In addition, multimodal sensing studies are often evaluated based on incremental improvements in internal validation performance, while issues such as true information complementarity, robustness under missing modalities, computational cost, and economic feasibility remain underexplored [47,48,49,50,51].

Therefore, this review presents a reliability-oriented perspective on multimodal sensing and ML-based nondestructive assessment of agricultural products (Figure 1). Rather than simply summarizing sensing techniques and algorithms, this review establishes connections between sensing mechanisms, target attributes, data-processing strategies, and intelligent decision-making. The advantages and limitations of conventional ML and DL models under different data conditions are critically analyzed, while multimodal fusion is examined as a strategy requiring scientific validation rather than an assumed improvement. Particular emphasis is placed on model interpretability, uncertainty evaluation, calibration transfer, domain adaptation, and deployment scalability. Through this perspective, the review aims to provide insights into the development of robust, transferable, and practical intelligent sensing platforms for future agricultural quality and safety management.

Figure 1. Multimodal sensing and machine learning for nondestructive agricultural product assessment.

2. Analytical Targets as Measurement and Decision Constraints

Rather than treating agricultural quality and safety attributes as a catalogue of measurable indicators, this section classifies analytical targets according to four decision-relevant properties: the causal proximity between the target and the measured signal, the spatial and temporal heterogeneity of the target, the likely sources of confounding, and the level of evidence required for the intended use. A sensing system may be appropriate for rapid grading when it captures a stable surrogate of product quality, but the same system may be insufficient for safety or regulatory decisions when the target is present at trace levels or is only indirectly related to the measured response. This distinction provides the basis for selecting sensing modalities, modeling strategies, and validation protocols in the following sections.

2.1. External and Internal Quality Attributes

The target of nondestructive inspection should be defined before selecting a sensor or model. External quality includes color uniformity, dimensions, shape, surface roughness, bruising, cracking, insect damage, mold colonies, and other visible defects. These attributes are naturally represented by spatial patterns and are therefore suited to machine vision, multispectral imaging, or hyperspectral classification maps. Internal quality includes water content, soluble solids, acidity, protein, fat, starch, firmness, maturity, internal browning, and tissue breakdown [52,53,54,55,56,57]. Such properties influence absorption and scattering but may be only indirectly related to the measured optical signal. Their predictability depends on penetration depth, sample geometry, temperature, skin thickness, and the strength of the calibration reference (Table 1).

Table 1. Comparison of major sensing technologies for nondestructive quality and safety assessment of agricultural and food products.

Technology	Information Source	Typical Targets	Advantages	Limitations	Typical Use	Ref.
Machine vision	Spatial patterns: color, texture, shape	Size, color, defects, foreign objects	Low cost, fast, intuitive, simple calibration	Surface information only	Sorting, grading, defect detection	[58]
MSI	Discrete bands (reflectance/emittance) at selected λ	Pigments, moisture, surface defects	Moderate cost, simple calibration	Limited spectral resolution	Sorting, moisture mapping	[59]
HSI	Continuous spectrum for each pixel	Composition, moisture, chemical distribution	Rich information, high spectral resolution	High data volume, complex analysis	Quality prediction, material mapping	[60]
NIR spectroscopy	Molecular overtones and combination absorptions	Moisture, protein, fat, carbohydrates	Fast, robust, minimal sample prep	Overlapping bands, matrix effects	Quality parameters, composition mapping	[61]
Raman/SERS	Molecular vibrational fingerprints	Molecular structure, additives, contaminants	High specificity, structural information	Weak signals, fluorescence interference	Trace contaminants, adulteration	[62,63]
Fluorescence spectroscopy	Electronic transitions of fluorophores	Spoilage, metabolic products, mycotoxins	Highly sensitive, selective probes	Quenching, background interference	Spoilage detection, mycotoxins	[64]
E-noses	Volatile compounds	Freshness, spoilage, fermentation stage, allergens	Sensitive to volatiles, integrated responses	Cross sensitivity, drift, humidity effects	Freshness grading, spoilage monitoring	[58]
Other biosensing	Biorecognition events	Pathogens, toxins, adulterants	High specificity, potential for field use	Requires labels or surface preparation	Safety screening, rapid detection	[65]

Abbreviations: NIR, near-infrared; SERS, surface-enhanced Raman scattering; E-noses, electronic noses; MSI, multispectral imaging; HSI, hyperspectral imaging.

2.2. Freshness and Spoilage Assessment

Freshness and spoilage represent dynamic quality states resulting from complex physicochemical and biological changes rather than individual measurable analytes. These changes are manifested through alterations in microbial activity, physicochemical properties, biochemical composition, and sensory characteristics [66,67,68,69,70]. Multimodal sensing approaches integrating spectral, imaging, and gas-sensing information provide opportunities to capture complementary indicators associated with spoilage progression [71,72,73,74,75]. However, freshness prediction models are often established using specific reference parameters and should not be directly interpreted as universal microbial detection systems. Therefore, model validation against independent biological or chemical measurements remains essential for ensuring reliable spoilage assessment.

2.3. Food Safety Hazards and Screening Specificity

Safety assessment requires greater analytical rigor because hazardous factors are often present at low levels, unevenly distributed, and difficult to distinguish from complex sample backgrounds. Conventional spectral approaches may capture indirect variations associated with overall matrix changes rather than the specific hazard itself, leading to potential overestimation of predictive capability [76,77,78,79,80]. Therefore, highly specific sensing strategies based on molecular recognition or targeted signal amplification are often required to improve hazard identification accuracy [81,82,83,84,85,86]. However, additional procedures involved in these approaches may compromise the principle of strictly nondestructive analysis. Accordingly, intelligent screening systems should be regarded primarily as rapid risk-assessment tools, while regulatory decisions still require confirmation using reliable reference methods.

2.4. Authenticity and Provenance Assessment

Authenticity and provenance assessment aims to verify the identity, origin, and production characteristics of agricultural products by capturing subtle differences in their inherent composition and structural features [87,88,89]. These tasks often involve complex patterns associated with genetic background, cultivation conditions, processing practices, and supply-chain information. However, authenticity models are particularly susceptible to spurious correlations caused by uncontrolled factors, such as seasonal variations, packaging conditions, instrument differences, or laboratory-specific characteristics. Therefore, high predictive accuracy obtained from randomly partitioned datasets may not necessarily indicate true discrimination capability. Robust model development requires group-aware validation strategies and evaluation using independent origins, production batches, or external datasets to ensure reliable generalization under real-world conditions [90,91].

2.5. Analytical Task Types

Multimodal sensing systems for agricultural assessment generally involve four major analytical task types: classification, regression, localization, and anomaly detection [92,93,94,95,96,97]. Classification aims to assign samples into predefined categories, whereas regression establishes quantitative relationships between sensor responses and continuous quality or safety attributes. Localization provides spatial information regarding the distribution or severity of defects, while anomaly detection identifies samples that deviate from the characteristics represented in the training dataset. Among these tasks, anomaly detection remains relatively underexplored despite its importance for real-world applications, where systems frequently encounter new varieties, variable storage conditions, unexpected contamination events, or sensor abnormalities [98]. Therefore, reliable deployment requires evaluation beyond closed-set accuracy and should consider model robustness under uncertain and previously unseen scenarios.

3. Intelligent Sensing Technologies and Detection Mechanisms

The development of intelligent nondestructive assessment systems relies on sensing technologies capable of capturing complementary information from different physical and chemical domains. Individual sensing modalities provide distinct responses related to the external appearance, internal structure, molecular composition, physiological status, and safety characteristics of agricultural products. However, no single sensing technique can simultaneously satisfy the requirements of comprehensive characterization, high sensitivity, rapid analysis, and practical deployment. Therefore, the integration of multiple sensing modalities has emerged as an effective strategy to overcome the limitations of individual approaches by combining complementary signals and improving analytical robustness.

3.1. Machine Vision and Digital Imaging

Conventional red-green-blue imaging (RGB) measures reflected light in broad red, green, and blue bands [99,100,101]. Its primary information is morphological and colorimetric rather than molecular. After illumination correction and segmentation, images support size measurement, shape analysis, color grading, defect segmentation, and texture characterization [102,103,104,105]. The technology is inexpensive, fast, and readily integrated with conveyors. Its main limitation is the inability to detect internal composition or invisible contamination unless these induce external symptoms. Illumination geometry, gloss, shadows, sample orientation, camera exposure, and background strongly affect performance. Color charts and radiometric control are therefore essential when models are expected to transfer between sites.

3.2. Visible-near-Infrared Spectroscopy

Vis-NIR measures reflectance, transmittance, or interactance over wavelength. The visible region contains information from pigments and color-related absorption [106,107,108,109]. The near-infrared region is dominated by broad overtone and combination bands of O-H, C-H, and N-H vibrations, making it responsive to water, carbohydrates, proteins, and lipids [110,111,112]. Scattering carries additional information on tissue microstructure, particle size, porosity, and cellular integrity. Because the bands overlap substantially, multivariate calibration is required [113]. Penetration depends on wavelength, optical geometry, and matrix; interactance and transmittance can probe deeper than simple surface reflectance but often require more controlled hardware. Vis-NIR is well suited to bulk composition, maturity, soluble solids, moisture, protein, fat, and authenticity [114,115]. It is less suitable for trace hazards that do not materially alter the optical matrix. Spectral sensitivity to temperature and path length means that operational control is part of the measurement, not merely a preprocessing issue.

3.3. Multispectral and Hyperspectral Imaging

MSI records a limited number of selected bands [116,117,118], whereas HSI acquires tens to hundreds of contiguous narrow bands [119,120,121,122], producing a three-dimensional data cube with two spatial dimensions and one spectral dimension. The central advantage is simultaneous localization and chemically sensitive contrast. A bruise, fungal colony, foreign material, or region of internal disorder may be mapped rather than represented by a single averaged spectrum [123]. Push-broom systems are common for conveyor inspection, while snapshot systems offer higher temporal resolution at reduced spectral flexibility.

Hyperspectral data are information-rich but operationally expensive. Acquisition generates large files, and spectral redundancy increases computational burden and overfitting risk. Spatial resolution, spectral resolution, signal-to-noise ratio, exposure time, and conveyor speed are coupled engineering choices [124,125]. The most deployable workflow often uses laboratory hyperspectral data to identify informative wavelengths and then implements a lower-cost multispectral device. Deep learning reconstruction of dense spectra from RGB or sparse bands is another emerging route, but reconstructed spectra must be validated for the downstream task and should not be assumed to preserve chemically meaningful detail [126,127,128,129].

3.4. Raman Spectroscopy and Surface-Enhanced Raman Scattering

Raman spectroscopy measures inelastic light scattering associated with molecular vibrations [130,131,132]. Raman bands are comparatively narrow and chemically interpretable, providing stronger molecular specificity than diffuse near-infrared absorption. Water usually causes less interference than in infrared spectroscopy, which is beneficial for aqueous foods. Raman systems can identify adulterants, additives, pesticides, toxins, and compositional markers [133,134,135]. Limitations include weak native scattering, fluorescence background, heating, heterogeneous sampling, and sensitivity to focus and surface geometry [136].

SERS amplifies Raman signals near metallic nanostructures and can support trace-level detection. Its analytical value is highest when target capture, substrate uniformity, and spectral classification are controlled. However, substrate batch variation, hotspot heterogeneity, matrix interference, extraction recovery, and reproducibility remain major obstacles [137,138]. Many SERS procedures involve extraction or contact sampling, so the term “nondestructive” should be used carefully. For industrial screening, robust substrates and internal standards may be more important than an increment in model complexity [139,140,141].

3.5. Fluorescence and Color-Sensor Arrays

Fluorescence sensing measures emitted light after excitation. Signals may arise from intrinsic fluorophores, microbial metabolites, oxidation products, pigments, or target-responsive probes [142,143,144,145]. Fluorescence HSI can localize biochemical changes and often offers high sensitivity, but excitation wavelength, photobleaching, quenching, background fluorescence, and matrix opacity affect the response [146,147,148,149]. Colorimetric and fluorescent sensor arrays use multiple partially selective sensing elements to create a response pattern [150]. They are attractive for packaging-integrated freshness monitoring and rapid contaminant screening, although shelf life, calibration stability, and environmental cross-sensitivity require careful evaluation.

3.6. Electronic Noses and Volatile Sensing

Electronic noses (E-noses) consist of partially selective gas sensor arrays that generate collective response patterns [151,152,153,154,155], enabling the characterization of volatile fingerprints associated with complex biological and chemical processes. They are particularly suitable for monitoring dynamic changes related to ripening, fermentation, oxidation, spoilage, species identification, and freshness evaluation because these processes produce characteristic variations in headspace composition [156,157,158,159,160]. Unlike chromatographic techniques, E-noses generally provide pattern-based recognition rather than direct identification or quantification of individual volatile compounds. However, their performance can be strongly influenced by environmental conditions, sampling procedures, sensor drift, and signal instability. Therefore, standardized measurement protocols, appropriate calibration strategies, and robust machine learning models are critical for achieving reliable and transferable volatile sensing applications.

3.7. Comparative Interpretation

No sensing modality is intrinsically superior; its value depends on whether the measured signal is causally and operationally linked to the intended decision. RGB imaging is highly deployable for visible defects but cannot provide direct molecular specificity [161]. Vis–NIR spectroscopy is efficient for bulk compositional attributes, yet its broad and overlapping bands make trace-hazard detection and cross-matrix transfer challenging. Hyperspectral imaging adds spatially resolved chemical contrast, but the resulting gain may be offset by calibration demands, data volume, and throughput limitations [162]. Raman and SERS can provide stronger molecular specificity, although weak signals, substrate variability, matrix interference, or sampling requirements may limit routine deployment [163]. Therefore, sensor selection should begin with the dominant failure mode of the intended application rather than with the nominal analytical capability of the instrument (Table 2).

Table 2. Comparative characteristics of intelligent sensing technologies for agricultural quality and safety assessment.

Technology	Primary Mechanism/Information	Best-Suited Targets	Main Strengths	Key Limitations	Typical Deployment	Ref.
RGB machine vision	Broadband reflectance; color, morphology, texture	Size, color, shape, surface defects, grading	Low cost, high speed, mature hardware	Surface only; illumination and pose sensitivity	Conveyor grading, smartphone inspection	[58,164]
Vis-NIR	Pigment absorption; O-H, C-H, N-H overtone/combination bands; scattering	Moisture, soluble solids, protein, fat, maturity	Rapid bulk prediction; compact devices available	Overlapping bands; indirect for trace hazards; matrix dependence	Handheld probes, at-line and in-line monitoring	[61,165]
Multispectral/HSI	Spatially resolved reflectance or transmittance spectra	Bruises, internal disorders, fungal damage, composition maps, foreign material	Localization plus chemical-sensitive contrast	Large data volume, cost, calibration and speed trade-offs	Laboratory mapping, push-broom sorting, selected-band systems	[166,167]
Raman/SERS	Molecular vibrational fingerprints; plasmon-enhanced scattering for SERS	Adulterants, pesticides, toxins, additives, molecular identity	High specificity; narrow bands; water-compatible	Weak signal, fluorescence, substrate variability, sample-interface demands	Targeted screening, portable confirmation, swab/extract workflows	[137,138,168]
Fluorescence/sensor arrays	Intrinsic or probe-mediated emission, quenching and color response	Microbial metabolites, oxidation, toxins, freshness indicators	High sensitivity and imaging compatibility	Photobleaching, background, probe stability, matrix opacity	Packaging indicators, laboratory or portable screening	[142]
E-noses	Pattern of partially selective gas-sensor responses	Ripeness, spoilage, fermentation, species and process differences	Sensitive to headspace dynamics; relatively low cost	Humidity/temperature effects, drift, indirect specificity	Storage monitoring, freshness screening, process control	[156,157,158]

Abbreviations: RGB, red-green-blue; Vis-NIR, visible-near-infrared; SERS, surface-enhanced Raman scattering. E-noses, electronic noses; HSI, hyperspectral imaging.

4. From Data Processing to Trustworthy Models

Data processing is not a neutral technical step: each transformation, feature-selection procedure, and model-selection decision can either improve robustness or introduce hidden optimism. The central question is therefore not which algorithm yields the highest internal score but whether the full analytical pipeline remains stable when the product, batch, instrument, operator, or environment changes.

4.1. Data Preprocessing, Feature Extraction, and Variable Selection

Data preprocessing aims to minimize irrelevant variations while retaining information associated with target attributes [169,170,171]. For spectral data, commonly applied procedures include reference correction, smoothing, denoising, scatter correction, normalization, baseline adjustment, and derivative transformation. However, each preprocessing strategy introduces specific assumptions and may influence model performance. Therefore, preprocessing procedures should be selected based on nested validation rather than optimized according to final test performance [172,173,174,175]. Importantly, all data-dependent operations, including scaling, dimensionality reduction, feature selection, outlier identification, and augmentation parameter estimation, must be performed exclusively within the training folds during resampling [176,177,178]. Otherwise, information leakage from validation samples may lead to overestimated model performance, particularly in high-dimensional spectroscopic datasets with limited sample sizes.

Image-based analysis requires additional procedures to ensure reliable feature representation, including radiometric correction, background removal, segmentation, region-of-interest extraction, and geometric alignment [179,180,181]. A critical consideration is maintaining biological independence during dataset construction. Pixel-level samples generated from the same agricultural product are highly correlated and should not be randomly distributed between training and testing sets. Instead, data partitioning should be conducted according to independent units, such as individual samples, batches, fields, or production lots, before pixel extraction. Feature extraction can rely on handcrafted descriptors or data-driven representations [182,183,184,185]. While conventional image and spectral features remain useful for interpretation, deep learning models can automatically learn hierarchical patterns from complex multimodal inputs [186,187,188]. However, dimensionality reduction approaches such as PCA [189] should not be assumed to retain predictive information simply because they explain large variance.

Feature selection and variable optimization are particularly important for high-dimensional sensing systems because they reduce redundancy, alleviate collinearity, and support the development of cost-effective multispectral devices. The robustness of selected variables should be evaluated across repeated validation procedures, as unstable features may reflect sampling variability rather than meaningful relationships. Features consistently identified across different datasets and supported by known physical or chemical mechanisms provide stronger evidence for practical implementation [190]. Data augmentation can further improve model robustness when generated variations represent realistic acquisition conditions [191], such as illumination changes, sensor noise, or spectral fluctuations. Conversely, unrealistic augmentation may introduce artificial patterns unrelated to underlying mechanisms. Class imbalance and abnormal samples should be carefully managed through appropriate sampling strategies, robust evaluation metrics, and mechanistic investigation rather than automatic removal (Figure 2).

Figure 2. Online full-transmission visible-near-infrared (Vis–NIR) spectroscopy and spectral-modelling workflow for the non-destructive prediction of soluble solids content (SSC) in tomatoes. (A) Conveyor-belt-based full-transmission Vis–NIR detection system and two tomato orientations (Z1 and Z2) used for spectral acquisition. (B) Wavelength-selection workflows based on backward variable selection–partial least squares (BVS–PLS) and simulated annealing (SA). (C) Representative multi-point full-transmission spectra and averaged spectra collected under the two measurement orientations. Adapted from Cai et al. (2024) [192] under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

4.2. Chemometrics and Conventional Machine Learning

Classical chemometrics provides the methodological foundation for many spectral-based agricultural assessments. Principal component analysis (PCA) transforms the original variables into mutually orthogonal principal components ordered by explained variance and is widely used for data visualization, exploratory analysis, dimensionality reduction, and outlier investigation [193]. However, components explaining substantial variance should not be regarded as direct evidence of predictive separability. Partial least-squares regression (PLSR) constructs latent variables by maximizing the covariance between spectral information and target responses, making it particularly suitable for high-dimensional and collinear datasets with limited sample sizes [194]. Its classification counterpart, partial least-squares discriminant analysis (PLS-DA), applies the PLS latent-variable framework to class-coded responses for supervised discrimination. Appropriate calibration of decision thresholds and cautious interpretation of classification probabilities remain important considerations [195].

Conventional machine learning algorithms further extend chemometric analysis by modeling nonlinear relationships and complex feature interactions. Support vector machines (SVMs) identify a maximum-margin separating hyperplane, whereas support vector regression (SVR) extends the regularized kernel-based formulation to continuous responses [196]. Kernel functions enable flexible nonlinear modeling while regularization can reduce overfitting in small- and medium-sized datasets [197,198]. Random forests aggregate predictions from randomized decision trees, while gradients are boosting constructs an additive ensemble by sequentially reducing prediction loss [199]. These approaches can integrate heterogeneous spectral, image, and metadata features and provide useful variable-importance information; however, their performance and interpretation may be compromised when sample sizes are limited or predictors are highly correlated. Simpler models, such as linear discriminant analysis, logistic regression, and k-nearest neighbors, remain valuable baseline approaches when interpretability and operational transparency are required.

The suitability of conventional models depends strongly on the characteristics of the dataset and analytical objective. For relatively small spectral datasets containing several hundred samples, chemometric methods and traditional machine learning algorithms may achieve comparable or superior performance compared with deep learning approaches [200,201,202,203,204]. However, complex imaging tasks, large-scale multimodal datasets, and spatially resolved analysis often require more advanced representation-learning strategies. Feature engineering remains critical for conventional models [205,206], where physically meaningful preprocessing, stable variable selection, image descriptors, and auxiliary information can improve prediction accuracy while reducing hardware complexity. Nevertheless, extensive feature optimization increases the risk of model-selection bias. Therefore, fair comparisons require consistent data partitioning, leakage-free preprocessing, comparable optimization procedures, and reporting of performance variability rather than only the best-performing results (Figure 3).

Figure 3. Hyperspectral-imaging setup and spectral analysis for the non-destructive assessment of eggshell quality and crack detection. (A) Reflectance spectra of chicken eggs acquired by the hyperspectral imaging system. (B) Regression-coefficient profiles and selected characteristic wavelengths for predicting (a) L, (b) a, (c) b color values, and (d) eggshell strength. (C) Schematic of the impact device used to generate cracked egg samples. (D) Line-scan hyperspectral imaging system comprising a camera, imaging spectrograph, lens, dual light sources, and a motor-driven conveyor platform. Adapted from Xie and He [207], Scientific Reports, 2016, licensed under CC BY 4.0.

4.3. Deep Learning for Spectral and Image Analysis

At the methodological level, deep learning models are multilayer differentiable functions trained by back-propagation to minimize a task-specific loss, thereby learning successive representations directly from data rather than relying entirely on manually designed features. Convolutional neural networks use local receptive fields and shared weights to learn repeated spectral or spatial patterns efficiently, which provides the architectural basis for their application to sensing data [208]. For spectral analysis, one-dimensional convolutional neural networks (1D CNNs) treat spectra as sequential signals and automatically extract local absorption and scattering patterns through convolution operations [209,210,211]. These models can capture nonlinear relationships among neighboring wavelengths and provide efficient end-to-end learning. However, their performance is highly dependent on dataset size and diversity, and insufficient samples may lead to overfitting or the extraction of instrument-specific artifacts rather than meaningful chemical information [212,213,214].

For image-based sensing, two-dimensional CNNs have been widely applied to RGB images, multispectral data, Raman imaging, and selected-band spectral image [215,216,217]. By learning hierarchical spatial features, CNNs can effectively characterize appearance variations, structural patterns, and defect-related information. Transfer learning from large image datasets can improve performance when target information is dominated by visual morphology or color characteristics; however, its effectiveness may be limited for spectral representations with statistical properties different from natural images. Three-dimensional CNNs further integrate spatial and spectral dimensions for hyperspectral data analysis, enabling extraction of joint spectral–spatial features [218]. Nevertheless, their computational demands and sensitivity to patch construction require careful experimental design, and specimen-level splitting must be performed before patch generation to prevent information leakage [219,220].

Sequential deep learning architectures, including recurrent neural networks (RNNs), gated recurrent units, and long short-term memory networks, have been explored for spectral interpretation and dynamic quality monitoring [221,222]. Their application is most appropriate when temporal measurements or evolving sensor responses are available, rather than simply relying on the sequential arrangement of wavelength variables. More recently, attention-based architectures and transformers have attracted increasing interest because they can model long-range dependencies that may not be effectively captured by local convolution operations [223,224]. In a transformer, an input spectrum or image is first divided into tokens, such as wavelength intervals for spectral data or image patches for visual data. Each token is projected into query, key, and value representations, and self-attention assigns data-dependent weights to interactions among all tokens. This mechanism enables the model to identify relationships among distant wavelength regions, spatial locations, or spectral–spatial features. Positional information is further incorporated to preserve the sequential order of wavelengths or the spatial arrangement of image patches. Multi-head attention can learn several complementary relationships simultaneously; however, the computational cost increases with token number, and large models may be prone to overfitting when agricultural datasets are limited or insufficiently diverse.

Cross-modal attention extends this principle to multimodal sensing [225]. Rather than simply concatenating features from different sensors, representations from one modality can serve as queries, while representations from another modality provide keys and values. Such a design enables the model to learn which spectral, imaging, volatile-sensing, or metadata features are most relevant to each other for a specific prediction task. Although cross-modal attention can learn latent correspondences between heterogeneous measurements, reliable application still requires rigorous sample-level pairing, temporal synchronization where applicable, modality-specific normalization, and an explicit strategy for missing or unreliable modalities. Attention weights may support model inspection, but they should not be interpreted as direct causal explanations without further validation using independent chemical, physical, or biological evidence.

Deep generative models provide additional tools for representation learning, denoising, anomaly detection, and data augmentation [226]. An autoencoder contains an encoder that compresses an input signal into a low-dimensional latent representation and a decoder that reconstructs the original input. The reconstruction error can be used to identify atypical spectra or images, while the latent representation can reduce dimensionality before subsequent prediction. A variational autoencoder further models the latent representation as a probability distribution and introduces Kullback–Leibler regularization to obtain a continuous and sampleable latent space [227]. This property can facilitate data generation and uncertainty-aware representation learning; however, reconstructed or generated spectra should be checked carefully to ensure that diagnostically important absorption features and physically meaningful variation are preserved.

Generative adversarial networks learn through competition between a generator, which produces synthetic samples, and a discriminator, which distinguishes synthetic from real samples [228]. Diffusion models instead gradually corrupt real data through a forward noising process and learn a reverse denoising process to generate new samples. These approaches may be useful for class-balanced augmentation, denoising, and simulation of rare but realistic quality or safety conditions. Nevertheless, GANs may suffer from unstable training or limited diversity, whereas diffusion models generally require substantial computational resources [229]. Therefore, synthetic samples should be generated using training data only and evaluated using distributional similarity, preservation of key spectral or image characteristics, class-conditional diversity, and performance improvement on a completely independent real-world test set. Synthetic data should complement, rather than replace, representative experimental measurements.

Overall, the comparison between deep learning and conventional machine learning should consider multiple dimensions rather than accuracy alone (Table 3). DL is particularly advantageous when complex spatial–spectral structures, large-scale data diversity, or multimodal interactions must be modeled. In contrast, conventional approaches may remain preferable for limited datasets, interpretable applications, and resource-constrained deployment scenarios. Therefore, model selection should be guided by data characteristics, analytical objectives, and practical requirements rather than the assumption that increasing model complexity inevitably improves performance.

Table 3. Comparison of chemometric, machine learning, and deep learning model families.

Model Family	Strengths	Weaknesses	Preferred Data Regime	Interpretability/Deployment	Ref.
PCA/clustering	Exploration, compression, outlier discovery	Unsupervised; variance may not be predictive	Small to medium datasets	High transparency; not a final predictive proof	[230,231]
PLSR/PLS-DA	Handles collinearity; sample efficient; established in spectroscopy	Mostly linear latent structure; threshold calibration needed	Small to medium spectral datasets	Loadings and coefficients aid interpretation; lightweight	[232]
SVM/SVR	Strong nonlinear performance with regularization	Kernel and parameter sensitivity; limited probabilistic output	Small to medium high-dimensional data	Moderate interpretability; feasible on edge devices	[233,234]
Random forest/gradient boosting	Nonlinear interactions; mixed features; robust tabular fusion	Can overfit correlated small datasets; importance may be biased	Engineered spectral, image and metadata features	Good deployment profile; explanation tools available	[235]
1D/2D/3D CNN	Automatic local feature learning; spatial or spectral structure	Data and tuning demands; possible shortcut learning	Medium to large structured datasets	Saliency possible; compression often required	[235,236]
RNN/temporal models	Sequential and dynamic signal modeling	Training complexity; often unnecessary for static spectra	Time-series sensing or ordered response curves	Moderate latency and interpretability	[237]
Transformer/cross-attention	Long-range dependencies and flexible multimodal fusion	High data and compute requirements; attention is not explanation	Large or pretrained multimodal datasets	Powerful but deployment and audit are demanding	[238]

Abbreviations: 1D/2D/3D CNN, one-dimensional, two-dimensional and three-dimensional convolutional neural networks; PCA, principal component analysis; PLSR, partial least squares regression; PLS-DA, partial least squares discriminant analysis; SVM, support vector machine; SVR, support vector regression; CNN, convolutional neural network; RNN, recurrent neural network.

5. Multimodal Data Fusion

Multimodal fusion should be treated as a decision-design problem rather than an automatic route to higher accuracy. Before fusion is implemented, the intended decision, the cost of false acceptance and false rejection, the independent information contribution of each modality, and the availability of synchronized measurements should be defined. Fusion is justified only when it improves a decision-relevant property, such as sensitivity to a critical hazard, robustness to environmental variation, uncertainty estimation, or continuity of operation after partial sensor failure. If the additional modality provides only redundant information or introduces unstable batch-related signals, a simpler unimodal system may be preferable [239,240,241].

5.1. Preconditions for Fusion

Data-level fusion directly integrates raw or minimally processed measurements from multiple sensing sources [242]. Typical examples include combining spectral regions, synchronizing image and spectral signals, or integrating parallel sensor-response streams. This strategy preserves rich original information and allows early discovery of cross-modal relationships. However, successful implementation requires appropriate alignment of sampling scales, temporal synchronization, normalization procedures, and missing-data management. Direct concatenation of high-dimensional measurements may introduce redundancy, increase computational burden, and aggravate overfitting, particularly when sample numbers are limited. Therefore, data-level fusion is most suitable when sensing modalities are physically compatible and sufficient samples are available to support the expanded feature space.

5.2. Selecting the Appropriate Fusion Level

Data-level fusion is most appropriate when modalities are acquired from the same sample under synchronized and physically aligned conditions and when the sample size is sufficient to support the enlarged input space [243,244]. Feature-level fusion is appropriate when each modality requires independent preprocessing or representation learning, but stable modality-specific features can be extracted and audited before integration. Decision-level fusion is preferable when sensors differ in sampling frequency, data format, reliability, or availability [245,246], because each modality can retain an independent prediction pathway and serve as a fallback when another modality fails. These alternatives should not be regarded as a hierarchy; the appropriate level depends on the measurement architecture and decision requirement.

5.3. Risk-Aligned Error Control and Decision Rules

Decision-level fusion integrates independent model outputs rather than raw measurements or intermediate features. Common strategies include weighted averaging, voting, stacking, Bayesian integration, confidence-based combination, and rule-based decision systems [247,248]. Because each modality is processed separately, this approach provides flexibility when sensors have different acquisition frequencies, formats, or availability. It also improves interpretability by maintaining modality-specific predictions throughout the decision process. However, decision-level fusion may lose fine-grained interactions between modalities that occur at earlier stages of representation learning. In addition, unreliable combination of poorly calibrated probabilities may reduce overall performance, highlighting the importance of confidence estimation and model calibration.

The acceptable error profile must be specified before model optimization. For high-risk safety screening, false-negative risk should be minimized, and uncertain samples should be referred to confirmatory analysis rather than automatically accepted. For grading or sorting, a predefined false-rejection rate may be acceptable when it is justified by product value, throughput, and downstream handling capacity. In all cases, probability calibration, uncertainty estimation, and explicit escalation rules are required before modality outputs are combined into an operational decision.

5.4. Modality Robustness and Fallback Design

Recent advances in attention mechanisms and joint representation learning have enabled deeper integration of heterogeneous sensing information by learning relationships across modalities [249,250]. For example, spectral information can enhance the interpretation of visually ambiguous regions, while volatile signals may provide additional evidence for biochemical changes associated with quality deterioration [251]. Nevertheless, practical multimodal systems must account for variable reliability among sensing sources. Strategies such as modality gating, attention weighting, missing-modality masking, mixture-of-experts architectures, uncertainty estimation, and modality dropout can improve robustness. These mechanisms are particularly important for industrial environments, where sensor contamination, signal saturation, hardware failure, and communication interruptions are unavoidable.

5.5. Verification of Fusion Benefit

Performance improvements after multimodal fusion may originate from different mechanisms and should therefore be interpreted cautiously. A focused comparison of directly comparable studies included in this review illustrates the scale of reported internal fusion gains. For mango-quality classification, the integration of computer vision and a colorimetric sensor array increased prediction accuracy from 92.5% for the best single-modality models to 97.5%¹⁵⁵, corresponding to an absolute gain of 5.0 percentage points [252]. For green-tea-quality classification, multisensory fusion increased prediction accuracy from 78% using the optimal single-sensor information to 92%, corresponding to a gain of 14.0 percentage points. Thus, the directly comparable examples identified in the reviewed literature showed internal prediction-set gains of 5.0–14.0 percentage points (median: 9.5 percentage points).

These gains should not be interpreted as a pooled estimate of causal complementarity. The two studies differed in commodity, target definition, feature construction, and validation design, and the reported performance was based on prediction sets rather than independent batch, season, or site-level external tests. Consequently, the observed improvement may reflect genuinely complementary information, but it may also be influenced by shared storage or maturity conditions, metadata-related confounding, unequal feature-selection procedures, or differences in model complexity and tuning. Multimodal studies should therefore report the best unimodal comparator, the absolute incremental gain, controlled modality-ablation results, group-aware data partitioning, and external validation before attributing performance improvement to complementary sensing information.

5.6. Deployment Trade-Offs

Multimodal fusion provides the greatest value when individual sensing technologies cannot fully describe the underlying biological or chemical processes (Figure 4). Freshness evaluation represents a typical example, where different sensors capture complementary changes associated with visual appearance, internal composition, biochemical transformation, and volatile evolution [58,253,254,255]. Similarly, authenticity assessment may benefit from integrating compositional, structural, and contextual information, whereas safety screening may combine rapid broad-spectrum sensing with highly specific confirmation approaches. However, multimodal systems inevitably increase hardware complexity, maintenance requirements, and computational demands. Therefore, fusion should be adopted only when additional information improves decision reliability, robustness, or practical value beyond what can be achieved by a simpler sensing solution.

Figure 4. Decision-oriented framework for multimodal fusion: define the analytical objective and error cost, demonstrate complementary information, assess alignment and data sufficiency, select an appropriate fusion level, verify the gain through ablation and external validation, and deploy only when reliability improves beyond a simpler baseline.

6. Representative Applications Across Agricultural Products

6.1. Fruits and Vegetables

Fruits and vegetables represent one of the most extensively investigated commodity groups because their quality characteristics are highly heterogeneous and directly associated with consumer acceptance, processing performance, and economic value [58,256,257]. Imaging-based methods are widely used for external quality evaluation, whereas spectroscopic and hyperspectral approaches provide information related to internal composition, physiological status, and quality deterioration [222,258,259,260]. Multimodal sensing has further improved the characterization of complex quality changes by integrating complementary visual, spectral, and biochemical information. However, practical deployment remains challenging due to variations in cultivar, growing environment, seasonal conditions, and postharvest handling. Models developed under controlled laboratory conditions may therefore exhibit limited performance when applied to diverse commercial samples [261,262].

6.2. Grains and Oilseeds

Grains and oilseeds are important targets for nondestructive assessment because their quality is closely related to composition, storage stability, and safety risks [263,264,265,266,267,268]. Spectroscopic techniques, particularly near-infrared sensing, have been widely adopted for rapid evaluation of bulk properties, while hyperspectral imaging enables spatially resolved inspection and kernel-level classification [269,270,271,272]. Machine learning models have shown promise for quality grading, defect recognition, and risk screening. However, safety-related prediction requires careful interpretation because optical responses may reflect indirect changes associated with matrix variation, microbial growth, or physical damage rather than the target hazard itself. Therefore, representative sampling, robust validation, and confirmatory reference analysis remain essential for reliable application.

6.3. Meat and Aquatic Products

Meat and aquatic products require rapid and reliable assessment because their quality changes dynamically during storage and distribution. Multimodal sensing approaches have been applied to evaluate appearance, composition, freshness, authenticity, and spoilage-related changes [127,157,273,274,275,276]. Imaging methods provide visual information, while spectroscopy and volatile sensing capture complementary chemical and biochemical variations. However, these products are highly sensitive to environmental and processing conditions, including storage temperature, packaging atmosphere, sample geometry, and measurement protocols [196,277]. Such factors can introduce substantial domain shifts between laboratory datasets and industrial scenarios. Therefore, model development should consider realistic storage conditions and independent validation to ensure reliable transferability.

6.4. Dairy Products, Edible Oils, and Processed Foods

Dairy products, edible oils, and processed foods provide diverse opportunities for nondestructive composition analysis, quality monitoring, and authenticity assessment. Spectroscopic approaches can effectively characterize compositional variations and detect quality changes when target-related signals are sufficiently expressed. Raman and surface-enhanced Raman techniques offer additional molecular specificity for targeted analysis [278,279,280]. However, processed foods often contain complex matrices involving multiple ingredients, heterogeneous structures, and variable processing histories. Consequently, models developed for one formulation or production condition may not generalize well across different brands, raw materials, particle structures, or manufacturing processes. Reliable deployment requires evaluation of matrix effects and validation across diverse real-world samples.

6.5. Tea, Medicinal Plants, Spices, and Specialty Products

Tea, medicinal plants, spices, and other specialty products are frequently investigated for authenticity, origin verification, quality grading, and processing characterization. Their complex chemical compositions generate characteristic spectral and volatile patterns that can support discrimination among different product categories [165,166,281,282,283,284,285,286,287,288,289]. However, provenance-related models are particularly vulnerable to confounding factors, including harvest season, supplier differences, storage conditions, and instrumental variation. High classification accuracy obtained from randomly divided datasets may therefore overestimate true geographical or biological discrimination capability.

A worked example can clarify this risk. Consider a tea-origin dataset containing samples from three geographical regions, with four independent harvest lots per region and multiple spectra acquired from each lot [290]. If the spectra are randomly divided at the spectrum level, measurements from the same harvest lot may occur in both the training and test sets. The resulting accuracy then evaluates recognition of another spectrum from an already represented lot, rather than provenance discrimination for a new harvest lot. A more appropriate design assigns all spectra from each lot to a single partition and holds out complete lots, while retaining each geographical class in the training data. Where possible, the held-out groups should also represent different seasons, suppliers, or acquisition sites.

The practical consequence of this distinction has been demonstrated in a related Vis–NIR agricultural sensing study. Egg-level stratified five-fold validation produced an accuracy of 0.915 ± 0.015, whereas leave-one-incubation-batch-out validation yielded 0.846 ± 0.014, corresponding to an absolute decrease of 6.9 percentage points [291,292]. This comparison does not itself demonstrate data leakage, because measurements from the same egg were kept within one partition; rather, it quantifies the performance reduction that can arise when a model is evaluated on a previously unseen batch. Accordingly, reliable authenticity assessment should report group-wise or prospective validation results in addition to random or within-group validation.

6.6. Cross-Commodity Matrix Effects and Transferability

Despite successful demonstrations across various agricultural products, transferability remains a major challenge for intelligent sensing systems. Matrix characteristics influence signal generation and model performance through variations in composition, structure, morphology, and environmental conditions. Factors such as water distribution, particle characteristics, surface properties, and temperature-dependent responses can alter spectral, imaging, and volatile signals. Therefore, models should not be assumed to be universally transferable across commodities or application scenarios without sufficient evidence. Establishing robust calibration strategies, domain adaptation methods, and independent validation protocols is essential for achieving reliable cross-matrix deployment (Table 4).

Table 4. Representative application matrix linking products, targets, sensing combinations, and generalization risks.

Product Matrix	Representative Targets	Suitable Sensing Combinations	Modeling Emphasis	Generalization Risk	Ref.
Fruits and vegetables	Maturity, soluble solids, firmness, bruises, disease	RGB + Vis-NIR/HSI; optional fluorescence	Spatial-spectral regression and defect segmentation	Cultivar, season, orchard, temperature, skin thickness	[293,294,295]
Grains and oilseeds	Moisture, protein, fungal damage, mycotoxin risk, foreign material	NIR + HSI; Raman/SERS for targeted hazards	Kernel-level classification; bulk calibration; anomaly detection	Lot, cultivar, storage, uneven contamination	[296,297,298]
Meat and aquatic products	Freshness, oxidation, fat, tenderness, microbial spoilage	RGB/HSI + fluorescence + E-noses	Multimodal temporal prediction and confidence-based screening	Packaging atmosphere, cut geometry, temperature, headspace protocol	[299,300,301]
Milk, dairy, oils and processed foods	Composition, adulteration, oxidation, process state	NIR/Raman + imaging + metadata	Nonlinear regression, authenticity classification, transfer learning	Brand, formulation, particle size, processing conditions	[302,303,304]
Tea, herbs, spices and specialty products	Origin, variety, grade, processing, adulteration	NIR/HSI + E-noses + metadata	Domain-aware classification and open-set recognition	Harvest year, supplier, packaging, provenance confounding	[305,306,307]

Abbreviations: RGB, red-green-blue; Vis-NIR, visible-near-infrared; SERS, surface-enhanced Raman scattering. E-noses, electronic noses; HSI, hyperspectral imaging.

7. Model Reliability, Industrial Translation, Challenges, and Future Directions

7.1. Model Evaluation, Explainability, and Reliability

Reliable evaluation is fundamental for determining whether machine learning models can generalize beyond experimental datasets. Appropriate performance assessment should consider the analytical task, data distribution, and practical decision requirements rather than relying solely on overall accuracy. Classification studies should include class-specific metrics, confusion matrices, and imbalance-aware indicators, whereas regression models should report complementary measures such as prediction error and explained variance [308,309]. Data partitioning strategies must reflect the actual deployment scenario [310,311]. Samples originating from the same biological source, production batch, or environmental condition should remain within the same data partition to avoid information leakage. Random splitting may substantially overestimate performance when hidden group effects dominate [312]. Therefore, independent test sets, nested validation strategies, and external evaluations are essential for assessing true model generalization.

External validation provides a more realistic assessment of robustness by introducing variations in season, origin, operator, instrument, and production conditions. Such evaluation is particularly important for agricultural products because biological variability can arise from environmental factors, cultivation practices, and storage history [313]. When performance degradation occurs, recalibration should not be performed directly on the test set but should be quantified and analyzed. Calibration transfer methods, including standardization approaches, spectral augmentation, representation alignment, and transfer learning, can improve model adaptation across different acquisition conditions [314,315]. However, transfer strategies should be evaluated against appropriate baselines and should clearly report the amount and type of additional data required for successful adaptation.

Beyond accuracy, uncertainty estimation and interpretability are essential for safe deployment. Industrial sensing systems should quantify prediction confidence and recognize samples outside the model’s reliable operating range. Methods including ensemble learning, Bayesian approximation, conformal prediction, and out-of-distribution detection provide potential solutions for estimating different sources of uncertainty [316]. Explainable artificial intelligence approaches can further improve model transparency by identifying influential features, spectral regions, or image areas [317]. However, explanations should be validated for stability, reproducibility, and consistency with known physical mechanisms. Attention maps or visual explanations alone should not be interpreted as direct evidence of causal relationships, particularly in highly correlated spectral datasets [318].

Reproducible research practices are necessary to accelerate translation from laboratory models to practical systems. Comprehensive reporting should include sample characteristics, reference methods, sensor configurations, acquisition conditions, preprocessing procedures, data partition strategies, augmentation methods, model architectures, optimization procedures, and computational settings. Reporting only the highest-performing model among multiple trials may introduce selection bias and limit reproducibility [319,320]. Whenever possible, sharing raw signals, metadata, trained models, and external evaluation results can facilitate independent verification and improve confidence in model reliability (Table 5).

Table 5. Minimum and preferred practices for reliable model development and reporting.

Reliability Item	Minimum Acceptable Practice	Preferred High-Rigor Practice	Ref.
Sampling unit	Split by independent sample	Split by batch/site/season and retain a prospective set	[308,309]
Preprocessing	Fit transformations on training data only	Use a fully nested pipeline with locked external test data	[310,311,312]
Model comparison	Same partitions and metrics	Same inputs, tuning budget, seeds, uncertainty and computer reporting	[313]
Classification metrics	Accuracy plus class-wise sensitivity/recall	Balanced accuracy, MCC, PR-AUC, calibration and decision-cost analysis	[314,315]
Regression metrics	RMSE/MAE and R2	Bias, slope, residual plots, prediction intervals and application thresholds	[316]
External validation	Independent production batch	Cross-season, cross-origin, cross-instrument and multisite validation	[317]
Explainability	Feature or region attribution	Stability tests, perturbation tests and physical/chemical consistency	[318]
Reproducibility	Complete methods and hyperparameters	Code, raw data, metadata, trained model and external predictions	[319,320]

Abbreviations: MCC, Matthews correlation coefficient; PR-AUC, area under the precision–recall curve; RMSE, root mean square error; MAE, mean absolute error; R², coefficient of determination.

7.2. From Laboratory Models to Practical Equipment

The transition from laboratory-based models to practical sensing systems requires simultaneous consideration of hardware adaptation, computational efficiency, and operational reliability. Portable and handheld platforms, including compact spectrometers, smartphone-based imaging systems, and miniaturized sensor arrays, provide opportunities for field, warehouse, and on-site inspection applications. However, miniaturization inevitably alters optical configuration, spectral resolution, signal quality, and calibration stability. Therefore, laboratory-developed models cannot be directly transferred to portable devices without evaluating instrument differences and applying appropriate calibration or transfer strategies [321,322,323,324,325,326]. Successful deployment requires optimization of both sensing hardware and analytical models according to the intended application scenario.

Online inspection and automated sorting impose additional requirements beyond prediction accuracy. Industrial systems must integrate sensing, data processing, sample tracking, and mechanical decision-making within limited time windows. Factors such as acquisition speed, processing latency, false acceptance, false rejection, maintenance requirements, and system throughput are equally important for practical performance [327]. Edge computing and lightweight artificial intelligence can reduce computational demands through model compression, pruning, quantization, knowledge distillation, and compact architectures [328]. A promising strategy is to use high-dimensional sensing platforms for initial information discovery, followed by identification of essential features and development of simplified sensors optimized for real-time decisions [329].

From an industrial engineering standpoint, deployment feasibility is tightly governed by capital expenditure (CapEx), recurring operational/maintenance expenses (OpEx), and line inspection throughput. In commercial conveyor-sorting operations (e.g., fruit and vegetable packing lines), sorting rates typically demand processing speeds between 5 to 15 items per second (or conveyor velocities of 0.5–2.0 m/s) with per-item inference latencies strictly below 30–50 ms [330].

Single-modality machine vision systems operating in the visible domain represent the lowest economic threshold and effortlessly achieve sorting throughputs exceeding 10–20 items/s [331]. Point-scan Vis–NIR spectrometers fall into a moderate price regime and can acquire integration spectra within 10–100 ms, making them viable for high-speed online grading [332].

In contrast, push-broom hyperspectral imaging (HSI) systems, while providing chemical and spatial distribution maps, incur substantially higher hardware investments. Furthermore, due to multi-band exposure times and massive data cube serialization (often generating hundreds of megabytes per second), full-band HSI typically operates at restricted speeds of 1–3 items/s, falling short of high-speed industrial conveyor demands unless paired with stable characteristic band selection down to multispectral setups [333].

Gas-sensing arrays (E-noses) and colorimetric arrays face a different bottleneck: while sensor component costs are low, chamber dead-volume equilibration and sensor baseline recovery impose typical acquisition cycles of 30 s to 5 min per measurement, rendering them unsuitable for continuous in-line conveyor sorting and restricting their economic deployment to batch storage monitoring, headspace warehouse surveillance, or at-line QA screening. Therefore, multimodal fusion must be quantitatively evaluated by the marginal accuracy/safety benefit per dollar of added sensor cost and per millisecond of added latency.

Long-term operation requires robustness against environmental variability and regulatory constraints. Factors such as illumination changes, temperature, humidity, vibration, contamination, and sample positioning can introduce signal drift and reduce model reliability. Therefore, practical devices should incorporate calibration standards, quality-control procedures, environmental compensation, and continuous monitoring mechanisms to identify sensor degradation or out-of-distribution samples. Furthermore, commercial and regulatory acceptance depends on traceability to reference methods, quantified uncertainty, documented maintenance procedures, and clearly defined application boundaries. Validation should therefore match the intended purpose, whether grading, process monitoring, screening, or regulatory decision-making, and performance claims should remain consistent with available evidence (Table 6).

Table 6. Principal barriers to industrial translation and corresponding technical priorities.

Translation Barrier	Operational Consequence	Priority Solution	Ref.
Instrument and batch drift	Calibration decay and unstable thresholds	Transfer standards, invariant representations, drift alarms and scheduled maintenance	[334,335]
Environmental variation	False rejects/accepts under light, temperature or humidity changes	Shielding, reference correction, compensation models and stress testing	[336,337]
Large hyperspectral data	Latency, storage and hardware cost	Stable wavelength selection, multispectral redesign, compression and edge acceleration	[338]
Weak hazard specificity	Risk of confusing indirect matrix effects with contaminants	Target-specific Raman/SERS/biosensing and confirmatory reference analysis	[339]
Missing or failed modalities	Multimodal model cannot produce a decision	Modality dropout, reliability gating, fallback unimodal models and sensor diagnostics	[339,340]
Uncertain predictions	Unsafe automated decisions	Calibrated confidence, conformal prediction, abstention and human/reference escalation	[341]
Lack of standards and traceability	Low regulatory and commercial acceptance	Intended-use validation, reference-method traceability, uncertainty budgets and reporting standards	[342]

7.3. Evidence-Based Barriers to Reliable Translation

A first barrier is the mismatch between experimental datasets and the population encountered after deployment. Agricultural sensing datasets are frequently small, homogeneous, and structured by product batch, season, cultivar, instrument, operator, or storage condition. Random sample-level splitting can therefore produce optimistic internal estimates when correlated measurements from the same independent unit are represented in both training and test sets. The practical consequence of validation design has been demonstrated in a Vis–NIR study of 1600 eggs from four incubation batches: egg-level five-fold validation achieved an accuracy of 0.915 ± 0.015, whereas leave-one-batch-out validation yielded 0.846 ± 0.014 [291]. This difference does not by itself prove data leakage because individual eggs were kept within one partition; however, it quantifies the reduction in performance that can occur when a model is evaluated on a previously unseen batch.

A second barrier is measurement transfer. A model may perform well with one instrument configuration yet fail to transfer directly to another device or laboratory because of wavelength shifts, optical geometry, detector response, sample presentation, or environmental variation. In a multi-instrument study of edible-oil authenticity screening, inter-instrument differences were sufficient to compromise direct transfer of classification models developed using raw low-field NMR data [343]. This finding illustrates why calibration transfer, reference standards, and cross-instrument validation must be treated as part of the analytical method rather than as optional post-processing steps.

A third barrier concerns analytical specificity and the quality of the reference method. Broad spectral or volatile fingerprints may correlate with a hazard-associated matrix change without directly measuring the hazard itself. This problem is especially important for trace contaminants, complex food matrices, and safety decisions with asymmetric false-negative risk. Targeted SERS approaches can improve molecular specificity, but practical implementation remains limited by substrate reproducibility, nonspecific binding, matrix interference, and incomplete standardization [344]. Therefore, high laboratory sensitivity should not be interpreted as sufficient evidence for routine screening unless selectivity, reproducibility, recovery, and comparison with an independent reference method are demonstrated.

A fourth barrier is operational integration. A deployable system must maintain traceable sample handling, stable acquisition, calibration control, data integrity, model monitoring, latency-compatible inference, and defined action rules over extended operation [345]. A highly accurate laboratory model may still be unsuitable for industrial use if it cannot identify sensor drift, missing modalities, out-of-distribution samples, or uncertain predictions. Reliable translation therefore requires co-design of the sensor, reference method, model, quality-control procedure, maintenance plan, and human or confirmatory escalation pathway.

Addressing these barriers requires solution pathways that are matched to their underlying causes. Dataset–deployment mismatch should be addressed through representative sampling across batches, seasons, sites, and instruments, together with group-aware and prospective validation; the trade-off is greater experimental cost and longer data-collection cycles, which should be justified by improved external performance. Instrument and environmental drift require reference standards, calibration transfer, drift monitoring, and scheduled recalibration, but these measures increase maintenance requirements and depend on the availability of representative transfer samples. Weak analytical specificity should be addressed through targeted sensing strategies and confirmatory reference analysis; however, greater specificity may reduce throughput, increase cost, or compromise strictly nondestructive operation. For multimodal systems, reliability gating, unimodal fallback models, and sensor diagnostics can mitigate missing or failed modalities, although they increase system and validation complexity. Finally, calibrated uncertainty estimation, abstention rules, and human or reference-method escalation can reduce unsafe automated decisions, but they necessarily increase the proportion of samples requiring secondary assessment. The effectiveness of these measures should be verified using cross-domain validation, environmental stress testing, reference-method agreement, calibration analysis, and explicit monitoring of false-negative risk.

7.4. Priority Research Directions for Overcoming Identified Barriers

Future research should be prioritized according to the barriers identified above rather than according to algorithmic novelty alone. The first and most fundamental priority is the establishment of representative datasets, traceable reference measurements, and group-aware external validation protocols. These elements address dataset–deployment mismatch, label uncertainty, and hidden batch effects; without them, improvements in model architecture cannot be interpreted as reliable progress.

The second priority is robust measurement transfer and risk-aware model operation. Calibration transfer, environmental compensation, drift monitoring, uncertainty estimation, and out-of-distribution detection should be integrated into model development from the outset. These strategies directly address the loss of performance caused by instrument differences, changing acquisition conditions, and previously unseen samples. Their success should be assessed through cross-batch, cross-season, cross-instrument, and prospective validation rather than internal accuracy alone.

The third priority is reliability-aware multimodal and deployable system design. Multimodal fusion should be introduced only when complementary information has been demonstrated and when the added modality improves a decision-relevant property, such as sensitivity, robustness, uncertainty reduction, or fault tolerance. Lightweight models, selected-band sensors, and edge implementation are then required to balance analytical performance against throughput, maintenance, cost, and latency. Advanced methods, including self-supervised learning, transformers, and generative models, should be evaluated as enabling tools for these three priorities rather than as independent objectives.

Future research should therefore evaluate each proposed advance not only by its improvement in internal predictive performance, but also by the deployment barrier it addresses, the additional resources it requires, and the evidence demonstrating that its benefit is retained under realistic operating conditions.

8. Conclusions

Intelligent sensing integrated with machine learning provides a powerful framework for rapid, non-destructive, and comprehensive assessment of agricultural quality and safety. While conventional chemometric approaches remain highly effective for small-scale and high-dimensional spectral datasets, deep learning offers additional advantages for complex imaging tasks, large heterogeneous datasets, and multimodal representation learning. Multimodal fusion has the potential to enhance analytical performance and robustness by integrating complementary information across sensing modalities, but its success depends on appropriate data alignment, reliable validation, uncertainty assessment, and practical cost considerations. Future progress should move beyond maximizing laboratory accuracy toward developing trustworthy sensing systems characterized by robust generalization, interpretability, transferability, and operational stability. Achieving this goal requires coordinated advances in data quality, model evaluation, sensor engineering, and hardware–software integration to enable reliable deployment under diverse real-world agricultural conditions.