Integrated Process-Oriented Approach for Digital Authentication of Honey in Food Quality and Safety Systems—A Case Study from a Research and Development Project

Featured Application

The process-oriented approach proposed in this study aims to develop a prototype system for honey authenticity assessment based on AI-supported pollen analysis. Potential applications include the automation of botanical origin verification, detection of inconsistencies, and support in monitoring the food supply chain. The system is currently at the testing stage and requires further validation. However, its architecture—based on deep learning models and scalable, secure data processing solutions—has been designed in alignment with food quality assurance requirements. In the future, this solution may contribute to enhancing consumer safety, increasing market transparency, and enabling producers to obtain higher prices for products that meet defined quality standards.

Abstract

The increasing scale of honey adulteration poses a significant challenge for modern food quality and safety management systems. Honey authenticity, defined as the conformity of products with their declared botanical and geographical origin, is challenging to verify solely through documentation and conventional physicochemical analyses. This study presents an integrated, process-oriented approach for digital honey authentication, building on initial findings from an interdisciplinary research and development project. The approach includes the creation of a comprehensive digital pollen database and the application of AI-driven image segmentation and classification methods. The developed system is designed to support decision-making processes in quality assessment and VACCP (Vulnerability Assessment and Critical Control Points) risk evaluation, enhancing the operational resilience of honey supply chains against fraudulent practices. This study aligns with current trends in the digitization of food quality management and the use of Industry 4.0 technologies in the agri-food sector, demonstrating the practical feasibility of integrating AI-supported palynological analysis into industrial workflows. The results indicate that the proposed approach can significantly improve the accuracy and efficiency of honey authenticity assessments, supporting the integrity and transparency of global honey markets.

1. Introduction

Food authenticity is a critical aspect of quality and safety management systems in global supply chains. It involves not only the precise identification of the raw material composition and geographical origin of products but also the prevention of practices that can mislead consumers [1]. According to the definition, food authenticity refers to the consistency of actual product characteristics, such as botanical composition, production technology, and origin, with the declared information. It also implies the absence of intentional practices that might deceive consumers, including composition adulteration, incorrect labeling, or unauthorized health claims [2,3,4]. In the context of global supply chains, where products often travel thousands of kilometers before reaching consumers, authenticity is essential for ensuring food safety, protecting public health, and building consumer trust [5].

Honey, one of the most frequently adulterated food products, presents particular challenges in terms of authenticity. These challenges arise from the high market value of honey, the diversity of its varieties, and the difficulty in precisely determining its botanical and geographical origin [6,7]. Additionally, the globalization of trade and complex supply chains create numerous opportunities for introducing products of questionable origin into the market [8]. Therefore, it is essential to develop advanced analytical methods and quality management tools that enable effective fraud detection and origin verification for honey [9,10,11]. However, some modern analytical techniques, such as metagenomic DNA analysis (MDA) or nuclear magnetic resonance (NMR), remain controversial within the industry. These concerns primarily relate to their practical application, costs, and the availability of reliable reference databases [12]. This requires further validation studies and standardization of analytical protocols before full implementation [13]. Modern forms of honey adulteration include the following:

• Addition of sugar syrups—significantly alters the chemical composition and sensory properties of honey (the most common form of adulteration).
• Blending of honeys—mixing honeys from different origins to reduce production costs while maintaining a higher market price.
• Incorrect labeling—falsely declaring geographical or botanical origin to increase the market value of the product.

Results from a recent EU-wide control program, coordinated by the European Commission, indicated that up to 46% of honey samples from non-EU sources exhibited characteristics indicative of adulteration—primarily through the addition of syrups or incorrect origin declarations [13,14]. In light of these findings, the issue of honey authenticity has become not only a technological challenge but also a strategic concern in terms of consumer trust and fair competition.

Although numerous analytical methods are available, including isotopic analysis (IRMS), NMR, FTIR spectroscopy and pollen analysis, none provide full reliability under operational conditions, particularly for small- and medium-sized enterprises (SMEs) [6,15]. Effective honey authenticity verification, therefore, requires not only high-quality data but also the ability to efficiently integrate these data into quality management systems [16].

In response to this challenge, integrated process-oriented approaches that combine laboratory validation with food safety management systems and international quality standards (BRCGS, IFS Food, FSSC 22000), vulnerability assessment of product adulteration (VACCP), and advanced digital technologies have emerged as promising solutions [17]. Special attention is given to artificial intelligence (AI) tools, including deep learning methods and image processing techniques, such as the segmentation and classification of microscopic pollen images. AI enables faster, more precise, and automated fraud detection, supports the identification of botanical origin, and reduces the risk of human error—a critical factor in global supply chains [18,19,20].

The present study responds to critical limitations identified in current honey authentication practices by proposing an AI-enhanced approach grounded in digital palynology and supported by systemic data integration principles. While numerous analytical techniques have been explored in the literature, most lack scalability, automation, or interoperability with food quality assurance frameworks—particularly under the operational constraints of small- and medium-sized enterprises. This project addresses these gaps by investigating the feasibility of integrating image-based pollen analysis with deep learning models and IT architecture designed for industrial traceability.

The following sections provide the theoretical and regulatory context (Section 2 and Section 3), outline the methodological framework (Section 4), and synthesize recent technological developments relevant to food authentication (Section 5). This structure supports a critical exploration of both the conceptual underpinnings and practical constraints of automated honey authenticity assessment.

The aim of this article is to present a process-oriented, proof-of-concept solution for the digital authentication of honey botanical origin, developed within the NUTRITECH.I-004A/22 R&D project. Although not yet implemented in industry, the proposed approach reflects the current direction of innovation in food safety technologies and offers a foundation for future applications within certification, traceability, and fraud prevention systems.

2. Authenticity Management in Normative Systems and Food Safety Standards

In the era of globalized supply chains, food authenticity has become one of the key areas of quality management. It is understood as the consistency of the declared characteristics of a product with its actual properties, including geographical and botanical origin, production method, ingredient integrity, and the absence of intentional consumer deception [3,21,22]. For the honey sector, characterized by high commercial value, complex supply chains, and a high risk of fraudulent practices, this issue is particularly significant.

Within normative systems and food safety management standards [23,24,25], the requirement is to identify and assess the risk of fraud. One of the tools supporting these efforts is the VACCP (Vulnerability Assessment and Critical Control Points) framework, which extends the classic HACCP approach by including an assessment of vulnerabilities to economic and reputational risks. VACCP involves identifying points susceptible to fraud, conducting risk analyses, and implementing appropriate preventive and corrective actions [26,27].

However, it should be noted that not all advanced analytical techniques, such as LC-HRMS, NMR, or DNA-based methods (MDA), are universally accepted as standard methods within these systems. This is due to the lack of fully harmonized procedures, analytical complexity, and variability in results, which can depend on laboratory conditions and sample type. Additionally, some of these methods are not yet formally recognized by regulatory bodies, as highlighted by industry organizations such as FEEDM [28] and the European Commission [12].

The effectiveness of authenticity management largely depends on the quality of input data, the availability of current market analyses, supply chain transparency, and the degree of integration with other quality system components [29,30]. Key factors in this context include the following: traceability, which is the ability to track the origin of a product throughout the supply chain; documented verification of origin; supplier oversight; and validation of applied fraud detection methods [31]. A summary of the main categories and components of authenticity management systems is presented in Table 1.

The modern scientific literature emphasizes that effective authenticity management cannot be limited to technical aspects alone. Equally important are organizational factors, such as management commitment, the maturity of organizational culture, staff competencies, and effective communication with stakeholders [32,33]. In this context, the concept of “food integrity culture” is gaining recognition, reflecting an organization’s sensitivity to fraud risks and its ability to proactively manage them.

Despite the growing awareness of this issue, the implementation of VACCP systems still faces several challenges. Among the most frequently reported barriers are the lack of standardized methods for authenticity validation, limited access to analytical tools (especially in small- and medium-sized enterprises, SMEs), low operational awareness among employees, and poorly adapted risk assessment checklists [26,27,34,35]. Additionally, there is a need for broader use of predictive tools, such as risk modeling based on digital data and better integration of information from documentation and laboratory analyses.

Table 1. Key elements of food authenticity management systems.

Category	Description/Examples
Normative Systems	ISO 22000:2018 [23], IFS Food v.8, BRCGS v.9—standards that include requirements for authenticity assessment, traceability, and food fraud vulnerability management
Management Tools	VACCP (Vulnerability Assessment and Critical Control Points), TACCP (Threat Assessment and Critical Control Points), traceability systems—enable the identification of risks and preventive measures across the food supply chain
Operational Procedures	Fraud vulnerability assessment (VACCP), supply chain mapping, supplier verification (due diligence), key performance indicator (KPI) monitoring, and risk analysis
Supporting Technologies	Advanced analytical techniques, Business Intelligence systems, predictive algorithms
Competency Requirements	Staff training, risk documentation, internal and external audits, development of food safety culture
Supporting Elements	Data integration (traceability, data analytics, risk assessment), digital decision-making support, AI-based analysis and machine learning models

Source: own study based on [23,24,25,26,27,36,37].

Table 1. Key elements of food authenticity management systems.

Category	Description/Examples
Normative Systems	ISO 22000:2018 [23], IFS Food v.8, BRCGS v.9—standards that include requirements for authenticity assessment, traceability, and food fraud vulnerability management
Management Tools	VACCP (Vulnerability Assessment and Critical Control Points), TACCP (Threat Assessment and Critical Control Points), traceability systems—enable the identification of risks and preventive measures across the food supply chain
Operational Procedures	Fraud vulnerability assessment (VACCP), supply chain mapping, supplier verification (due diligence), key performance indicator (KPI) monitoring, and risk analysis
Supporting Technologies	Advanced analytical techniques, Business Intelligence systems, predictive algorithms
Competency Requirements	Staff training, risk documentation, internal and external audits, development of food safety culture
Supporting Elements	Data integration (traceability, data analytics, risk assessment), digital decision-making support, AI-based analysis and machine learning models

Source: own study based on [23,24,25,26,27,36,37].

In this context, tools supported by Industry 4.0 technologies, particularly artificial intelligence (AI), are gaining importance. Automated analytical techniques, such as pollen analysis, spectroscopy, and predictive modeling, can significantly improve the objectivity, speed, and repeatability of authenticity verification, while also supporting real-time decision making in quality management. The integration of data from various sources (traceability, chemical analyses, risk models) is currently one of the most promising directions for the development of modern food quality systems.

3. Limitations of Classical Fraud Detection Methods and Their Potential Integration into Quality Management Systems

Despite significant advances in food analytics, honey authenticity verification still relies primarily on classical physicochemical, spectroscopic, chromatographic, and documentary methods. While each of these approaches has important advantages, their effectiveness can be limited under industrial conditions, where rapid, scalable, and repeatable verification is required [11,38,39,40]. In particular, techniques such as LC-HRMS and NMR, despite their high precision and potential for detecting adulteration, are not yet formally accepted within many normative standards. This is due to the lack of clearly defined validation protocols, difficulties in harmonizing results, and the absence of full acceptance by regulatory bodies [12]. Similarly, DNA-based techniques (MDA), though promising, remain controversial in the industry due to the potential degradation of genetic material, variability in results depending on analytical conditions, and a lack of widespread acceptance in the sector [28]. The most commonly used methods include the following:

• Isotopic Analysis (IRMS)—used to detect the addition of plant-derived sugars. However, this method is costly and can be challenging to interpret, particularly when analyzing honeys with similar isotopic profiles [39,41].
• Spectroscopy (FTIR, NIR, UV-Vis)—allows for rapid chemical profiling, but the effectiveness of this approach depends on the quality of calibration models, reference databases, and measurement accuracy [42,43,44].
• Chromatography (HPLC, GC-MS)—enables the identification of specific botanical markers and detection of adulterants but requires time-consuming sample preparation and expert interpretation of results [45,46,47,48].
• DNA Analysis (MDA)—allows for highly precise botanical identification but re-quires clean samples, is susceptible to genetic material degradation, and remains controversial in the industry due to inconsistent results under varying analytical conditions [28,49].
• High-Resolution Spectroscopic Techniques (NMR, LC-HRMS)—offer high analytical precision, but the lack of standardized validation protocols and accreditation requirements limit their acceptance in normative systems [12,50].
• Documentary Analysis (traceability)—a fundamental tool for quality audits, but without the support of laboratory and digital analytics, its ability to detect fraud is limited [26,27].

In the honey sector, particular challenges include the detection of blends with similar chemical profiles and the precise determination of geographical origin. These issues arise from similarities in pollen composition, chemical variability, and the complex matrix of honey, which can complicate data interpretation, even when using advanced techniques such as spectroscopy or DNA analysis [39,47].

Moreover, many of these methods are costly, require specialized equipment and expertise, and are often inaccessible to small- and medium-sized enterprises (SMEs). The lack of formal accreditation (e.g., ISO/IEC 17025 [51]) and the need for extensive reference databases also pose significant barriers to the widespread adoption of these methods. An overview of these methods, along with their advantages, limitations, and certification requirements, is presented in Table 2.

Table 2. Advantages and limitations of selected methods for honey authenticity assessment.

Method/Technique/Tool	Advantages	Limitations	Accreditation and Validation Requirements
Pollen Analysis (Classical)	Established, low-cost, allows assessment of botanical origin	Subjective, time-consuming, requires experienced specialists, prone to interpretative errors	Typically does not require certification, but validation is recommended
Isotopic Analysis (IRMS)	High specificity in detecting plant sugar adulterants	High costs, difficulties in interpreting results for honeys with similar isotopic profiles	Requires accreditation (ISO/IEC 17025)
Spectroscopy (FTIR, NIR, UV-Vis)	Enables rapid chemical profiling, automation, low unit costs	Effectiveness depends on the quality of calibration models and reference databases, sensitivity to interferences	Requires accreditation (ISO/IEC 17025)
High-Resolution Spectroscopic Techniques (LC-HRMS, NMR)	High precision and sensitivity, capable of detecting a wide range of chemical markers	Lack of standardized validation protocols, high costs, need for extensive reference databases, limited availability in routine laboratories, no clear standards	Requires accreditation (ISO/IEC 17025), lacks clear standards
Chromatography (HPLC, GC-MS)	Accurate detection of botanical markers, residues of adulterants, and foreign compounds	Potential masking by honey matrix components, requires sample preparation and expert interpretation, high costs	Requires accreditation (ISO/IEC 17025)
DNA Analysis (MDA)	High precision, capable of detecting a wide range of species, botanical and geographical origin identification	Requires clean samples, highly sensitive to contamination, needs well-developed reference databases, susceptible to DNA degradation, lacks standardized protocols	Requires accreditation (ISO/IEC 17025), industry controversies
Omics Technologies (Metabolomics, Proteomics, Genomics)	High specificity and sensitivity, allows simultaneous analysis of multiple compound classes	High costs, advanced infrastructure, requires specialized expertise	Requires accreditation and validation (ISO/IEC 17025)
Chemometrics + Hybrid Models	Allows integration of multiple data sources (NIR, pollen, isotopes), high predictive accuracy	High computational and competency requirements, needs interlaboratory validation, standardization challenges	Requires accreditation and validation (ISO/IEC 17025)
Physicochemical Analysis	Fast, standardized, allows detection of deviations from reference profiles	Does not allow definitive detection of adulteration without support from other methods, limited specificity	Typically does not require certification, but validation is recommended
System Elements for Authenticity Management	Support the implementation of VACCP/TACCP, based on risk analysis, audits, and continuous improvement	Require organizational maturity, management commitment, and data availability, high implementation costs	Requires certification (e.g., ISO 22000, FSSC 22000, BRCGS)
Documentation Analysis (Traceability)	Low cost, easy to implement, fundamental to quality systems	Highly susceptible to fraud without analytical or audit support, requires reliable source data	Not applicable

Source: own study based on [12,26,27,28,38,39,40,41,42,43,44,45,46,47,49,50,52].

Table 2. Advantages and limitations of selected methods for honey authenticity assessment.

Method/Technique/Tool	Advantages	Limitations	Accreditation and Validation Requirements
Pollen Analysis (Classical)	Established, low-cost, allows assessment of botanical origin	Subjective, time-consuming, requires experienced specialists, prone to interpretative errors	Typically does not require certification, but validation is recommended
Isotopic Analysis (IRMS)	High specificity in detecting plant sugar adulterants	High costs, difficulties in interpreting results for honeys with similar isotopic profiles	Requires accreditation (ISO/IEC 17025)
Spectroscopy (FTIR, NIR, UV-Vis)	Enables rapid chemical profiling, automation, low unit costs	Effectiveness depends on the quality of calibration models and reference databases, sensitivity to interferences	Requires accreditation (ISO/IEC 17025)
High-Resolution Spectroscopic Techniques (LC-HRMS, NMR)	High precision and sensitivity, capable of detecting a wide range of chemical markers	Lack of standardized validation protocols, high costs, need for extensive reference databases, limited availability in routine laboratories, no clear standards	Requires accreditation (ISO/IEC 17025), lacks clear standards
Chromatography (HPLC, GC-MS)	Accurate detection of botanical markers, residues of adulterants, and foreign compounds	Potential masking by honey matrix components, requires sample preparation and expert interpretation, high costs	Requires accreditation (ISO/IEC 17025)
DNA Analysis (MDA)	High precision, capable of detecting a wide range of species, botanical and geographical origin identification	Requires clean samples, highly sensitive to contamination, needs well-developed reference databases, susceptible to DNA degradation, lacks standardized protocols	Requires accreditation (ISO/IEC 17025), industry controversies
Omics Technologies (Metabolomics, Proteomics, Genomics)	High specificity and sensitivity, allows simultaneous analysis of multiple compound classes	High costs, advanced infrastructure, requires specialized expertise	Requires accreditation and validation (ISO/IEC 17025)
Chemometrics + Hybrid Models	Allows integration of multiple data sources (NIR, pollen, isotopes), high predictive accuracy	High computational and competency requirements, needs interlaboratory validation, standardization challenges	Requires accreditation and validation (ISO/IEC 17025)
Physicochemical Analysis	Fast, standardized, allows detection of deviations from reference profiles	Does not allow definitive detection of adulteration without support from other methods, limited specificity	Typically does not require certification, but validation is recommended
System Elements for Authenticity Management	Support the implementation of VACCP/TACCP, based on risk analysis, audits, and continuous improvement	Require organizational maturity, management commitment, and data availability, high implementation costs	Requires certification (e.g., ISO 22000, FSSC 22000, BRCGS)
Documentation Analysis (Traceability)	Low cost, easy to implement, fundamental to quality systems	Highly susceptible to fraud without analytical or audit support, requires reliable source data	Not applicable

Source: own study based on [12,26,27,28,38,39,40,41,42,43,44,45,46,47,49,50,52].

The recent scientific literature highlights that the effectiveness of these methods can be significantly enhanced through complementary application and integration within food quality and safety management systems [5,27,53]. Particularly promising is the combination of pollen, spectroscopic, and chemometric analyses with traceability data, integrated within advanced risk assessment systems based on artificial intelligence (AI) and omics techniques (e.g., proteomics, genomics), as well as hybrid models [54]. Such an approach supports dynamic fraud risk assessment, continuous improvements in preventive measures, and alignment with international standards, such as VACCP, BRCGS, IFS Food, FSSC 22000, and ISO/IEC 17025, which often require method accreditation and validation [11].

However, some of these modern techniques, such as DNA analysis (e.g., MDA, PCR), LC-HRMS, and NMR, while offering precise botanical and geographical origin identification, remain controversial within the honey sector. This is primarily due to the lack of fully harmonized procedures, challenges related to sample degradation, and high implementation costs, which can limit their availability for small- and medium-sized enterprises (SMEs) [12,28]. Additionally, many of these methods still lack formal accreditation in line with ISO/IEC 17025 [51], which can restrict their acceptance in certification and auditing systems.

Key challenges include limited access to technology for smaller producers, high investment costs, the absence of open, interoperable reference databases, and stringent accreditation and validation requirements [27,35]. For example, pollen analysis—although considered a reference method for botanical origin assessment—is often criticized for its high degree of subjectivity, limited repeatability, and low specificity, particularly when analyzing honey blends. The need for standardizing this technique and harmonizing analytical protocols has been repeatedly emphasized by the scientific community [55,56].

As a response to these challenges, there is growing interest in AI-based solutions that enable the automation of analytical processes. An example is the use of convolutional neural networks (CNNs) and segmentation algorithms (e.g., UNet) for automated analysis of microscopic pollen images, which significantly improves repeatability, reduces analysis time, and minimizes the impact of human error [57,58,59]. These approaches can also be integrated with hybrid models and transfer learning techniques, allowing for their application even with small datasets, thus improving classification accuracy [60,61].

Despite the development of chemometric tools and multivariate models, there is still a lack of universal platforms that integrate data from various sources into a single decision-making system. Therefore, it is essential to develop systems that allow for the automated integration of spectroscopic, pollen, and documentation analysis results, providing a foundation for robust and scalable authenticity control systems.

Given the limitations of classical analytical methods, one of the most promising directions for future development is the automation of honey pollen analysis using deep learning tools. Implementing an AI-based approach can not only increase repeatability and objectivity but also enable integration with digital traceability systems and comprehensive food quality and safety management systems. The practical application of this approach—as developed within a recent research and development project—is presented in the following section.

Despite numerous advances in analytical methods and their gradual integration into food quality systems, there remains a critical gap in practical, scalable, and automated solutions for honey authenticity verification, especially in the context of SMEs. Current methods often lack repeatability, are resource-intensive, and are poorly suited for real-time decision making. Moreover, the absence of standardized, interoperable platforms hampers integration with traceability and certification workflows [11,12,26,27,28,55,56]. These constraints highlight the need for novel, AI-supported solutions capable of automating key steps in the authentication process, particularly pollen identification, and embedding them within broader food safety and quality assurance systems. The present study directly addresses this gap through a process-oriented implementation of digital palynology supported by deep learning and industrial traceability frameworks.

4. AI-Supported Pollen Analysis

Pollen analysis remains one of the key tools in verifying honey authenticity, as it allows for the precise determination of botanical and geographical origin. This technique involves identifying pollen grains present in honey samples and matching them to plant taxa characteristic of specific regions [56,62]. Despite its high diagnostic value and widespread application in scientific research and quality audits, classical pollen analysis has significant limitations, including being time-consuming, subjective, and highly dependent on the operator’s experience [41].

In response to these challenges, automated tools for pollen analysis based on artificial intelligence (AI) have been developed. In particular, deep learning and CNNs have proven effective for the automatic identification, segmentation, and classification of pollen grains in microscopic images. These systems can process large datasets with high repeatability and accuracy, often exceeding 90%, making them an attractive solution for both industrial and research applications, especially in the context of verifying the authenticity of premium honeys [59,63,64,65].

In automated pollen analysis, a critical step is the precise segmentation of pollen grains from the image background, typically performed using architectures like UNet. This is followed by the extraction of key features, including morphometric (shape, size, contour), textural, and pixel intensity characteristics, which are then used to develop classifiers capable of accurately assigning samples to specific botanical taxa [66,67]. This approach reduces subjectivity and minimizes interpretative errors, which are particularly important for honeys from regions with high biodiversity, where the complexity of pollen profiles can pose a challenge for traditional methods [64].

Increasingly, the integration of image analysis with spectroscopic techniques, such as FTIR and NIR, is gaining importance, as it enables the differentiation of samples based on their chemical composition and optical properties [67]. Hybrid models combining visual and spectroscopic features show significantly higher classification accuracy, making them a promising solution for validating the authenticity of honeys from different geographical regions and with diverse botanical compositions [68].

AI-supported automated pollen analysis not only improves the efficiency of the analytical process but can also be integrated with traceability systems and analytical platforms, allowing for the automated detection of discrepancies between the declared origin and the actual pollen profile of the product [11,65,69]. The use of deep learning algorithms, such as CNNs and transfer learning, enables dynamic model adaptation to new data, which is essential in the context of changing botanical diversity and regional differences. This makes it possible not only to rapidly and accurately process images but also to detect anomalies and identify potential fraud that might be missed by traditional methods. This approach supports VACCP, FSSC 22000, and BRCGS systems, contributing to the operational resilience of supply chains and strengthening consumer confidence in authentic honey products [18,26,65].

The application of AI in pollen analysis represents a significant advancement in the detection and prevention of honey adulteration. Deep learning models enable the identification of hidden patterns and subtle anomalies that can indicate potential inconsistencies, thus improving the quality of data used in audits, certifications, and quality management decisions.

In light of these advancements, it was essential to situate our proposed system within the context of existing CNN-based pollen classification frameworks to clarify its specific contributions and enhance its practical relevance. In recent years, several peer-reviewed models have been developed:

-

PollenNet achieved approximately 98% accuracy, precision, recall, and F1-score, utilizing semantic segmentation combined with explainable AI techniques [19].

-

A hybrid ensemble of EfficientNet and SE-ResNeXt networks, trained on a 40-taxa pollen image dataset, reported ~97–97.3% accuracy and F1-score [61].

-

Furthermore, RCANet-style architectures, implementing dual residual attention modules, delivered F1-scores between 97.8 and 98.7% on geographically diverse datasets [60].

Our system differed from and expanded upon these approaches in two key ways:

• It utilized a ResNet152 backbone for deep residual feature extraction, which had proven effective in fine-grained image classification tasks [62].
• It employed Mask R-CNN for instance-level segmentation, which was crucial for resolving overlapping pollen grains, unlike the semantic segmentation employed in other models [63].

Moreover, our pipeline was fully integrated within a microservices-based IT architecture designed for compliance with food safety standards (ISO 22000, BRCGS), enabling industrial and regulatory applicability—an aspect typically absent in purely algorithmic studies.

While a direct numeric comparison was limited by dataset differences, our system achieved 98.4% accuracy and 98.3% F1-score, placing it competitively among the state-of-the-art models and further distinguishing it by its end-to-end deployability for honey authenticity control.

This comparative section clarified the scientific advancements of our approach and underscored its practical novelty as a deployable digital palynology tool for honey quality assurance.

The next section will discuss how these tools can be integrated with the VACCP approach and other food safety and quality management systems to more effectively man-age fraud risk in the honey sector.

5. Fraud Risk Management in the Honey Sector—The VACCP Approach and Innovative Technologies

The effective prevention of honey fraud requires not only modern analytical tools but also well-designed risk management systems. Critical here are systems like HACCP (Hazard Analysis and Critical Control Points), which focus on identifying and controlling health hazards throughout the supply chain, as well as VACCP (Vulnerability Assessment and Critical Control Points) and TACCP (Threat Assessment and Critical Control Points). These extend the HACCP concept by including the assessment of economic, reputational, and intentional threats. VACCP focuses on identifying and minimizing the risk of economic and reputational fraud, taking into account the integrity of the supply chain, while TACCP is primarily concerned with protection against intentional acts such as sabotage or terrorism. These systems have been adopted in many international standards, such as BRCGS (British Retail Consortium Global Standard), IFS Food (International Featured Stand-ards Food), and FSSC 22000 (Food Safety System Certification), which integrate HACCP, VACCP, and TACCP into comprehensive food safety management systems [3,11,26].

The potential for integrating advanced digital technologies, such as artificial intelligence (AI) and predictive analytics, with the VACCP approach is discussed further in this section, with a focus on increasing the operational resilience of the honey sector against fraudulent practices.

The implementation of the VACCP system constitutes a targeted response to the in-creasing scale of food fraud, affecting high-risk commodities, such as olive oil, meat, fish, spices, and honey. According to data published by FoodChain ID, these products represent some of the most frequently adulterated food categories, highlighting the critical need for robust vulnerability assessment and mitigation systems [70,71]. In countries such as Italy, Spain, and France, sector-specific VACCP guidelines have been developed, providing structured threat scenarios and preventive control measures, as emphasized in the FoodDrinkEurope recommendations [72]. Similar strategies have been adopted at the national level, an illustrative example being the United Kingdom, where the National Food Crime Unit (NFCU) operates as a dedicated agency responsible for detecting and counteracting food-related criminal activities [73]. An essential element supporting risk evaluation in this context is the use of analytical tools such as the Food Fraud Database, which compiles detailed records of documented fraud incidents. This resource facilitates the identification of vulnerability hotspots within the food supply chain, thereby enhancing the effectiveness of preventive strategies implemented by food business operators [74].

The effectiveness of the VACCP system depends largely on its alignment with the specific risks of a given sector. In the honey industry, this includes challenges such as the globalization of supply chains, a large number of intermediaries, and the difficulty of detecting certain types of fraud, such as blending honeys with similar chemical profiles [5,11,75]. Given these challenges, a process-oriented approach based on data triangulation, including chemical, palynological, documentary, and external data, is recommended [13,76]. The VACCP vulnerability assessment process typically includes six key steps: (1) gathering information on the product and supply chain; (2) identifying potential authenticity threats; (3) assessing vulnerability to fraud; (4) prioritizing identified risks; (5) defining preventive and corrective actions; and (6) conducting periodic system reviews to evaluate effectiveness [3,26,77].

In practice, tools such as checklists, risk matrices, and scoring systems are widely used in the assessment of food fraud vulnerability. However, there is a risk that these tools may be overly generic, which limits their effectiveness in sectors characterized by highly specific risks, such as the apicultural industry [78]. In Poland, despite growing awareness among food business operators, the implementation of the VACCP system continues to face organizational, competency-based, and technological barriers. Studies indicate that vulnerability assessments are often conducted in a formalistic manner, lacking meaningful integration with broader quality management systems [79,80]. The need for sector-specific guidance and the improved operationalization of fraud prevention measures remains particularly urgent in high-risk supply chains such as those involving honey [81,82].

For the honey sector, it is recommended to update existing checklists by including specific indicators, such as the presence of atypical pollen, inconsistencies between declared and actual botanical composition, supplier trade history, and the use of AI-supported analytics. These measures align with international recommendations on supply chain integrity and traceability, particularly those outlined by FAO/WHO and EFSA [83,84,85].

In practical applications, predictive tools and digital risk assessment systems are becoming increasingly important. The integration of data from multiple sources, such as chemical analyses, AI results, traceability documentation, and commercial data, enables the creation of dynamic decision support models. This approach supports both internal audits and regulatory oversight, increasing operational resilience and transparency in the supply chain.

VACCP typically considers four key categories of vulnerability: (1) economic attractiveness—the potential economic benefit of fraud; (2) technical feasibility—the ability to carry out and conceal fraudulent activities; (3) effectiveness of oversight systems—the strength of existing control measures; and (4) supply chain transparency—the degree of visibility and traceability within the supply chain. These categories are widely recognized in the literature as fundamental components of food fraud vulnerability assessments [3,86]. A detailed overview of these vulnerability areas, along with proposed preventive measures within the VACCP system, is presented in

Table 3. Vulnerability areas in the honey sector and proposed preventive actions within the VACCP system.

Vulnerability Area	Risk Factor	Corrective Actions
Economic Attractiveness	High price of Manuka, organic (BIO), and regional honeys	Documentation verification, laboratory analyses, origin certification, traceability systems, market control (e.g., price monitoring)
Technical Feasibility of Fraud	Addition of sugar syrups, blending with imported honeys	NMR testing, sugar analysis (HPAEC-PAD, FTIR), pollen analysis (AI), isotopic profiling, detection of adulteration using spectroscopic techniques
Insufficient Supply Chain Oversight	Lack of uniform control standards outside the EU, regulatory disparities	Supplier certification (BRCGS, IFS Food), compliance audits, strengthening traceability mechanisms, collaboration with local inspections and certification bodies
Complex Supply Chain	Lack of transparency and traceability, large number of intermediaries	Digital traceability systems, batch geolocation, blockchain, electronic supplier records, early warning systems
Variable Botanical and Geographical Origin	Diverse pollen sources, difficulty verifying origin	Pollen reference databases, pollen geolocation, chemical and isotopic profiling, DNA analysis, botanical validation
Low Operational Awareness	Lack of knowledge about specific risks	Staff training, internal audits, awareness programs, industry-specific checklists, risk indicators
Low Data Quality	Incomplete or outdated analytical data	Regular database updates, data flow automation, integration with ERP systems, quality control of input data

Source: own study based on [82,83,84].

.

Recent studies suggest that the implementation of advanced digital technologies, particularly artificial intelligence (AI), can significantly improve the effectiveness of VACCP systems in the honey sector. AI-based predictive models enable not only the verification of consistency between declared origin and reference data (e.g., pollen profiles) but also early anomaly detection and fraud identification, supporting early-warning systems [59,60].

Integrating these tools with traceability platforms and Business Intelligence (BI) systems enables dynamic risk management, early threat detection, and continuous monitoring of supply chain integrity. This is crucial for preventing fraudulent practices in international trade and for maintaining the transparency and credibility of quality assurance systems, ultimately supporting consumer trust in authentic honey products.

6. Development of a Systemic Approach to Honey Authenticity Assessment—Case Study from a Research and Development Project

Between 2022 and 2024, the first two phases of a research and development project were conducted, with the primary objective of creating a comprehensive, automated system for honey authenticity assessment. This system integrates pollen analysis with artificial intelligence (AI) technologies, creating a solution that addresses both laboratory and operational requirements. The project adopted a process-oriented approach, including data standardization, laboratory automation, IT infrastructure design, and integration with food quality and safety management systems.

A critical milestone of the project was the creation of a large-scale database of microscopic pollen images from honey samples representing four major geographical regions: Europe, Asia, Africa, and the Americas. In total, over 100,000 images were collected in accordance with the PN-A-77626 [87] standard and the European standard CELEX-32001L0110-PL-TXT [88], ensuring technical consistency and high reference value [59].

Sample images were captured using in-house equipment as well as obtained from partner organizations. The imaging and sample preparation procedures were developed in accordance with the Polish Standard PN-88/A-77626. According to this standard, after dissolving the honey sample in distilled water, the mixture was centrifuged (3000 RPM, 10 minutes) to separate the pollen sediment from the rest of the solution. Decantation was then performed, carefully pouring off the upper liquid layer and leaving only the pollen sediment at the bottom of the tube. Another portion of distilled water was added to the sediment, followed by repeated centrifugation and decantation. The purified pollen sediment was then transferred onto a microscope slide and covered with a coverslip.

Original images were taken using a Delta Optical ProteOne microscope equipped with DLT-Cam Pro 5 MP and DLT-Cam Pro 12 MP microscope cameras.

Each image was annotated with comprehensive metadata, including geographic origin, imaging conditions, and taxonomic classification, providing a robust foundation for subsequent analyses.

The collected material underwent standardized digital processing. This process began with image preprocessing, including noise reduction, scaling, and correction for hardware variability, to ensure precise pollen grain isolation. The analytical workflow was divided into three main stages:

• Preprocessing—Image quality enhancement, including noise reduction, scaling, and equipment variation correction, to ensure precise feature extraction.
• Pollen Grain Segmentation—Accurate isolation of individual pollen grains from the image background using architectures like Mask R-CNN ResNet50 FPN.
• Feature Extraction and Classification—Analysis of morphometric (shape, size, contour) and textural features (e.g., GLCM texture indices), which served as input variables for training classification models.

The project utilized two specialized CNNs, Mask R-CNN ResNet50 FPN for pollen segmentation and ResNet152 for botanical classification, supported by Hjorth descriptors to capture complex textural characteristics. This division of tasks between specialized networks allowed for better control over each analysis stage, significantly improving overall model performance.

Figure 1 presents the architecture of the proposed system, illustrating the full pipeline process. The schematic diagram facilitates understanding of how each stage of the pipeline transforms raw input data into the final output. This visual representation breaks the system down into its parts, which, in turn, provides insight into the functions and interconnections of the various modules, thus helping explain the workflow and operations in the system pipeline.

The classification models achieved accuracy exceeding 95%, as confirmed on an independent validation set. These results were presented in a peer-reviewed publication [59], providing strong scientific validation for the approach.

6.1. Methodology

The convolutional neural network (CNN) approach was selected due to its demonstrated effectiveness in extracting complex morphological features from biological images, which is essential for accurate pollen classification

The analytical module of the honey authentication system employs a convolutional neural network (CNN) framework to classify the botanical origin of honey based on microscopic pollen images using two independent processes: segmentation and classification.

Segmentation process

The segmentation pipeline is designed to extract individual pollen grain images from complex microscopic images. To achieve accurate results, the system utilizes the Mask R-CNN ResNet50 FPN model, which combines object detection (bounding boxes) with instance segmentation (pixel-level masks for each detected object). Due to its ability to extract features at different levels of depth, ResNet-50-FPN is frequently used for segmentation tasks. ResNet-50 has an effective FPN, which both captures low-level and high-level image features and helps to semantically segment and detect objects at multiple scales. FPN boosts a model’s multi-scale object detection abilities by merging high-resolution fine-detail shallow features with semantically rich deep features.

The process consists of the following stages:

• Data Loading:
  • Images and corresponding segmentation masks are loaded.
  • Each mask encodes different object instances using different colors (background is black (#000000)).
• Transformations:
  • Images are converted to tensors.
  • During training, random horizontal flips are applied for data augmentation.
• Model Training:
  • The Mask R-CNN model is initialized with weights pretrained on the COCO dataset.
  • The classifier and segmentation heads are adapted to the number of classes (in this case: background + 1 object class).
  • Training is performed for 10 epochs using the SGD optimizer and a step learning rate scheduler.
• Prediction:
  • The trained model is loaded and used to segment new images.
  • For each image, segmentation masks and bounding boxes are generated.
  • Results are saved as cropped image fragments and preview images with marked bounding boxes.

Network Parameters:

5.

Model:

• Architecture: Mask R-CNN with ResNet-50-FPN backbone.
• Pretraining: Weights pretrained on the COCO dataset.
• Number of classes: 2 (background + 1 object class).
• Hidden layer: 256 neurons.

6.

Optimizer:

• Type: SGD
• Learning rate: 0.005
• Momentum: 0.9
• Weight decay: 0.0005

7.

Batch size:

• Training: 2
• Testing: 1
• Number of epochs: 10

The segmentation model achieved strong performance on the test dataset, with an average pixel-wise accuracy of 95% and an averaged F1-score of 0.93, indicating both high precision and recall in distinguishing and segmenting the target objects.

Classification Process

The ResNet152 architecture was chosen for its depth and proven performance in transfer learning tasks, enabling the model to leverage pretrained knowledge and adapt efficiently to the specific characteristics of pollen imagery. This combination allows for robust feature representation and improved classification accuracy compared to traditional machine learning methods. The input dataset comprised labeled pollen images obtained from honey samples collected across five regions of Poland.

The training data were divided into 19 groups (pollen type, number of images): CALLUNA (2199 images), RHAMNACEAE (259 images), VITEX (685 images), PLANTAGINACEAE (402 images), ECHIUM (2207 images), FAGOPYRUM (2271 images), BRASSICA_NAPUS1 (757 images), ARTEMISIA (2221 images), EUCALYPTUS (1713 images), DANDELION (891 images), TILIA (2171 images), PINUS (1771 images), PHACELIA (2214 images), LAVANDULA (1393 images), CHENOPODIUM (386 images), HELIANTHUS (2816 images), CIRSIUM (647 images), AMBROSIA (857 images), SALIX (1088 images).

To avoid overfitting and improve generalization, each group was expanded with augmented images to a total of 2500 images per class, and color information was removed. The augmentation process generates new variants of existing images (such as rotation, flipping, and brightness adjustment), allowing the model to encounter a wider variety of cases. The augmentation algorithm randomly applied these transformations to the pollen images, and 80% of images are used for training and 20% for validation. The split is performed within each class, so both training and validation sets maintain the same class distribution as the original dataset.

The image processing pipeline includes noise reduction using a bilateral filter (parameters: diameter 9, sigmaColor and sigmaSpace set to 75) and histogram normalization with the CLAHE method (clipLimit 2.0, tileGridSize 8 × 8).

OpenCV libraries were used for filtering, CLAHE, and segmentation, while PIL was utilized for augmentation.

The detailed data preprocessing pipeline was built as follows:

• Noise Reduction with Bilateral Filter. The bilateral filter reduces noise while preserving edges, which is particularly important for biological images, implemented as:
  • Framework: OpenCV (cv2.bilateralFilter)
  • Parameters:

    ■

    BILATERAL_DIAMETER: 9 (diameter of the pixel neighborhood)

    ■

    BILATERAL_SIGMA_COLOR: 75 (color weight)

    ■

    BILATERAL_SIGMA_SPACE: 75 (spatial weight)

• Histogram and Intensity Normalization with CLAHE (Contrast-Limited Adaptive Histogram Equalization). CLAHE is a standard histogram normalization technique used to enhance contrast and equalize image brightness, especially useful in microscopic imagery:
  • Framework: OpenCV (cv2.createCLAHE)
  • Parameters:

    ■

    CLAHE_CLIP_LIMIT: 2.0 (maximum histogram clip limit)

    ■

    CLAHE_TILE_SIZE: (8, 8) (tile grid size)

• Randomized Data Augmentation. Augmentation is used to increase the diversity of training data by applying transformations like rotation, flipping, and scaling, which helps models generalize better to real-world variations and reduces overfitting.
• Geometric Augmentation:
  • Library: Pillow 10.3.0 (ImageOps.mirror, Image.rotate)
  • Functions (25% probability per filter):

    ■

    Mirroring (ImageOps.mirror)

    ■

    Rotation (Image.rotate)

• Brightness, Contrast, and Sharpness Augmentation:
  • Library: Pillow 10.3.0 (ImageEnhance)
  • Functions (25% probability per filter):

    ■

    ImageEnhance.Contrast (up to 120%)

    ■

    ImageEnhance.Brightness (up to 110%)

    ■

    ImageEnhance.Sharpness (up to 110%)

The model was implemented using Python 3.10.12 and TensorFlow 2.14. Data preprocessing and image augmentation were performed using the OpenCV 4.9.0.80 and Keras 3.3.3 libraries. Training and testing were conducted on a NVIDIA RTX 4090 GPU server with 16 GB of RAM.

Keras enables the rapid construction, training, and evaluation of deep neural networks. Pre-built modules for transfer learning were utilized, allowing for flexible adaptation of the model architecture to the specific image classification task.

Model Architecture

• Base model: ResNet152 (pretrained on ImageNet)
• Network parameters:
  • Removal of the fully connected layer (include_top = False)
  • Addition of custom layers:
  • GlobalAveragePooling2D
  • Dense layer (Dense (1024, activation = ‘relu’))
  • Output layer (Dense (num_classes, activation = ‘softmax’))
  • Layer freezing strategy:
  • Initial phase: all ResNet152 layers frozen
  • Fine-tuning phase: layers 1–482 frozen, layers 483 and above trainable,
  • Input size: 256 × 256 px

Training workflow, two-phase training process:

• Initial phase:
  • Only the newly added layers are trained
  • Hyperparameters:

    ■

    Optimizer: Adam (default LR = 0.001)

    ■

    Number of epochs: 75 (with Early Stopping applied)

    ■

    Batch size: 32

    ■

    Loss function: categorical_crossentropy

• Fine-tuning phase:
  • Partial unfreezing of the base model
  • Hyperparameters:

    ■

    Optimizer: Adam with a lower learning rate (1 × 10⁻⁴)

    ■

    Number of epochs: 8

    ■

    Early Stopping: monitoring val loss (patience = 5 epochs)

Model validation was carried out using a stratified test set comprising 20% of the dataset (n = 9500), ensuring representation across all pollen classes. Evaluation metrics included overall accuracy (98.4%), recall (99.1%), and F1-score (98.3%). Confusion matrices revealed consistent performance across most classes, with minor misclassifications between morphologically similar pollen types (e.g., Brassica vs. Sinapis).

To enrich classification accuracy, Hjorth descriptors (Activity, Mobility, Complexity) were employed to capture the key textural features of the segmented pollen images. These signal-derived metrics have been adapted to 2D image matrices and previously validated for botanical classification purposes [59]. Their integration enhanced the CNN’s ability to distinguish subtle differences in pollen structure.

Compared to traditional melissopalynological analysis using light microscopy (ICBB 2004) [89], the developed system offers higher reproducibility, reduced subjectivity, and substantial time savings (seconds per image vs. 15–30 min per sample). While human expertise remains indispensable in complex cases, the automated approach supports standardization and scalability in routine quality control.

By modularizing the analytical process—segmentation via Mask R-CNN ResNet50 FPN and classification via ResNet152—the system achieves high adaptability and precision at each stage. All methodological elements were informed by previous studies and iterative testing, ensuring reproducibility and scalability.

6.2. System Architecture and Implementation

Beyond the analytical model, the project also addressed infrastructure requirements critical to building a robust, scalable, and secure solution for honey authenticity assessment. A dedicated server platform, based on microservices architecture, was developed to orchestrate image transmission, task distribution, and GPU-based parallel processing. This approach supports modularity, fault tolerance, and seamless integration with food quality and traceability systems, in line with best practices for interoperable digital infrastructure [90].

The server application was developed using Python and deployed through LXC containers across a network of NVIDIA GPU nodes operating in a secured cloud environment. Data encryption protocols, audit logging, and role-based access controls were implemented to ensure compliance with data protection regulations such as the General Data Protection Regulation (GDPR) and international food safety standards, including HACCP. These features are essential in modern food authenticity systems, which must handle sensitive data while ensuring transparency and traceability across the supply chain [69].

Scalability is achieved via containerized microservices managed by a load-balancing system, enabling horizontal expansion, high availability, and operational continuity even during high-volume analysis. The architecture allows for independent updates and testing of analytical modules without disrupting the broader system, supporting continuous improvements and resilience.

Although detailed benchmarks, load simulations, and throughput metrics will be provided in a forthcoming technical publication, preliminary internal evaluations confirm the system’s capability for the real-time processing of microscopic images using deep learning models. This is particularly important given the computational demands of pollen grain segmentation and classification workflows, where CNN architectures such as Mask R-CNN and ResNet152 are employed [60].

To illustrate the full integration of the analytical and infrastructure components, Figure 2 presents a schematic overview of the end-to-end digital workflow. It delineates the modular sequence of operations, including image acquisition, preprocessing, segmentation, feature extraction, classification, and result integration, into quality management and traceability frameworks. The visual representation highlights both analytical precision and systemic interoperability—key pillars of a future-proof authenticity assessment platform.

The model presented in Figure 2 includes five main stages:

• Data acquisition and standardization—Collection of honey samples (domestic and foreign), microscopic imaging of pollen grains, and metadata annotation in accordance with PN-A-77626 [87] and CELEX-32001L0110-PL-TXT [88] standards.
• Automation of Laboratory Analysis—Digital Image Processing—Preliminary filtration, pollen grain segmentation, and extraction of morphometric and textural features using CNNs—a technique proven effective in recent studies, showing classification accuracies exceeding 90% in morphologically diverse samples [59].
• Development of IT Infrastructure (Technical Components)—Implementation of a microservices architecture, image queuing, and processing systems on dedicated GPU servers, including data transmission encryption and API documentation for automated data flows.
• Validation and Quality Control of Data—Comparison of CNN model results with expert assessments, qualitative tests of server application and data security, and continuous model updates to improve analysis accuracy and precision.
• Integration of the Digital Honey Pollen Analysis System with Quality, Safety, and Authenticity Management Systems—Potential Applications—Integration with standards such as ISO 22000, FSSC 22000, IFS, BRCGS, and VACCP support, including risk assessment and detection of inconsistencies. Future development may include Business Intelligence (BI) and interoperable databases.

Each of these stages was designed to enable scalability, implementation across different organizational types, and integration with existing audit and certification systems.

Figure 2 illustrates not only the sequential analytical workflow but also the systemic architecture of the proposed solution. It represents a digitally enabled, interdisciplinary approach to honey authenticity assessment, bridging traditional palynological expertise with AI-driven process automation.

6.3. Discussion and Preliminary Evaluation

Figure 1 illustrates not only the sequence of analytical steps but also the strategic architecture underlying the developed system. The model represents a systemic and interdisciplinary response to persistent challenges in food authenticity assessment, particularly in the honey sector.

The developed approach effectively addresses reproducibility, automation, and scalability—key issues that have long hindered traditional palynological methods. It reflects a shift from manual, expert-dependent evaluation toward algorithm-supported decision making, integrating data quality, analytical precision, and system interoperability.

At a strategic level, the project combines classical palynological knowledge with Industry 4.0 technologies, resulting in a robust tool for supply chain risk management. The use of advanced deep learning models, such as ResNet152 for classification and Mask R-CNN for pollen segmentation, ensures accurate processing of microscopic images and aligns with the latest progress in biological image analysis [60].

Moreover, this integration modernizes analytical workflows and directly addresses the known limitations of manual pollen identification, including subjectivity, time consumption, and inconsistency, as emphasized in previous research [50,67]. The implemented IT infrastructure, based on a microservices architecture and integrated with secure cloud-based systems, supports efficient data scaling, high availability, and compliance with industrial data standards [90].

By harmonizing metadata structures and embedding secure data protocols, the system achieves interoperability with quality management frameworks and readiness for high-throughput industrial application. Its design anticipates integration with traceability systems, third-party certification platforms (e.g., IFS Food, BRCGS, FSSC 22000), and fraud prevention protocols (e.g., VACCP), offering valuable support to both food industry stakeholders and regulatory authorities.

Automated control processes and intelligent data analytics enhance early risk detection, supporting supply chain resilience and strengthening consumer trust in high-value natural products [69]. The results of this project confirm that effective fraud prevention in the honey sector requires an interdisciplinary and integrated approach, combining data engineering, AI-based analytics, and quality assurance within a unified architecture.

Preliminary validation of the system, as outlined in Section 6.1, showed classification accuracy exceeding 90% on a representative test set, with strong performance across morphologically diverse pollen types. However, these results reflect internal evaluation protocols and limited sample scope. Comprehensive model benchmarking, including confusion matrix analysis, cross-validation, and long-term performance tracking, will be presented in a forthcoming technical publication focused on optimization and scalability.

Nonetheless, the current prototype has been successfully deployed in a controlled R&D environment, confirming its operational feasibility and alignment with food authenticity verification processes. The outcomes also corroborate the findings of Punyasena et al. [67], who confirmed the applicability of CNNs to morphologically complex pollen datasets, further supporting the system’s adaptation to European honey matrices characterized by high botanical diversity.

6.4. Limitations and Future Directions

While the developed system represents a significant advancement in automated honey authenticity assessment, several limitations warrant discussion. First, the current classification models were trained and validated on a dataset primarily composed of pollen images from Polish and selected international honeys. Although morphologically diverse, this dataset may not fully capture the global variability in pollen types, particularly in honeys from tropical and subtropical ecosystems, which exhibit higher botanical complexity [67].

Second, system performance has so far been assessed under controlled R&D conditions. Real-world validation—incorporating variation in sample preparation, imaging equipment, and user proficiency—is necessary to confirm generalizability and robustness. This challenge of external reproducibility is a well-known issue in digital palynology and food image analysis [50].

Moreover, while the use of Hjorth descriptors and deep CNNs has demonstrated high classification accuracy, certain edge cases (e.g., morphologically similar pollen grains) may require multimodal approaches. Future enhancements may include integration with spectral, isotopic, or physicochemical data to support complex authenticity profiles, in line with emerging trends in food forensics [60,69].

Interoperability also remains a critical concern. Despite adopting a microservices architecture and metadata harmonization protocols, successful integration into existing industrial systems (ERP, LIMS, QMS) will require further standardization efforts and compliance with sector-specific interoperability frameworks [90].

Future research will, thus, focus on five key areas: (1) expanding the reference database through international collaboration, (2) external system validation in industrial environments, (3) model benchmarking using standardized public datasets, (4) integration of multimodal data, and (5) dissemination of technical validation results in a dedicated publication.

Ultimately, the presented system offers a scalable, reproducible, and modular foundation for next-generation authenticity control. Its strategic alignment with Industry 4.0, food traceability, and AI-based quality assurance positions it as a promising tool in addressing the complex challenges of food fraud prevention and supply chain transparency [69].

7. Summary and Conclusions

This study presents a novel digital system for the assessment of honey botanical authenticity, integrating image-based palynological analysis with deep learning techniques and modular IT infrastructure. The solution was developed and validated as part of the NUTRITECH.I-004A/22 project, addressing critical gaps in traditional verification methods, which rely heavily on documentation and physicochemical tests and often fail to detect complex cases of honey adulteration.

The interdisciplinary system combines microscopic pollen imaging, automated feature extraction using Hjorth descriptors, and classification via CNNs. Supported by a scalable microservices platform, the system enables real-time image processing and delivers classification accuracy exceeding 90% under controlled laboratory conditions. These findings confirm the feasibility of integrating AI-supported palynological analysis into industrial food quality control workflows.

At the same time, this study highlights several methodological and operational challenges. The current dataset, while diverse, does not yet capture the full global variability of honey pollen. The model’s robustness outside the development environment remains to be demonstrated, and future work must address integration with industrial systems, regulatory frameworks, and diverse imaging conditions. Nevertheless, the system provides a solid foundation for future advancements in automated food authenticity assessment.

These findings allow for several broader conclusions that underline the contribution of this work to the field of food authenticity assessment:

• Standardization of data and analytical protocols is fundamental to reproducibility in AI-based food authentication systems. Consistent results require harmonized imaging, annotation, and preprocessing procedures. This study shows that variability in input data quality significantly affects model performance, underlining the need for unified protocols across laboratories and applications.
• Deep learning applied to microscopic pollen images enables accurate and scalable verification of botanical origin. The tested CNN models achieved high classification accuracy while reducing analysis time and operator dependence. This confirms their utility as a viable alternative to manual palynology, especially in high-throughput or industrial settings.
• Open and validated reference datasets are essential for model generalization and benchmarking. The effectiveness of AI models depends on access to diverse and standardized training data. Structured databases ensure transferability across regions and products, facilitating regulatory acceptance and scientific transparency.
• Systemic integration of AI, palynology, and quality management creates robust tools for honey authenticity assessment. The interdisciplinary approach adopted in this project supports the development of scalable, secure, and auditable systems aligned with food safety standards and fraud risk prevention frameworks.

Further development efforts, grounded in the findings of this study, will focus on the following:

Extending the pollen image database to cover additional botanical groups and underrepresented geographical regions;

➢

Incorporating multimodal data (e.g., spectroscopic profiles) into hybrid classification models;

➢

Designing sector-specific digital tools for VACCP-based risk assessment and fraud prevention;

➢

Publishing technical validation guidelines to support implementation by laboratories, certification bodies, and industry auditors.

This work contributes to the scientific and technological advancement of food authentication systems and supports broader efforts in digital transformation, traceability, and consumer protection within the agri-food sector.