A Comprehensive Review of AI Methods in Agri-Food Engineering: Applications, Challenges, and Future Directions

Abstract

The deep integration of artificial intelligence (AI) is a core driver for digitalization and intelligence in agricultural and food engineering, boosting production efficiency, resource optimization, and product quality. This review systematically analyzes AI’s application scenarios, technical pathways, and challenges across the agricultural value chain. It aims to develop a structured taxonomy of AI-driven technical application mechanisms in agriculture, highlighting their roles in optimizing core agricultural processes. A systematic literature review was conducted using reputable databases, including Google Scholar, IEEE Xplore, ScienceDirect, Web of Science, SpringerLink, and Scopus, focusing on peer-reviewed articles from the last decade. Findings show that AI-enhanced techniques improve product quality and safety inspection efficiency. However, challenges like multi-source data synchronization barriers, high intelligent equipment costs, and model adaptability limitations in complex agricultural environments remain. This review contributes to the field by providing a unified framework for understanding AI applications in agri-food engineering, identifying key research gaps, and highlighting pathways for sustainable technology adoption that can benefit diverse agricultural stakeholders.

1. Introduction

The global agricultural and food system stands at a critical crossroads, confronting unprecedented pressures from a growing population. The United Nations projects a global population of 10 billion by 2050, requiring a 70% increase in food production [1]. Concurrently, traditional agricultural practices—reliant on empirical decision making, labor-intensive manual operations, and fixed-resource allocation—have become unsustainable. They suffer from low resource use efficiency, high post-harvest losses, and limited capacity to adapt to dynamic field conditions [2]. Against this backdrop, AI has emerged as a transformative and irreplaceable driver for agricultural digitalization and intelligence, making it a “critical domain” for addressing core challenges of food safety, resource optimization, and sustainable production.

The core workflow of agricultural and food production spans the entire lifecycle, encompassing pre-production, in-production, and post-production phases [3]. Traditional agricultural technologies present notable bottlenecks at each stage. These challenges include inaccurate crop phenotypic due to manual observation, inefficient use of resources resulting from fixed-threshold irrigation and fertilization, delayed or reactive decision making—such as the application of pesticides only after pest or disease outbreaks—and increased production costs arising from labor shortages [1].

In response, modern agricultural technologies are focused on systematically optimizing these processes to enhance overall productivity and operational efficiency. The modernization of agricultural and food production exhibits a clear trajectory, progressing from mechanization to automation, and is now steadily advancing toward intelligent agriculture and food engineering. Figure 1 depicts the evolution of agriculture development over time. Beginning with the Mechanization phase developed in the 1950s, the Automation phase evolved from the 1990s to 2010s, the Internet of Things (IoT) stage evolved from the 2010s to 2020s, and finally, the trajectory has been developing toward AI fusion.

Currently, the emergence of advanced communication infrastructure, including 5G networks with ultra-low latency and massive connectivity capabilities, enables real-time data transmission from distributed agricultural sensors and UAV-based remote sensing platforms [3]. For instance, in the domain of agricultural machinery path planning, dominant strategies include graph search-based optimization techniques and real-time control algorithms [4,5,6]. Figure 2 shows the scene of 5G-enabled smart IoT infrastructure for agriculture, highlighting how 5G connectivity serves as the backbone to facilitate real-time data flow that empowers farmers with insights for precision irrigation, health monitoring, and automated resource management, ultimately leading to enhanced crop yield and operational efficiency. Although these technological advancements provide a wealth of data for the application of AI and create possibilities for system-level agricultural optimization, the main focus must still be on how AI algorithms can effectively utilize these data streams to address key agricultural challenges.

The rapid advancement of AI technologies has positioned agriculture as a critical domain for transformative innovation, addressing fundamental challenges in food safety, resource optimization, and sustainable production. Modern agricultural systems generate vast quantities of heterogeneous data from diverse sources, including soil sensors, satellite imagery, weather stations, and crop monitoring devices [7]. This data richness creates unprecedented opportunities for AI-driven insights that can revolutionize agricultural decision making, yet current research reveals significant knowledge gaps in integrating these technologies effectively across the agricultural value chain [8,9]. In other words, AI applications in agriculture face unique challenges that distinguish this domain from other sectors. Agricultural environments are characterized by high variability in environmental conditions, complex biological processes, and long feedback cycles that complicate model training and validation. Traditional agricultural decision making relies heavily on empirical knowledge and reactive approaches, creating substantial opportunities for AI systems that can provide predictive insights and optimize resource allocation in real time.

Current reviews have thoroughly examined individual AI technologies and their specific agricultural applications, demonstrating significant advances in machine learning for crop yield prediction [10,11], pest detection [12,13], and automated irrigation systems [14]; the comparison of several reviews are shown in Table 1. However, existing research predominantly focuses on isolated applications rather than examining the interconnected nature of agricultural systems and the potential for holistic optimization, while the comparison among the application of AI in different fields would provide a more panoramic view. Moreover, the challenges encountered in various application scenarios are not fully explored, and a comprehensive analysis on the future research directions could also benefit potential readers [15].

To address these critical knowledge gaps, this review aims to provide a comprehensive, structured, and crossdimensional analysis of AI applications in agricultural and food engineering. The main contributions of this review are summarized as follows:

• Firstly, unlike previous reviews that focus on single AI subfields such as examining only deep learning for disease detection [16], this work systematically maps several core AI pillars to key agricultural processes across the entire value chain of pre-production (crop variety selection, land preparation), in-production (sowing, precision irrigation/fertilization, pest management), and post-production (storage, supply chain traceability). This taxonomy clarifies how different AI techniques address specific agricultural tasks, providing a unified framework for understanding technical pathways.
• Secondly, beyond technical challenges, this review identifies and addresses the fundamental barriers that have hindered widespread AI adoption in agriculture. The primary challenges stem from the inherent complexities of agricultural environments, including high data variability due to diverse climatic conditions, soil types, and crop varieties, which make it difficult to develop generalizable models. Small sample problems persist due to the cost and time-intensive nature of collecting high-quality agricultural datasets, while the lack of standardized data formats across different farming systems creates integration difficulties. Additionally, the economic constraints faced by smallholder farmers limit access to expensive AI-powered technologies, and the digital divide in rural areas poses significant deployment challenges.
• Thirdly, looking toward the future, this review establishes the directions for AI-driven agricultural transformation by linking current applications to specific bottlenecks in traditional farming practices. The identified directions include the establishment of robust dataset construction and standardization systems, the development and deployment of emerging AI technologies that can democratize access, and the enhancement of AI security and explainability. This comprehensive analysis bridges the gap between academic research and practical implementation, serving as a reference for researchers, policymakers, and agricultural technology developers in advancing sustainable and intelligent agriculture.

Table 1. Comparison of various reviews.

Reference	Remote Sensing	IoT Hardware	ML/DL Algorithm	Soil/Crop Monitor	Irrigation Optimization	Pest/Disease Detect	IoT Security	Dataset Build
[10]			✓					✓
[11]	✓		✓	✓	✓
[12]		✓	✓
[13]			✓			✓		✓
[14]		✓	✓				✓	✓
[17]	✓		✓	✓
[18]		✓	✓				✓
This work	✓	✓	✓	✓	✓	✓	✓	✓

Table 1. Comparison of various reviews.

Reference	Remote Sensing	IoT Hardware	ML/DL Algorithm	Soil/Crop Monitor	Irrigation Optimization	Pest/Disease Detect	IoT Security	Dataset Build
[10]			✓					✓
[11]	✓		✓	✓	✓
[12]		✓	✓
[13]			✓			✓		✓
[14]		✓	✓				✓	✓
[17]	✓		✓	✓
[18]		✓	✓				✓
This work	✓	✓	✓	✓	✓	✓	✓	✓

To develop this taxonomy and ensure a structured and logical flow, this paper follows a systematic approach. As shown in Figure 3, the remainder of this paper is organized as follows. Section 2 introduces the fundamental concepts and core techniques of AI. Section 3 discusses the systematic literature review approach, including search criteria, inclusion/exclusion methods, and classification strategies used to develop the proposed taxonomy. Section 4 surveys its applications in agricultural and food engineering, covering scenarios such as agricultural product quality monitoring, safety analysis, and agricultural production process management. Section 5 examines emerging challenges and future directions associated with AI in the agri-food domain. Finally, Section 6 summarizes the study’s key findings.

2. Overview of AI

AI is an interdisciplinary domain that seeks to enable machines to exhibit human-like intelligence by integrating methodologies from computer science, mathematics, psychology, linguistics, and related fields. Its primary goal is to develop computer systems capable of replicating human cognitive functions—such as perception, reasoning, learning, and decision making—to perform complex tasks, including image recognition, natural language understanding, and autonomous control.

The evolution of AI has undergone several cycles of breakthroughs and dormancy. The term “AI” was formally introduced at the Dartmouth Conference in 1956, which marked the beginning of the symbolic paradigm, emphasizing logic-based reasoning as the dominant research focus during its early phase [19]. In the 1980s, the emergence of expert systems facilitated the adoption of AI technologies across various industrial applications.

The DL revolution that began in 2006 ushered in a connectionist paradigm centered on deep neural networks, which overcame longstanding limitations in perception tasks such as speech and image recognition. This transition propelled AI into a new era of rapid advancement fueled by the convergence of big data and high-performance computing [20]. More recently, the rise of generative AI models and the progress in large-scale multimodal architectures signal a paradigm shift from task-specific intelligence toward the pursuit of artificial general intelligence [21].

The technical foundation of AI consists of three core pillars [22]: machine learning, which aims to learn data-driven patterns through algorithms and is typically categorized into supervised, unsupervised, and reinforcement learning paradigms; computer vision (CV), which facilitates the interpretation and generation of visual data such as images and videos; and natural language processing (NLP), which addresses the representation, understanding, and generation of human language. Additionally, technologies such as knowledge graphs, robotics, and expert systems have further broadened the application landscape of AI [21]. The hierarchical structure diagram of AI technology is shown in the Figure 4.

At the application level, AI has demonstrated widespread integration across diverse sectors, ranging from IoT security [18], renewable energy [23], and agriculture [24] to biomedicine [25], green development [26], construction [27], energy systems [28], WSN security [29], and mental health [30]. For instance, AI has achieved high diagnostic accuracy in analyzing chest CT scans and predicting chronic kidney disease [25] and has also shown efficacy in diagnosing schizophrenia using fMRI and natural language-based assessments [30]. In energy applications, AI has enabled accurate forecasting of wind power generation, leading to reductions in grid dispatch costs [23], and has improved short-term photovoltaic output predictions by reducing forecast errors [28]. Nevertheless, the advancement of AI brings forth pressing concerns in ethics, security, and societal implications—ranging from data privacy breaches and algorithmic bias to workforce disruptions and limited transparency and controllability in AI decision making.

3. Research Methodology

This systematic review employs a comprehensive approach to identify, screen, and analyze the academic literature concerning AI applications in agri-food engineering. The methodology follows established guidelines for systematic literature reviews to ensure rigor, transparency, and reproducibility in the research process.

3.1. Literature Search Strategy

The literature search was conducted across multiple prominent academic databases to ensure comprehensive coverage of relevant publications. Primary databases included Google Scholar, IEEE Xplore, ScienceDirect, Web of Science, SpringerLink, and Scopus, supplemented by specialized agricultural databases such as CAB Abstracts and AGRICOLA to capture domain-specific research that might be absent from general academic repositories.

The search strategy employed a systematic approach beginning with the core search string: (“Agri-food” OR “Agriculture” OR “Agricultural Engineering”) AND (“AI” OR “Artificial Intelligence” OR “Machine Learning” OR “Deep Learning”). To enhance precision and capture the breadth of AI applications in agriculture, additional specific terminology was systematically incorporated across four key categories.

Application domains covered “crop monitoring”, “quality assessment”, “precision agriculture”, “pest detection”, “yield prediction”, “supply chain”, and “traceability”. AI methodological approaches encompassed “neural networks”, “support vector machines”, “random forest”, “ensemble learning”, “transfer learning”, and “federated learning”. Sensing and detection technologies included terms such as “computer vision”, “hyperspectral imaging”, “near-infrared spectroscopy”, “SERS”, “electronic nose”, and “biosensors”. Technology integration aspects included “IoT”, “blockchain”, “robotics”, “automation”, and “edge computing”, with explicit emphasis on studies where these technologies function as integral components of AI-driven systems in agri-food engineering. This ensures the search focused on integrated frameworks, for example, AI-powered robotics for precision farming, as well as IoT sensors feeding data to machine learning models for crop monitoring rather than standalone implementations of enabling technologies without AI integration. The core focus remained on AI methodologies as the driving force, with enabling technologies considered only in their role of supporting or enhancing AI applications.

Boolean operators and wildcards were strategically employed to optimize search sensitivity while maintaining specificity. Search terms were adapted to accommodate the specific syntax requirements of each database platform to ensure comprehensive coverage.

3.2. Inclusion and Exclusion Criteria

The study employed stringent criteria to ensure the selection of high-quality, relevant research. Except for a few early classic literature sources on the development of AI, the search time window was restricted from 1 January 2015 to 1 May 2025, and the last search was conducted on 31 May 2025 to ensure the timeliness of the included studies. Eligible studies must center on AI technologies, including machine learning, deep learning, and AI-driven sensing analysis as their core innovation in agri-food engineering; those involving enabling technologies such as IoT or robotics must explicitly demonstrate how these technologies interface with or support AI functionalities, forming an integrated system rather than operating independently. Preference was given to research providing quantitative performance metrics or comparative analysis of AI methodologies, with clear alignment between the enabling technologies described and the AI applications they support.

Exclusion criteria targeted non-peer-reviewed publications, including preprints and gray literature sources, along with book chapters, editorials, and opinion pieces lacking empirical content. Publications with insufficient relevance to AI applications in agriculture, purely theoretical frameworks without practical validation, duplicate studies, and research with inadequate methodological detail were systematically excluded.

3.3. Screening and Selection Process

The literature screening process followed a systematic four-stage approach to ensure unbiased article selection. Initial database searches identified 386 potentially relevant articles. Title and abstract screening eliminated clearly irrelevant publications, resulting in 364 articles for detailed evaluation. Full-text assessment involved comprehensive evaluation against inclusion and exclusion criteria, considering methodological rigor, relevance to research questions, and contribution significance. Reference list reviews of selected articles identified additional relevant publications through backward citation analysis. Quality assessment evaluated methodological soundness, experimental design, statistical analysis, and result presentation.

We also evaluated the rigor and validity of the selected studies based on the following quality criteria:

• Clarity of objectives: Are the aims and research questions of the study explicitly stated and well motivated?
• Appropriateness of design: Does the research design comprehensively address the relevant aspects of AI applications in agriculture (e.g., data sources, sensing modalities, algorithms, deployment contexts) and enabling technologies (e.g., IoT, connectivity, security protocols)?
• Methodological transparency: Are the methods, models, and experimental setups described in sufficient detail to enable reproducibility?
• Evidence and benchmarks: Are the datasets, benchmarks, and evaluation metrics credible, representative, and aligned with the stated objectives?
• Validity of conclusions: Are the findings supported by adequate data analysis, and do the conclusions logically follow from the presented evidence?

These criteria guided our appraisal to ensure that only high-quality and methodologically sound studies were synthesized in this review. After applying the eligibility criteria and conducting the screening process, the final set of selected studies were obtaine and are shown in Figure 5.

The final corpus comprised 175 articles deemed most relevant and methodologically sound for comprehensive analysis, representing the highest quality research in AI applications across agri-food engineering domains. The selected literature sources were systematically categorized using a comprehensive framework encompassing multiple analytical dimensions. Application domains included crop monitoring, quality assessment, safety analysis, precision agriculture, supply chain management, and environmental monitoring. AI methodologies covered machine learning algorithms, deep learning architectures, computer vision techniques, spectroscopic methods, and ensemble approaches. Technology integration aspects examined sensing platforms, IoT systems, robotics applications, blockchain implementations, and edge computing solutions.

This systematic methodology provides a robust foundation for analyzing current developments, emerging trends, and future prospects in AI applications for agri-food engineering, ensuring that conclusions rest upon rigorous analysis of high-quality, relevant research contributions.

4. The Application of AI in Agri-Food Engineering

This section presents a comprehensive overview of the current landscape of applications of AI methods in the agri-food engineering domains.

4.1. Agricultural Product Quality Monitoring

AI has fundamentally transformed agricultural product quality monitoring, shifting from labor-heavy manual checks to advanced real-time, non-destructive automated systems. These systems use AI algorithms to analyze complex multidimensional data beyond human sensory ability, delivering objective quality assessment across the agricultural value chain. Figure 6 shows the distribution of the number of papers cited in this section. This trend highlights a collective research shift toward complex, automated feature extraction and modeling for solving problems primarily in quality inspection and material classification.

4.1.1. Ingredients and Condition Inspection

In the early stage, spectroscopic methods combined with chemometrics [31,32,33,34,35,36] were commonly used for food quality assessment. Later, traditional machine learning approaches [37,38,39,40] became more popular, and in recent years, deep learning architectures [41,42,43,44,45] have gained prominence, particularly due to their effectiveness in handling large datasets and enabling multimodal sensor fusion for ingredient analysis and condition inspection. These methods differ in their comparative effectiveness, with varying accuracy ranges, computational requirements, and application suitability.

To measure food quality as an efficient, effective, rapid, and inexpensive technique for fast detection, Near-infrared Spectroscopy (NIRS) plays a key role in the identification of food standard for the evaluation of food quality and safety. The literature uses NIRS, which is simple, fast, non-destructive, accurate, and reliable, combined with Partial Least Squares Regression (PLSR) in stoichiometry, without destroying the original detection substances. In [32,34], Tahir et al. combined PLSR with NIRS, FT-IR, and Raman spectroscopy to model antioxidant compounds in roselle and honey, achieving high correlations across multiple indices ($R^2>0.9$), thereby confirming the effectiveness of PLSR in antioxidant capacity evaluation. In addition, studies [35,36] applied Competitive Adaptive Reweighted Sampling (CARS), Partial Least Squares (PLS), and RF-PLS to predict phenolic and amino acid contents in tea and matcha samples, respectively, achieving excellent modeling performance, which highlights the broad applicability of the PLS framework across different agricultural products.

Machine learning involves monitoring of the agri-food product. In [39], a multivariate artificial olfaction system based on a colorimetric sensor array was developed, and the integration of the K-Nearest Neighbor (KNN)–AdaBoost algorithm enabled 100% accurate identification of volatile changes during black tea fermentation. Moreover, in [40], integrated differential pulse voltammetry (DPV) with machine learning algorithms such as ANN, KNN, and Support Vector Machine (SVM) for both geographical origin identification and quantitative pungency prediction of Sichuan pepper. The models achieved 100% accuracy on the test set, surpassing conventional NIR-based methods while also exhibiting strong robustness and generalization capability.

Deep learning techniques also have proven effective in various fields. Studies [41,42] further extended to ANN and Convolutional Neural Network (CNN) frameworks, enabling nonlinear modeling and image visualization for honey adulteration and matcha moisture prediction, where the RPD values reached up to 3.22. Specially, The network architecture in [41] employed was mainly a Multilayer Perceptron (MLP), with the optimal one being MLP (five input neurons using the first five components from UV-VIS and NIR spectra that account for over 99% variance, nine hidden neurons with Tanh activation, and six output neurons corresponding to six physical and chemical properties with Exponential activation), which was trained via the backpropagation algorithm on a dataset split of 70:15:15 (training:test:validation) and achieved validation correlation coefficients over 0.8 for most properties except moisture content. In [43], transfer learning was combined with 11 pretrained CNN models to accurately recognize pollen images from desert regions under small sample conditions. Among them, the ResNet101 model achieved a classification accuracy of 97.24%, significantly outperforming shallow models and custom CNN architectures. Additionally, the study [44] demonstrated the feasibility of combining visible–near-infrared transmission spectroscopy with an ANN-PLS algorithm for pineapple maturity and Sparse Subspace Clustering (SSC) prediction, achieving a validation R² of 0.76 and RMSEP of 0.7879 and highlighting the practical value of deep–shallow hybrid modeling approaches in fruit and vegetable component analysis. In [45], NIRS was combined with models such as decision trees and ANN to model the physicochemical properties of rice, where the RandTree model demonstrated superior stability and generalization ability due to its attribute selection mechanism.

In addition to the basic AI methods, multimodal fusion and feature selection methods also have great application scenarios in the field of food quality inspection. The literature [46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61] has offered key approaches to enhancing classification and prediction accuracy about complex perception scenarios, as well as multimodal information fusion. In [62], image features and odor response data of tomatoes were integrated to construct support vector classification and regression models, achieving a maturity classification accuracy of 94.20% and a hardness prediction R² of 0.9514. In [50,53], machine vision and electronic nose systems were further combined to build multimodal recognition systems, significantly improving the accuracy of freshness evaluation in spinach and browning detection in potatoes, with fusion-based models achieving up to 100% accuracy—substantially outperforming models based on single-sensor channels. A similar strategy was employed in [48] for tomato SSC prediction, where IRIV-CS variable selection and Support Vector Regression (SVR) modeling yielded a calibration R² as high as 0.9845, surpassing that of conventional SVR and PLSR approaches. In [46], mango color features and volatile information were fused, enabling the SVC model to achieve a classification accuracy of 97.5% for quality grading, while the SVR model reached correlation coefficients of 0.9241 and 0.8897 for TSS and hardness prediction, respectively, indicating significant performance gains. The study [49] combined mid-infrared spectroscopy with an improved Fuzzy C-means (FCM) algorithm, overcoming the inefficiency and destructive nature of traditional physicochemical methods and enabling 100% non-destructive identification of tea samples.

In terms of feature selection and model optimization, the study [57] compared four feature selection algorithms—Variable Importance in Projection (VIP), Successive Projections Algorithm (SPA), CARS, and Genetic Algorithm (GA)—and confirmed that the CARS-PLSR model performed best in predicting the moisture content in the red meat. This finding highlights the critical role of variable selection strategies in determining model performance for red meat analysis.

Across ingredient and condition inspection applications, three distinct performance tiers emerge: (1) PLSR-based spectroscopic methods achieve baseline accuracies of 85–93% with minimal computational overhead, making them suitable for single-compound quantification; (2) traditional ML approaches demonstrate 94–97% accuracy in classification tasks but require extensive feature engineering; and (3) deep learning methods reach 97–100% accuracy, particularly when leveraging transfer learning, though at 10–50× the computational cost. Critically, multimodal fusion strategies consistently outperform single-sensor approaches by 3–8 percentage points, with optimal results observed when combining spectroscopic data with image or volatile compound features. However, performance gains plateau beyond three sensor modalities, and real-world deployment remains constrained by the need for specialized hardware and domain-specific calibration datasets. Feature selection methods, particularly CARS and IRIV, prove essential for maintaining model parsimony, reducing dimensionality by 60–85% while preserving more than 95% of predictive power. According to the above research, multimodal fusion technology has achieved remarkable results in the field of food quality inspection, with significantly better performance than single sensor solutions. This is due to the organic combination of spectral technology and machine learning algorithms, as well as the synergistic effect of multi-sensor information.

4.1.2. Classification Identification

Beyond ingredient analysis and condition assessment, classification identification also represents a critical application domain in agricultural product analysis. In the early stage, Hyperspectral Imaging (HSI) techniques combined with chemometrics methods were widely applied in agricultural product classification [63,64,65,66,67], while in recent years, machine learning and deep learning approaches such as E-nose systems and multimodal fusion models have become more popular [68,69].

The combination of HSI and machine learning has become a particularly effective classification method. Sun et al. combined visible–near-infrared HSI with SPA and principal component analysis (PCA) for feature extraction, achieving 98.33% accuracy in black bean variety classification using Partial Least Squares Discriminant Analysis (PLS-DA) [63]. Wu et al. further advanced feature extraction methodologies by proposing a supervised discriminant analysis (SDA) algorithm that outperformed conventional PCA+Linear Discriminant Analysis (LDA) approaches for near-infrared spectral classification [64]. The integration of unsupervised learning frameworks was demonstrated by Wu et al., who employed a Fuzzy Discriminant C-means (FDCM) clustering model to achieve 97% classification accuracy for apple varieties [65].

Multi-source feature fusion has proven particularly effective in fine-grained classification tasks. Jin et al. developed a MSC-PCA-SVM model that fused spatial and spectral features from hyperspectral images, achieving 97.64% accuracy in wheat seed variety identification [66]. The potential of intelligent optimization algorithms was further demonstrated by Sun et al., who utilized the Artificial Bee Colony (ABC) algorithm to optimize SVM performance, achieving 100% classification accuracy for watermelon seed vigor assessment in 4.19 s [67].

Traditional meat detection methods face limitations, including low detection speed, poor field adaptability, and limited modeling flexibility. Electronic nose systems combined with supervised learning algorithms have emerged as promising alternatives for intelligent meat classification. Setyawati et al. developed a metal oxide sensor array-based E-nose system, where the LDA model under “five-window + maximum value feature” configuration achieved 99.9% validation accuracy and 100% test accuracy for beef, chicken, and pork floss classification [68].

Biomarker-based approaches have also shown considerable promise. Garcia et al. employed fatty acid and volatile compound data to develop classification models for light lamb carcasses, with both KNN and ANN achieving 98% accuracy [69]. The ANN architecture in the study adopted a Multilayer Perceptron paradigm with an input layer, three hidden layers (150, 100, and 50 neurons, respectively, each using ReLU activation), an output layer with 3 neurons (for three lamb categories, using Softmax activation), 500 training epochs, Adam optimizer, and dropout layers after the first two hidden layers to prevent overfitting. Notably, the ANN model achieved equivalent performance using only 7 variables compared to the 13 required by KNN, demonstrating superior feature compression capabilities. The relatively lower SVM performance (86%) suggests that nonlinear neural networks and KNN are better suited for handling complex biochemical attribute correlations under constrained feature dimensions. These studies collectively demonstrate that integrating electronic nose systems or biochemical indicators with machine learning approaches significantly enhances both classification accuracy and practical deployability in meat product identification applications.

Table 2. Applications of AI methods in agricultural products.

Product	Parameter	Best Model	Reference
Honey	Antioxidation	LDA	[33]
	Antioxidation	PLSR	[34]
	Adulteration Detection	ANN	[41]
Apple	Classification	PCA-LDA	[64]
	Classification	FDCM	[65]
Black Tea	Fermentation Parameters	CARS-PLS	[35]
	Fermentation Parameters	KNN-AdaBoost	[39]
	Pesticide Residue	1D-CNN-RF	[70]
Green Tea	Tea Polyphenol	ACO-ELM	[52]
	Approbation	PFCM	[49]
	Quality Rating	PCA-LDA	[38]
	Quality Rating	VISSA-SVM	[51]
	Moisture Content	CARS-SVR	[58]
Maize	Moisture Content	CARS-SPA-LS-SVM	[54]
	Moisture Content	CARS-PLSR	[57]
	Moisture Content	VIP-GA	[56]
	Non-Destructive Inspection	PLS-DA	[71]

presents the applications of AI methods in several agricultural products. Through the above research, it can be seen that the fusion technology of HSI and machine learning has achieved remarkable results in the field of agricultural product classification and recognition, with classification accuracy generally reaching over 95% and significantly better performance than traditional detection schemes. This is due to the technological advantages of spectral spatial feature fusion, intelligent optimization algorithms, and multi-sensor collaboration. However, HSI methods are still limited by high equipment costs, complex data processing, and real-time requirements, and further improvement is needed in large-scale application scenarios.

4.2. Agricultural Product Safety Analysis

AI is a critical tool for agricultural product safety analysis, addressing global food supply concerns over chemical residues, microbial contamination, and foodborne pathogens. Paired with advanced techniques like SERS and mass spectrometry, it revolutionizes detecting and quantifying health risks. AI-driven systems use advanced machine learning to predict microbial growth and assess toxicological risks with sensitivity and specificity exceeding traditional lab methods. Figure 7 shows the distribution of the number of papers cited in this aspect. It indicates a clear research trend toward leveraging sophisticated, data-driven models to solve pressing issues in crop disease management and environmental pollution monitoring, ultimately aiming to enhance the safety and quality of agricultural products.

4.2.1. Disease Detection

In the early stage, traditional approaches such as manual inspection, image recognition, and spectral analysis were commonly used for crop disease detection [72,73,74,75,76,77], while in recent years, advanced sensing techniques integrated with machine learning and deep learning models have become more popular, enabling the development of robust automated quality assessment and intelligent diagnostic systems for agricultural products [78,79,80,81,82,83,84].

Contemporary research has demonstrated the superior performance of deep learning architectures in plant disease identification. Pandian et al. developed a 14-layer DCNN for multi-crop disease classification, achieving 99.97% accuracy across 58 disease types spanning 16 crops through strategic data augmentation using DCGAN and NST techniques [79]. The integration of advanced segmentation methods has further enhanced diagnostic precision, with Bukhari et al. demonstrating that U²-Net achieved 95.74% segmentation accuracy for wheat stripe rust lesions, which improved to 96.20% when combined with ResNet-18 [81]. Long et al. developed the CerealConv model specifically for cereal crops, achieving 97.05% classification accuracy and surpassing pretrained CNN networks, thereby validating the effectiveness of incorporating crop-specific domain knowledge [82].

The fusion of spectral and imaging data has proven to be highly effective for disease detection. Lu et al. combined HSI with preprocessing techniques and variable selection strategies (CARS, IRIV, GSA), enhancing the SVM classification accuracy to 98.33% [75]. Similarly, Xu et al. developed a dual-channel imaging system combining texture and polarization features, achieving 96.67% accuracy in citrus Huanglongbing identification through Random Forest modeling [76].

Transfer learning has demonstrated significant potential for cross-variety disease detection. Wu et al. employed DCNN architecture for rice bakanae disease identification across six cultivars, achieving >96.5% accuracy across all varieties, with a peak performance of 99.02% on the YY cultivar [85]. For small sample scenarios, Xu et al. proposed a dual transfer learning framework based on Vision Transformer, achieving 86.29% accuracy in 20-shot settings—representing 13% and 37% improvements over ResNet and MAE, respectively [86].

Overall, while early image- and spectrum-based approaches typically achieved accuracies in the range of 85–90%, the integration of deep learning, spectral–imaging fusion, and transfer learning frameworks now consistently delivers performance above 95%, with some models exceeding 99%, underscoring a clear paradigm shift toward robust, multimodal, and highly accurate intelligent crop disease detection. The evolution from single-modal visual feature analysis to multidimensional information fusion represents a paradigmatic shift in agricultural diagnostics. Current approaches effectively integrate imagery, spectral data, texture, and polarization features through sophisticated deep learning frameworks, transfer learning mechanisms, and data augmentation strategies. This integration has substantially improved model robustness against environmental variations, noise interference, and disease heterogeneity.

4.2.2. Pesticide Pollution Detection

In the early stage, pesticide residue detection in agricultural products mainly relied on traditional chemical analysis and feature selection techniques [87,88,89], while in recent years, non-destructive methodologies such as HSI, NIRS, surface-enhanced Raman spectroscopy (SERS), and computational image analysis integrated with ensemble learning and multi-band fusion models [70,90,91,92,93,94,95] have become more popular, enabling accurate identification and quantitative prediction of pesticide types, concentrations, and complex residue mixtures.

Early pioneering work by Sun et al. established multi-region hyperspectral feature extraction strategies for lettuce pesticide residue detection, utilizing wavelet transform and multiple feature selection algorithms to enhance SVM classifier performance, achieving 98.33% prediction accuracy [87,88]. Alternative spectroscopic modalities have shown comparable effectiveness, with Xin et al. demonstrating 97.78% prediction accuracy using polarized spectroscopy combined with SVM modeling [89].

Algorithm optimization has increasingly focused on ensemble learning and deep neural networks. Feng et al. employed LC-HRMS data with ensemble models (XGBoost, RF) for pesticide retention time prediction, achieving sub-minute error rates for over 90% of samples [90]. The integration of lightweight CNNs with traditional ensemble methods has proven particularly effective, as demonstrated by Sun et al., whose CARS-SPA feature extraction combined with 1D-CNN-RF ensemble achieved 99.05% test accuracy [70].

Multi-band fusion strategies have yielded significant improvements in detection performance. Hu et al. demonstrated that Short Wave InfraRed (SWIR) input to 1D-CNN models improved the accuracy results from 89% (single band) to 94%, with an F1 score of 0.94, confirming enhanced SWIR sensitivity to organic molecules [91]. Portable implementation has shown remarkable success, with Lapcharoensuk et al. achieving 100% accuracy in chlorpyrifos detection using compact NIR devices coupled with SVM and PC-ANN models [92].

As shown in

Table 3. Machine learning algorithms for agricultural product safety detection.

Reference	Modeling Algorithm	Accuracy on the Prediction Set
Sun2017 [87]	CMS-WT-SVM	100%
Sun2018 [88]	CARS-IRIV-GSA-SVM	98.33%
Xin2018 [89]	CARS-SVM	97.78%
Lapcharoensuk2023 [92]	SVM-PC-ANN	100%
Zhou2019 [96]	VISSA-GOA-SVM	98.57%

, most of the commonly used machine learning algorithms for agricultural product safety detection are based on SVM techniques, with a smaller proportion employing ANN. These AI-based methods have demonstrated strong capabilities in both classification and prediction tasks.

The convergence of SERS with deep learning represents a significant advancement in highly sensitive pesticide detection. Wang et al. developed a SERS-CNN ensemble platform that achieved 99.62% accuracy across 10 pesticide types [93], while Hajikhani et al. introduced the SERSFormer model, utilizing the Transformer architecture to achieve 98.4% classification accuracy and R² = 0.849 for quantitative prediction [94]. The integration of SERS with CNN and similarity algorithms on Thin-Layer Chromatography platforms has enabled zero-error recognition of mixed pesticide samples, demonstrating suitability for complex matrix analysis [95].

Heavy metal detection has emerged as a complementary application area for hyperspectral-based intelligent systems. Zhou et al. developed a comprehensive framework incorporating data-level fusion and grasshopper optimization algorithm optimized SVM classifiers for cadmium stress classification in lettuce, achieving 100% training and 98.57% prediction accuracy [96]. Advanced deep learning approaches have shown exceptional performance in multi-contaminant scenarios, with Zhou et al. combining wavelet transform with stacked autoencoders and SVR for simultaneous Cd-Pb concentration prediction, achieving an Rp² > 0.93, an RMSEP < 0.05 mg/kg, and an RPD > 3 [97]. Some popular models used for the residue detection of lettuces are summarized in

Table 4. Applications of AI methods in residue detection of lettuce.

Parameter	Best Model	Reference
Cyanidin	DBO-ELM	[60]
Pesticide Residue	CMS-WT	[87]
Pesticide Residue	CARS-IRIV-GSA-SVM	[88]
Pesticide Residue	CARS - SVM	[89]
Pesticide Residue	PC-ANN-SVM	[92]
Heavy Metal Residue	VISSA-GOA-SVM	[96]
Heavy Metal Residue	WT-SCAE-SVR	[97]

.

To conclude, the intelligent detection method based on multimodal fusion has achieved remarkable results in the field of pesticide residue detection, with the detection accuracy generally exceeding 97%, offering significantly better performance than traditional chemical analysis schemes. This is due to the application of HSI and high-sensitivity spectroscopic technologies such as SERS, the adoption of multi-band information fusion strategies, and the advantages of deep learning algorithms in feature extraction and pattern recognition. However, intelligent detection methods are still limited by factors such as high equipment costs, complex spectral data processing, and adaptability to different substrate samples, and they still face technical and economic feasibility challenges in large-scale practical applications.

4.3. Agricultural Production Process Management

AI has revolutionized agricultural production process management by integrating IoT sensors, satellite imagery, and real-time analytics to optimize decision making across the entire production lifecycle [98]. These systems enable environmental monitoring, production optimization, and supply chain coordination via intelligent automation and predictive modeling. AI-driven platforms use machine learning to analyze interactions between environmental variables, resources, and market dynamics, empowering farmers to make proactive decisions and mitigate risks. They boost resource efficiency, optimize yields, support sustainability, and drive a shift toward holistic, intelligent agricultural operations that balance productivity with environmental protection [99]. As depicted in Figure 8, which shows the publication distribution for AI in agricultural management, machine learning and deep learning are the dominant paradigms driving innovation. The convergence of AI with IoT and optimization algorithms further underscores a pivotal shift toward developing interconnected and intelligent agricultural systems.

4.3.1. Environmental Monitoring

In the aspect of agricultural environmental monitoring, the current research mainly focuses on sensor integration, real-time parameter prediction, and other fields of the Internet of Things. Multi-source data-driven machine learning models have emerged as the predominant approach to agricultural environmental monitoring, demonstrating substantial improvements in spatial accuracy, computational efficiency, and model generalizability compared to traditional physical and statistical methods [100,101,102,103,104,105]. Contemporary developments emphasize IoT integration for precision agriculture, with continuous high-resolution monitoring of critical parameters, including soil health, crop growth, and environmental conditions.

Remote sensing-based approaches have achieved remarkable success in spatial downscaling and environmental parameter estimation. Park et al. demonstrated the effectiveness of Random Forest algorithms for fusing multi-source remote sensing data, enabling downscaling of 25 km AMSR2 soil moisture data to 1km resolution, with R² = 0.96 and Root Mean Square Error (RMSE) = 0.06, representing over twofold accuracy improvement compared to standard regression methods [100]. Ecosystem-scale applications have shown equal promise, with Raza et al. developing a machine learning ensemble approach (MLEA) augmented with SHAP-based interpretability for net ecosystem exchange estimation in tea plantations, achieving superior performance over traditional single-model approaches while eliminating dependence on expensive instrumentation [104].

The integration of IoT technology and machine learning algorithms into agricultural monitoring systems has enabled intelligent control systems that enhance efficiency and sustainability. These frameworks utilize low-cost sensor networks incorporating nitrogen, phosphorus, and potassium (NPK), as well as electrochemical and hyperspectral sensors within IoT infrastructure to acquire comprehensive environmental data [103,106,107,108,109,110,111].

Lightweight AI models, including fuzzy logic systems, PLSR, and gradient boosting algorithms, are deployed at edge or cloud levels to enable real-time assessment capabilities. Machine learning is employed in multi-phase processes to optimize resource management while tracking critical growth parameters. Khaydukova et al. developed an integrated NPK analysis system combining electronic tongue sensor arrays with PLSR, achieving prediction errors within ±5% for major nutrients while eliminating laboratory-based testing requirements [103]. Similarly, Lavanya et al. implemented an IoT-enabled fertilization platform incorporating fuzzy logic decision-making engines, achieving 91% classification accuracy with sub-5-second response latency through colorimetric NPK sensing and embedded processing units [107].

The combination of remote sensing technology and machine learning algorithms provides comprehensive solutions for precision agriculture challenges, with predictive analytics and real-time decision-making capabilities optimizing resource efficiency. These integrated approaches support data-driven agricultural decision making through continuous monitoring, automated analysis, and responsive control mechanisms, establishing foundations for sustainable and precision-oriented farming systems. And machine learning methods driven by multi-source data have achieved remarkable results in the field of agricultural environmental monitoring, with prediction accuracy generally exceeding 90%, spatial resolution increased by more than two times, and performance results being significantly better than traditional physical and statistical methods due to the deep integration of lightweight AI models at edge remote sensing technology and IoT sensors.

4.3.2. Optimization of Production Decision Making

In terms of agricultural decision support system, current research focuses on precision fertilization, intelligent irrigation, and other fields, and common technical means include machine learning [112,113,114,115,116] and edge computing [117,118,119,120,121].

Machine learning and expert system-based approaches have become extensively adopted in agricultural decision support, demonstrating superior performance over traditional experience-based methods in precision and computational efficiency across fertilization scheduling, irrigation control, and crop selection applications [112].

Data-driven modeling approaches have shown significant potential for optimizing agricultural inputs and operations. Liao et al. developed regression-based predictive models for droplet size estimation in air-induced nozzle systems, enabling precision pesticide application without labor-intensive field trials [113]. Expert system integration has proven particularly effective for smallholder agriculture, with Firmansyah et al. implementing a rule-engine-driven system that integrates soil composition, crop type, and climatic data through Geographic Information System (GIS) inputs and forward-chaining logic, achieving <8% recommendation error and <10-second decision latency [114].

Comprehensive smart farming platforms have emerged that integrate multiple ML and DL algorithms for holistic agricultural management. Durai et al. constructed a platform encompassing crop recommendation, pest detection, fertilization planning, and cost prediction modules, achieving 97.5% accuracy in crop selection and 99.56% in fertilization guidance through multimodal data fusion [115]. Advanced hybrid architectures have addressed class imbalance challenges, with Bhat et al. proposing a DNN framework combining gradient boosting and Bayesian optimization with SMOTE-Tomek resampling, demonstrating robust performance across 12-class soil–crop suitability classifications [116].

The deployment of intelligent agricultural terminals and edge computing solutions has enabled real-time decision making in resource-constrained environments. Thorat et al. introduced the TPFCNN model integrating visual and soil sensor data for pest detection and fertilizer recommendation on low-power devices, achieving 91% classification accuracy with reduced latency compared to traditional methods [117]. Computer vision integration with robotic systems has enhanced operational precision, as demonstrated by Zhu et al.’s YOLO-based variable-rate spraying system for foliar fertilization, enabling adaptive responses to heterogeneous field conditions [118]. Elbeltagi et al. developed machine learning ensembles for vapor pressure deficit prediction in semi-arid regions, with Random Forest models achieving 0.9694 test correlation coefficient results for intelligent irrigation applications [119]. Edge computing implementations have demonstrated substantial resource efficiency improvements, with lightweight neural network and IoT-based systems achieving 30% water usage reduction in precision irrigation control [120,121].

Contemporary developments in agricultural decision support systems demonstrate the convergence of edge computing, multimodal sensing, rule-based reasoning, and ensemble modeling approaches. These integrated frameworks provide substantial advantages in scalability, efficiency, and responsiveness, offering promising foundations for expanding intelligent agriculture adoption, particularly among smallholder farming operations, where resource optimization and decision support are most critical. Although intelligent decision making methods have demonstrated excellent performance, these systems are still limited by factors such as insufficient model interpretability, strong dependence on data quality, and low technical acceptance among small farmers. Therefore, further reductions in technical barriers and usage costs are needed to facilitate their widespread promotion and application.

4.3.3. Supply Chain Management and Control

The convergence of blockchain technology and machine learning has catalyzed the widespread adoption of smart contract-driven ensemble learning methods for fraud detection and supply chain transparency. These integrated approaches demonstrate superior performance over conventional methods in accuracy, operational efficiency, and system transparency in financial transactions, healthcare claims, food authentication, and product quality monitoring [122,123,124,125,126,127].

Smart contract integration with machine learning models has achieved remarkable success in fraud detection applications. Ashfaq et al. integrated XGBoost and Random Forest models with blockchain-based smart contracts for Bitcoin fraud detection, achieving 98.7% classification accuracy and 0.92 AUC while addressing data sharing inefficiencies [122]. Explainable AI frameworks have shown particular promise for food safety applications, with Bhatia et al. implementing a Faster R-CNN architecture combined with blockchain and QR codes, achieving 99.53% accuracy in harmful substance detection with <360 ms inference latency under high-throughput conditions [126].

Blockchain-enabled traceability systems leverage IoT sensors and computer vision for comprehensive supply chain monitoring, applying ML models for anomaly detection and quality assessment while recording both raw data and model outputs on-chain [124,128,129,130]. Wu et al. developed the MBITTS system by integrating GBDT models (86% recall, F1 score 0.92) for temperature anomaly detection in tea cultivation and GPS signal drift monitoring during logistics, enabling full lifecycle traceability [124].

TinyML integration has enhanced on-device anomaly detection capabilities, with Tsoukas et al. combining blockchain with edge computing for temperature and humidity monitoring in cold-chain logistics, achieving F1 scores of 0.85 for equipment tampering detection [128]. Computer vision and machine learning methods have further advanced agricultural automation, with Wang et al. implementing SLIV-SVM (98.5% accuracy), TeaNet (100% accuracy), and SVM (97.5% accuracy) for comprehensive tea cultivation and processing workflows [131].

Privacy-preserving frameworks reconcile data sharing requirements with confidentiality constraints through federated learning and advanced optimization algorithms [125,130,132,133]. Federated learning implementations have demonstrated effectiveness under data isolation conditions, with Gavai et al. developing federated Bayesian networks achieving 0.90 AUC for fraud prediction and yielding superior sensitivity compared to centralized approaches [125]. Lei et al. implemented hybrid homomorphic encryption and differential privacy approaches, achieving GDPR compliance (L1 accuracy ±2 °C), a 92% traceability accuracy in cold-chain logistics, and a 35% reduction in cross-chain interoperability costs [130]. Advanced optimization frameworks have enhanced cryptographic security, with Aljabhan et al. proposing PCGSO-based systems reducing tamper sensitivity by 40% compared to conventional approaches while maintaining a 0.5% correlation between sanitized data and encryption keys [132].

Federated learning implementation in smart agriculture introduces significant network security vulnerabilities affecting both data integrity and operational safety. Non-independent and identically distributed (non-IID) data distribution degrades model accuracy and creates security vulnerabilities, while equipment heterogeneity exacerbates communication inefficiencies and creates additional attack vectors [134]. Critical infrastructure vulnerabilities have been identified in precision agricultural equipment, particularly through CAN bus exploitation. Freyhof et al. demonstrated how adversaries can manipulate fertilizer application parameters via CAN bus attacks, resulting in measurable yield reductions and economic losses that become compounded by limited cybersecurity awareness among agricultural operators [135].

Contemporary research has proposed various security enhancement approaches, including Local Adaptation Federated Learning models for improved IoT intrusion detection [136] and control flow certification with encryption protocols for model update integrity [137]. However, critical vulnerabilities persist in agricultural monitoring systems, particularly regarding secure transmission of sensor and remote sensing data, requiring enhanced protection against data leakage and tampering threats [138].

According to the above research, the intelligent supply chain method based on the integration of blockchain and machine learning has achieved remarkable results in the field of agricultural traceability and safety supervision. Its performance is significantly better than traditional centralized management solutions due to the deep integration of smart contracts and machine learning models, the collaborative training advantage of federated learning under privacy protection, and TinyML’s real-time anomaly detection capability at the edge. However, the blockchain intelligent supply chain system is still limited by factors such as network security vulnerabilities, data distribution heterogeneity, and device compatibility, and it faces the dual challenges of technical complexity and security protection in large-scale agricultural infrastructure deployment.

5. Existing Challenges

The widespread adoption of AI-driven technologies in agri-food engineering faces multifaceted challenges. Understanding these barriers is crucial for developing inclusive and responsible AI systems that can effectively contribute to global food security while promoting equitable development.

5.1. Dataset Construction Challenges

The core challenge in agricultural AI applications lies in constructing high-quality datasets. Agricultural data exhibit significant spatiotemporal heterogeneity characteristics, with enormous variations in climatic conditions, soil types, and crop varieties across different geographical locations. This diversity makes it extremely difficult to build standardized datasets with universal applicability [139].

Table 5. Existing deficiencies in agri-food datasets.

Dataset Name	Main Deficiency
Carrot-Weed [141]	Small scale (only 39 images)
Plant Seedlings [142]	Only image-level annotation (no target location information)
Synthetic SugarBeet Weeds [143]	Algorithm generation (non-real scene)
Leaf Counting [144]	Only leaf number classification is supported
DeepWeeds [145]	Only image-level labels (no pixel/target labeling)
Early Crop Weed [146]	Each graph contains only a single species.
Ladybird Cobbitry Brassica [147]	Without any label, the amount of data is very large (>2.8 TB)
MangoNet [148]	The user needs to cut the high-resolution image (4000 × 3000 → 200 × 200) by himself
3D Broccoli [149]	No annotation; only original point cloud and video data are provided
Apple Trees [150]	Only raw depth and color images are available
Sugarcane Billets [151]	Only image-level labels (no target location/segmentation labeling)
Oil Radish Growth [152]	The test set is unannotated (34 images)

shows the existing challenges in high quality datasets. It can be seen that many existing studies rely on datasets collected from specific regions or crop types, which restricts the generalization of their models across diverse geographical areas, such as soil parameter models that may only apply to certain climatic zones. Moreover, there is a notable scarcity of publicly available, annotated datasets. As a result, most research efforts rely on small-scale, self-collected data, leading to limited generalization across different crop species and disease types [140]. Therefore, it is of great importance to develop a geneal standard for agricultural datasets that could provide a wide variety of high quality data covering the whole scope of agricultural and food engineering, aiming to support the enhancement of the application of AI schemes.

5.1.1. Technical and Cost Barriers in Data Collection

Traditional agricultural data collection relies heavily on extensive manual field surveys and expensive specialized equipment, such as multispectral cameras and soil sensors, resulting in extremely high equipment procurement and maintenance costs. The seasonal and cyclical characteristics of agricultural data require long-term continuous monitoring, further escalating temporal costs. Additionally, the harsh agricultural environment leads to frequent equipment damage, with repeated maintenance and replacement adding operational burden [153].

5.1.2. Professional Knowledge Requirements and Annotation Challenges

Agricultural data annotation requires deep professional knowledge, particularly in tasks such as pest and disease identification and crop growth status assessment. Disease identification demands plant pathology expertise, as symptoms of different diseases often appear similar and require experienced specialists for accurate differentiation [72]. The limited number of agricultural experts and their uneven distribution make it difficult to meet large-scale data annotation demands, further exacerbating challenges in annotation costs and quality control [154].

5.1.3. Lack of Standardization and Interoperability

Current agricultural data lack a unified standardization system, with significant differences in data formats and annotation standards adopted by different research institutions and enterprises, severely affecting data sharing and reuse [153]. Although the International Organization for Standardization (ISO) is developing agricultural data standardization frameworks (ISO 11783 series standards), widespread adoption still requires time [155]. The agricultural industry chain is extensive, involving multiple segments with significant differences in data standards across segments, creating technical and institutional barriers to interdisciplinary integration [156].

5.1.4. Privacy Protection and Sharing Mechanisms

Agricultural data involve farmers’ commercial secrets, such as yield data and business decisions. The lack of effective privacy protection mechanisms and incentive sharing mechanisms limits the establishment of large-scale datasets. Federated learning technology provides new insights for addressing this issue, allowing for collaborative training without sharing raw data [157].

5.2. Algorithm Performance and Deployment Challenges

While AI adoption in agriculture has grown, its deployment remains constrained across critical domains—largely due to underlying technical barriers.

5.2.1. Computational Complexity and Real-Time Requirements

Agricultural AI applications need to process high-dimensional, multimodal data, such as high-resolution images and time-series sensor data, resulting in extremely high computational complexity [139]. Agricultural decisions have time-sensitive requirements, such as pest and disease warnings that need completion within 24 h, posing severe challenges to algorithm real-time performance. Deep learning models have massive parameter counts, making it difficult to meet real-time processing requirements with limited computational resources, while agricultural environments typically lack high-performance computing infrastructure [158].

5.2.2. Technology Acceptance Barriers for Smallholder Farmers

Approximately 80% of global agricultural operators are smallholder farmers with limited resources and low technology acceptance capacity, posing significant challenges for AI technology popularization [159]. Smallholder farmers face multiple difficulties, including economic resource constraints, insufficient technical literacy, and weak infrastructure, necessitating the development of low-cost, easy-to-operate AI solutions, such as smartphone-based crop diagnosis applications.

5.2.3. Few-Shot Learning Limitations

Agricultural data are influenced by multiple factors, including climate, soil, and crop types, making data collection difficult due to limited available training samples [160]. Long agricultural production cycles result in slow data accumulation, with extremely scarce data for low-frequency, high-impact events such as extreme weather events and emerging pest diseases. Strong regional differences in agricultural environments lead to poor model generalization across different regions, requiring substantial region-specific samples for adaptation [139].

5.2.4. Insufficient Explainability

Agricultural decisions directly relate to farmers’ economic benefits, requiring AI models to provide clear, understandable decision rationales [161]. However, deep learning models are typically viewed as “black boxes,” creating conflict between their complexity and agricultural decision making’s transparency requirements. Increasingly stringent regulatory requirements, particularly in food safety and environmental protection, demand complete transparency and traceability of algorithmic decision-making processes.

In short, AI deployment in agriculture faces interconnected technical challenges—from limited application in smallholder farming, aquaculture, and food processing to core bottlenecks like fragmented data, inefficient algorithms, poor field robustness, and high costs. While lightweight edge AI and cross-domain integration offer solutions, overcoming core technical hurdles like standardizing data, boosting algorithm adaptability, and optimizing cost–performance is key. Only by resolving these issues can AI move from labs to real agricultural scenarios, driving scalable intelligent transformation and supporting global food security and productivity goals.

5.3. Environmental Sustainability and Regulatory Framework

While AI promises transformative potential for agricultural efficiency, its environmental implications and regulatory landscape present complex considerations that require systematic analysis for responsible deployment.

5.3.1. Environmental Impact Duality

AI applications in agriculture present contradictory environmental effects that demand careful evaluation. On the positive side, precision agriculture systems can reduce fertilizer usage by 25% through targeted application, while AI-driven irrigation optimization can decrease water consumption by 30–50% in monitored fields [162]. Smart pest monitoring systems enable pesticide reduction of up to 20–40% by enabling precise timing and localized treatment [162], and yield prediction models help minimize food waste by optimizing harvest timing and supply chain planning.

However, AI deployment also creates environmental costs through substantial energy consumption from continuous data processing, cloud computing infrastructure, and sensor networks that require regular replacement. The carbon footprint of training large agricultural AI models can range from 17 to 78 CO₂-equivalent tons [163], while the embodied carbon in widespread sensor deployment across farms represents a significant but often overlooked environmental cost. These trade-offs require quantitative assessment to ensure AI adoption genuinely supports sustainability goals rather than creating technological solutions that offset their environmental benefits.

5.3.2. Regulatory Landscape Complexity

The regulatory environment for agricultural AI varies significantly across major markets, creating implementation challenges for global technology deployment. The European Union’s AI Act classifies certain agricultural applications as high-risk systems, requiring conformity assessments, risk management systems, and human oversight, particularly for food safety and environmental impact applications. GDPR implications extend to farm data collection and sharing, imposing strict consent and data protection requirements that affect AI system design [164].

The United States regulatory approach remains fragmented across agencies, with the FDA overseeing AI applications in food safety, USDA managing agricultural data policies, and EPA regulating environmental monitoring systems [165]. State-level regulations create additional complexity, with varying requirements for precision agriculture incentives and data sharing protocols. In India, the Digital India Agriculture initiative promotes AI adoption, while data localization requirements mandate domestic storage of agricultural data, and smallholder farmer protection measures impose specific safeguards for technology access and pricing [166].

5.3.3. Policy Implementation Gaps

Current regulatory frameworks often lag behind technological capabilities, creating uncertainty for AI developers and adopters. Cross-border data sharing restrictions limit the development of global agricultural intelligence systems, while inconsistent standards across jurisdictions increase compliance costs and implementation complexity [167]. Many regulations focus on data protection and algorithmic transparency but provide insufficient guidance on environmental impact assessment and sustainability metrics for AI deployment.

The absence of harmonized international standards for agricultural AI creates barriers to technology transfer and scaling, particularly affecting developing countries that could benefit most from AI-driven agricultural improvements. Regulatory gaps in liability frameworks for AI-driven agricultural decisions pose additional challenges for insurance and risk management in farming operations.

In summary, successful agricultural AI deployment requires balancing environmental benefits against computational costs while navigating complex, fragmented regulatory landscapes. Policymakers need comprehensive frameworks that promote innovation while ensuring environmental sustainability and farmer protection. Only through coordinated international standards and quantitative sustainability assessments can AI fulfill its potential to support global food security while advancing environmental stewardship goals.

6. Future Directions

The transformative potential of artificial intelligence in agriculture is increasingly evident, yet realizing this potential requires addressing fundamental challenges that span technical, practical, and societal dimensions. This section synthesizes critical future directions: the establishment of robust dataset construction and standardization systems to address current data limitations; the development and deployment of emerging AI technologies that can democratize access to intelligent agricultural solutions; and the enhancement of AI security and explainability to build stakeholder trust and ensure responsible technology adoption. Together, these directions form a comprehensive roadmap for advancing agricultural AI from experimental applications toward transformative solutions.

6.1. Dataset Construction and Standardization Systems

The development of robust AI systems for agricultural applications fundamentally depends on the availability of comprehensive, high-quality datasets that can capture the complexity and variability inherent in agricultural environments. However, as Section 5.1 shows, current agricultural AI research faces significant challenges related to data scarcity, heterogeneity, and a lack of standardization across different sources and formats. This subsection examines critical directions for addressing these foundational data challenges, including the establishment of standardized data collection frameworks, development of efficient annotation systems, creation of sustainable data sharing mechanisms, and utilization of synthetic data generation technologies to augment limited real-world datasets.

6.1.1. Multi-Source Data Integration and Standardization

Future efforts should establish standardized agricultural data collection and annotation systems, integrating multi-source heterogeneous data that include satellite remote sensing, drone imagery, IoT sensors, and meteorological data [154]. Through constructing unified data collection standards and quality control systems, the compatibility and comparability of data from different sources can be ensured. Automated data preprocessing technologies should be developed to achieve spatiotemporal registration and format unification of multi-source data, reducing data integration costs.

6.1.2. Crowdsourcing Annotation and Expert Knowledge Integration

Crowdsourcing annotation platforms should be developed to mobilize agricultural experts and farmers to participate in data annotation work, ensuring annotation accuracy through quality control algorithms. Hierarchical annotation mechanisms could be established, decomposing complex tasks into multiple simple subtasks to reduce annotation difficulty and improve efficiency. Active learning techniques may be utilized to optimize annotation strategies, prioritizing the annotation of samples that contribute most to model performance improvement, enhancing annotation resource utilization efficiency [168].

6.1.3. Data Sharing Incentive Mechanisms

Data sharing incentive mechanisms should be established to promote industry–academia collaboration and form large-scale, high-quality agricultural datasets. Privacy-preserving data sharing technologies could be developed, such as differential privacy and secure multi-party computation, to promote data circulation while protecting sensitive information [169]. Data value assessment systems and revenue distribution mechanisms can also be constructed to ensure data providers receive reasonable returns.

6.1.4. Synthetic Data Generation Technologies

Synthetic data generation technologies ought to be utilized to produce diverse training samples through simulation environments, alleviating real data scarcity issues. For example, Generative Adversarial Networks (GANs)-based agricultural image synthesis technologies could be developed to generate crop images at different growth stages and disease severity levels [170]. Digital twin models of agricultural ecosystems should be constructed to simulate agricultural production processes under different environmental conditions, providing rich virtual data for AI training.

6.2. Emerging AI Technologies in Agricultural Applications

As the agricultural industry seeks to balance productivity demands with sustainability goals while accommodating diverse stakeholder needs, emerging AI paradigms offer novel solutions that extend beyond traditional machine learning approaches. This subsection explores the key technological frontiers that are reshaping agricultural AI: the development of accessible, resource-efficient deployment solutions for smallholder farmers; the integration of large-scale foundation models to democratize agricultural expertise; and the application of meta-learning approaches to enhance system adaptability across diverse agricultural contexts and conditions.

6.2.1. Low-Cost Deployment Solutions

Addressing smallholder farmers’ resource limitations, lightweight AI models based on edge computing should be developed to reduce hardware costs and network dependencies. Smartphone-based agricultural AI applications are proposed to be developed, leveraging the ubiquity of mobile devices to achieve widespread technology adoption. Cloud-edge collaborative AI service systems should be constructed to reduce individual farmers’ technical barriers through resource sharing and load balancing [171].

6.2.2. Large Models Empowering Agricultural Knowledge Services

Agricultural-specific large language models should be constructed to integrate professional knowledge in the agricultural domain, providing intelligent technical consultation and decision support for farmers. Multimodal large models ought to be developed, combining text, images, videos, and other information forms to provide more intuitive and accurate agricultural guidance. Agricultural knowledge graphs could be established, utilizing large models’ reasoning capabilities to achieve digital inheritance and intelligent application of agricultural expert experience [172]. Natural language interfaces should reduce technical usage barriers, enabling farmers to obtain personalized agricultural advice through conversational methods.

6.2.3. Meta-Learning for Adaptability Solutions

Meta-learning-based agricultural AI systems is a potential solution to quickly adapt to agricultural needs under different regions, crops, and climatic conditions. Few-shot learning frameworks could be constructed to address AI applications in data-scarce scenarios such as rare pest identification and new variety characteristic prediction [173]. Meta-learning techniques should be utilized to achieve cross-regional and cross-crop knowledge transfer, improving AI models’ generalization and adaptability [174].

6.3. AI Security and Explainability Enhancement

As AI systems become increasingly integrated into critical agricultural decision-making processes, ensuring their reliability, transparency, and security has emerged as a paramount concern for widespread adoption and stakeholder trust. This subsection addresses the urgent need for developing trustworthy AI systems through three interconnected approaches: the creation of explainable AI frameworks tailored to agricultural domains, the implementation of comprehensive security assurance mechanisms to protect against vulnerabilities and attacks, and the establishment of robust ethical and regulatory frameworks to guide responsible AI deployment in agriculture.

6.3.1. Explainable AI Model Development

Agricultural-specific explainable AI frameworks are supposed to be developed, combining agricultural domain knowledge to construct easily understandable decision trees and rule sets. Attention mechanism-based visual explanation techniques should be developed to intuitively display image regions and features that models focus on [175]. Causal reasoning models ought to be constructed to reveal causal relationships among factors in agricultural production, providing scientific decision-making bases. Multi-level explanation systems should be established to provide appropriate explanation content and expression methods for different user groups [176].

6.3.2. Agricultural AI Security Assurance Technologies

Adversarial attack detection and defense technologies for agricultural AI are meant to be developed to ensure model stability under malicious attacks. Uncertainty quantification mechanisms for AI models should be established to provide confidence estimates for decisions, avoiding high-risk decision making. Security audit frameworks for agricultural AI systems could be constructed, establishing model performance monitoring and early warning mechanisms. Applications of federated learning and differential privacy technologies in agriculture should be developed to achieve collaborative learning while protecting farmer data privacy [172].

6.3.3. Ethical and Regulatory Frameworks

Ethical guidelines for agricultural AI applications should be established to ensure technology development aligns with sustainable agriculture and social equity principles. Responsibility attribution mechanisms for AI decision making could be developed, clarifying responsibility boundaries of AI systems in agricultural production [172]. Standardized testing and certification systems for agricultural AI should be constructed to ensure AI product quality and safety.

These directions will drive deep integration of AI technology with agricultural practices, injecting new momentum into modern agricultural development and ultimately achieving intelligent, precise, and sustainable agricultural production.

7. Conclusions

This review demonstrates the transformative potential of AI in agri-food engineering across quality monitoring, safety analysis, and production management. Multimodal fusion technologies combined with machine learning have enabled non-destructive, real-time assessment of agricultural products with remarkable accuracy improvements. AI-driven HSI and SERS technologies achieve over 95% classification accuracy in disease and contaminant detection, while IoT-AI integration optimizes environmental monitoring and supply chain operations.

However, critical obstacles impede widespread adoption. The scarcity of high-quality, standardized datasets represents the most significant bottleneck, particularly given the regional specificity of agricultural data and prohibitive annotation costs. This deficiency severely constrains model generalization across different geographic regions and crop varieties. Existing deep learning architectures remain computationally intensive and poorly suited for edge deployment in resource-constrained environments, creating a substantial gap between laboratory performance and field applicability. Economic and technical barriers disproportionately affect smallholder farmers who constitute 80% of global agricultural operators, as current AI technologies demand capital investments and technical expertise beyond their reach. Supply chain applications face persistent vulnerabilities related to data security and system interoperability.

Addressing these limitations requires targeted interventions. Multi-source data standardization frameworks must harmonize disparate datasets while preserving regional specificity. Crowdsourcing platforms with rigorous quality control offer scalable pathways for expanding dataset diversity, while generative adversarial networks enable synthetic data augmentation for rare events. Lightweight architectures optimized through network pruning and knowledge distillation can reduce computational requirements for edge deployment. Meta-learning frameworks provide crucial adaptability across diverse agricultural contexts with minimal additional training data.

Success requires sustained collaboration among researchers, developers, policymakers, and farming communities to ensure solutions address genuine operational needs. Validation studies in authentic agricultural settings must account for environmental variability and resource constraints. As global food security challenges intensify alongside mounting environmental pressures, the practical deployment of AI technologies becomes imperative for achieving sustainable, resilient agricultural systems capable of supporting growing populations while preserving ecological integrity.