Artificial intelligence, big data, and blockchain in food safety

2.1 Big data

The definition of big data is rapidly collected and complex data in huge quantities as defined by the World Health Organization (WHO) [10]. The storage of big data can reach terabytes (10¹² bytes), etabytes (10¹⁵ bytes), or even zettabytes (10²¹ bytes). As described in ref. [11], it has three important characteristics, namely, high volume, high velocity, and high variety information. Similarly, The European Commission (EC) [12] defined big data as large amounts of data generated from various types of sources. New tools and technologies such as powerful algorithms are needed to deal with these highly variable data.

In all the above definitions, volume represents the amount of data, velocity refers to the information generation and processing speed, and variety represents the variation in data formats (e.g., structured data that are organized by rows and columns, such as tables and databases, and unstructured data represents the information that cannot be represented in a structural manner, such as BBC News pages, and other social media postings [13]).

2.2 Artificial intelligence

AI, defined as “the simulation of human intelligence in machines that is programmed to think and act like humans”, is a hot research field in computer science. It contains a broad range of concepts such as machine learning (ML), deep learning, etc. AI has undergone several stages of development, namely, (i) knowledge-based rules, (ii) patterns using ML, (iii) automation using deep learning. In the first stage of AI, knowledge-based rules are first defined by human experts and then are used identify rules or make decisions. Specific patterns are designed by using techniques such as feature engineering and then ML are utilized to classify or recognize these defined patterns in the second stage. Then, deep learning techniques are utilized to identify and make decisions automatically in an end-to-end manner.

2.2.3 Deep learning

ML based AI has achieved great progress for the manner of learning from data. However, ML based methods are concentrated on the way of learning and ignore the processing of feature acquisition. Thus, deep learning (DL) based AI is investigated to extract high levels of information from large amounts of data. Different from traditional ML based methods, DL based algorithms are organized hierarchically with the increase of complexity. Multiple processing layers are stacked to learn data representations with multi-level of abstractions.

DL based algorithms have obtained unprecedented achievements on a wide range of applications, such as image classification, object detection [3], image segmentation [4, 5] in computer vision. There are mainly four models in DL based algorithms, namely, convolutional neural network (CNN), recurrent neural network (RNN), generative adversarial network (GAN), and deep reinforcement learning (DRL). CNN is a kind of ANNs that are constructed by multiple layers, including convolutional layers, pooling operation, and fully connected (FC) layers, and are used to learn spatial hierarchical features through back-propagation automatically and adaptively [16]. RNN is a neural network with feedback (closed loop) connections and can be used for learning sequential or time varying patterns [17]. GAN is a new framework with a generative model G and a discriminative model D, where G learns the data distribution, and D estimates the probability that a sample came from the training data rather than G [18]. DRL is a deep neural network version of reinforcement learning in the traditional ML based AI, which is implemented by using deep layers and reward mechanism [19] (see Figure 3).

Figure 3:

Machine learning (ML) and deep learning.

2.3 Blockchain

As defined in ref. [20], a blockchain is a digital, decentralized, and distributed ledger, which records and adds the transactions in a chronological order in order to create permanent and tamper-proof records. It is combined by different methods, techniques, and tools to solve a specific problem. Due to its ability to increasing transparency, increasing the immutability of transactions, and enhance trust among stakeholders in the food supply chain, blockchain has obtained wide attention. A blockchain is a growing record list linked through cryptography, namely, blocks. A block consists of a cryptographic hash of the previous block, a time stamp, and transaction data (a Merkle tree). Modification of the data is not resistant in a blockchain because the data cannot be changed retroactively without alteration of all following blocks once recorded [21].

In conclusion, blockchain has four main characteristics as follows. (i) Decentralization: Validation through the central trusted agency is needed in the traditional centralized transaction systems, which inevitably results in the additional cost for the system. Differently, in a blockchain network, any two peers (P2P) can realize the transactions without the authentication by the center server. The server costs and performance bottlenecks can be reduced and mitigated by this way with the use of blockchain. (ii) Persistency: Tamper is nearly impossible because each transaction in the network needs to be recorded and confirmed by every block distributed in the whole network. Moreover, each broad-casted block is validated by other nodes, and transactions is checked. Thus, it is easily to detect falsification. (iii) Anonymity: Generated addresses can be used by a user during the interaction with the network to avoid identity exposure. Private information of users are no longer kept in any nodes, which ensures certain privacy on the transactions. However, the perfect privacy cannot be guaranteed due to intrinsic constraint. (iv) Auditability: The previous records can be easily verified and traced by accessing the time stamp of the interactions in any nodes in the distributed networks. For example, in the bitcoin blockchain, every transaction can be accessed to previous records iteratively. This mechanism ensures the traceability and transparency of the data in the blockchain.

The blockchain is a chain structure constructed by the blocks using a hash algorithm, as shown in Figure 4. The first block, also called the block header, is the genesis block. The creation block mainly consists of Hash value, Mine difficulty, Nonce, Time stamp, and Merkle tree root data [22]. The middle block includes all transactions within a certain period including the hash value of the previous block, Nonce, and Times tamp. The blockchain generation process can be briefly summarized as follows: by adding a time stamp, all parties can keep the account and verify it at the same time, and confirm it every 10 min to form the entire network record [23].

Figure 4:

The generation process of the blockchain.

3.1 Data collection

Various types of data sources with useful information can be utilized for food safety, such as online databases, sensor data, genomics data, and social media data. However, the diversity of the data also brings challenges that it is difficult to identify relevant information within a data source and the relation with other data sources.

3.2 Data storage and transferring

Large-scale of data collected from Internet, genomics data, social media, and sensor data are managed in a structure manner using databases such as structured query language (SQL), and NoSQL (see Table 3). General data management methods, such as MySQL, Oracle, and PostgreSQL are utilized to realize data storage. However, these systems are not adequate to handle big data for the flexibility and reliability of these systems are limit due to the SQL. Thus, NoSQL systems are developed for next-generation databases that are open-source and horizontally scalable. For example, MongoDB, Cassandra, and HBase are representative databases using NoSQL.

Apart from data storage, it is more difficult for dealing with big data from different sources in a NoSQL cluster. Therefore, data transferring software is needed to transfer data, such as Aspera, Talend, and Elasticsearch.

3.3 Data analysis

Data analysis are operated after the data storage and moving the data into the processing unit. As shown in Table 3, a list of the most used analysis tools for ML and deep learning are presented. Python [44] is the basic programming language for data science and AI. Scikit-learn [45] is a ML module written in Python. It integrates a broad scope of advanced ML algorithms for supervised and unsupervised problems. Similarly, PyTorch [46] is a deep learning framework which supports lots of deep learning algorithms and computes in a parallel manner by using graphics processing units (GPUs), as well as other frameworks such as TensorFlow [47], Keras [48], MXNet [49].

For rule-based AI applications, the authors in ref. [50] developed an ES to investigate seed potato production. The ES constructed 127 rules diagnose 11 pathogenic diseases and six nonpathogenic diseases by two knowledge engineers and five domain experts in potato disease diagnosis. The proposed ES has shown great potential in managing the disease component of seed potato production. Similarly, Sarma et al. [51] designed an ES based on a logic programming approach for the diagnosis of common diseases in the rice plant. The system is composed of a knowledge base, inference engine, and user-interface. To identify the halal safety rating for food additives, Zakaria et al. [52] used an ES to offer the consumers a safety rating key according to their past food consumption record experience. Similarly, to monitor and predict the product quality in the production process, an intelligent ES was developed. Recommendations for the products and the defects for the final product can be identified by this system [53].

Compared to human ESs based on logic programming, ML methods can make the process of protecting food safety become more “intelligent”. When rule-based AI systems cannot work in complex scenarios, ML algorithms are deployed to make decisions by learning from data [54]. ML models, such as Naive Bayes, SVM, MLP, and ANN are widely applied for the applications of food safety. To automatically grade mangoes based on the features of fruit, Pise and Upadhye [55] proposed a ML system based on the features of size, shape, color, and surface pixel of mangoes using Naive Byes and SVM. Experimental results have shown that the proposed system can achieve better performance than rule-based AI systems. Similarly, to analyze the quality of wine, Shaw et al. [56] focused on the comparative study over different classification algorithms for wine quality analysis. In their study, SVM, random forest, and multilayer perceptron (MLP) were compared. ANN has also been utilized for dried vegetables quality detection in ref. [57].

The features used for food quality assessment are often defined by human experts, which may not proper for the classification results. Instead of using features of food independently, deep learning can learn features from the food images automatically, which can be more suitable for the process of protecting food safety. The authors in ref. [58] used the deep denoising auto-encoder to illustrate the influence of food contamination on gastrointestinal infections and provided a valuable tool for morbidity forecast. To detect the quality of food for safety, convolutional neural networks (CNNs), stacked auto-encoders (SAEs) are applications for the qualitative detection of vegetables, fruits, and meat. Liu et al. [59] developed a novel classification algorithm by using deep feature representation learned by the stacked sparse auto-encoder (SSAE). The proposed SSAE was then combined with CNN for the defect recognition of pickling cucumbers by learning from hyper-spectral images. Similarly, the authors in ref. [60] recognized the plum varieties by applying image analysis and deep CNNs. Computer vision and deep learning based methods were also applied in the quality assessment of meat products in refs. [61, 62]. The food supply chain is a complicated system consisting of various economic processes from primary farmers to customers (including production, transportation, restoration, retail). The multiple processes result in unreliable information from the supply chain, which can give rise to food fraud and food safety dilemmas. Thus, it is hard for official organizations or governments to achieve reliable information about food safety. Mao et al. [63] proposed a credit evaluation system using blockchain for the food supply chain with long term short memory network (LSTM). For more examples about deep learning for food, we recommend the readers for reviews such as refs. [64, 65]. Chen et al. [66] adopted deep learning for food quality assessment, where AI-based methods, image processing (IP) system, and sensor technology can be utilized for quality assessment in the FI. The IP systems and AI can be used for various purposes, such as classifying products based on size and shape, detecting product defects, the presence of microbes, and grading food quality (see Table 4).

3.4 Data visualization

Table 2 presents several visualization tools which are available to analyze and give summaries of a large volume of data. R [67] and Circos [68] are two most commonly used tools. R, an open-source programming language widely adopted in data science, is often applied to visualize and analyze data. Circos enables the visualization of data using a circular layout and can illustrate the relationship between objects or locations, which has become the main choice of genome chromosomes visualization. The above two tools require programming skills, several commercial visualization softwares such as IBM Many Eyes and Tableau are good choices for researchers without any programming experiences.

3.5 Data security

The complex food supply chain results in the unreliable information, which will influence data security. Similar to food safety, data security is also essential to the whole process of data processing. Thus, blockchain is utilized to guarantee data security during the collection, storage, and processing stages. As shown in Table 5, a blockchain is constructed based on its unique infrastructure, which contains six layers, namely, data layer, network layer, consensus layer, incentive layer, contract layer, and application layer [69, 70]. Each layer performs a core function, and all the layers realize a decentralized trust mechanism by cooperating with each other [71].

From the bottom to the top layer, the data layer briefly illustrates the physical form of blockchain technology. The main function of the network layer is to realize the information communication chain among nodes in the network [23]. The network layer mainly contains point-to-point transmission technology (also known as peer-to-peer [P2P] system), propagation mechanisms, and verification mechanisms. It has the functions of consensus algorithms, encrypted signatures, data storage, etc. The consensus layer enables highly decentralized nodes in a decentralized system to effectively reach a consensus on the validity of block data [72, 73]. Proof of Work (PoW), Proof of Stake (PoS), and Delegated Proof of Stake (DPoS) are three commonly used consensus mechanisms. The main purpose of the incentive layer is to provide certain incentives to encourage every block to participate in the security verification and to attract participants to contribute to the computing power [74]. Various script codes, algorithmic mechanisms, and smart contracts are the main elements of the contract layer, which establish regulated and auditable contract specifications. The last layer is the application layer, which mainly focuses on programmable finance, currency, and society, such as bitcoin.

Blockchain, a tool that builds trust among stakeholders with divergent benefits, has been applied in many different sectors since it was first used in the year 2008 [75, 76]. Blockchain technology can provide an unprecedented applications for innovation in different domains due to its decentralized, transparent, and tamper-proof features. Originally, it was intended to record financial transactions between individuals, and it has been widely applied with technological advances and increased interest from international companies [77]. At present, blockchain technology has been extended to many other fields, such as mutual insurance [78], citizenship identification [79, 80], as well as food safety [81]. In the research field of food safety, blockchain technology is mainly applied to guarantee the data generated in the food production process will not be falsified to ensure food safety, strengthen food traceability and increase consumers’ trust in food safety [82, 83].

As shown in Table 6, for plant food, each link transaction (such as purchase and use of seeds, pesticides, herbicides, and fertilizers) related to the plant culture can be verified by nodes in the network by using a “consensus mechanism” [84, 85]. Meanwhile, for animal food, the food for animal feeding will influence the quality of the animal food, such as pesticide or herbicide residues, heavy metal residues, and various other contaminants [86, 87]. Animals are inevitably interfered with by diseases and the use of veterinary drugs or antibiotics, resulting in veterinary drugs or antibiotic residues [88]. Moreover, the temperature is also a very important factor affecting the safety of animal food. The microbial contamination of meat products mainly occurs after the slaughter of livestock. When the conditions are abnormal, microorganisms are born, which affect meat quality and cause food safety problems [89]. Based on HACCP, blockchain, and Internet of Things, a distributed information system has been proposed for monitoring food production [90, 91]. In this system, blockchain technology can be used to control all aspects of food production, such as sorting, cleaning, processing, storage, transportation, and retail. In developing countries, producers usually seek to gain greater profits by mixing low-cost ingredients that are harmful to human health; deliberate deception is quite common at the expense of the food quality for sale [92].

4.1 Challenges

As shown in Figure 5, we illustrate the challenges of the applications for food safety of different technologies. Big data is mainly responsible for data acquisition and pre-processing. A large amount of data are generated during the food production process, which brings challenges for the process of data pre-processing. Moreover, these data are emerged from different sources as mentioned above, which will make the system hard to identify the source of the data and update data promptly. Most importantly, the data acquisition or collection process in charge of the data quality, which is important for data analysis. AI is utilized for data process or analysis for assessment of food safety. However, a large amount of data with labels are needed for the process of AI models, which means that human experts are required to mark the data with labels. Due to the data capacity and AI models, computing resources are required for training AI-based models, especially deep learning based models. Moreover, generalization problems often occur in AI models, where the trained models are not appropriate for different scenarios. For blockchain, which is used for data security, is the recent advanced technology. It may meet basic infrastructure problems, technical problems, and data authenticity problems.

Figure 5:

Challenges of big data, artificial intelligence (AI), and blockchain for food safety.

4.2 Futures

1. Basic infrastructures: At present, with the development of information and computer science, people are familiar with new technologies in their daily lives. However, there are still unknown information for people without any background knowledge. Thus, it will take time for people to accept these technologies, such as blockchain. Moreover, for these new technologies, infrastructures which can meet all the requirements of these technologies are still lacking. Therefore, it will take a long time to build the basic infrastructure system.

2. Technical challenges: As the food safety system is a complex system which contains both of hardware and software infrastructures, therefore it is hard to implement such system. Moreover, applications of big data, AI, and blockchain in the food safety system increase the difficulty of its deployment. With the development and application of 5G technology, the speed of data transmission will be greatly improved, and the shortcomings of these technologies in terms of speed could be overcome. However, it still needs hard effort for the applications of these techniques. Another important challenge is the current high development cost of the food safety system. Food safety systems using big data, AI, and blockchain may achieve significant cost savings by circumventing mediations, but in some situations, they may not have a competing advantage over current solutions in mature markets.

3. Compatibility and standardization: The compatibility and standardization of these systems across industries matter a lot in the real applications of systems. These architectures need to be constructed to support interpretability protection among industrial solutions for collaborative trust and information. Thus, compatibility and standardization of the different systems is another challenge.