Based on the domain analysis introduced above, the next section describes the conceptual framework developed for designing and implementing Digital Twins. Control is a basic concept in system dynamics. It ensures that the system’s objectives are achieved, even if disturbances occur. The basic idea of control is the introduction of a controller that measures system behaviour and corrects if measurements are not compliant with system objectives . Farm processes are ‘in control’ if the performance of its operations remain in a steady state. Therefore, the activities of these processes must include the cybernetic control functions necessary to demonstrate ‘cybernetic validity’. Basically, this implies that they must have a feedback loop in which a norm, sensor,discriminator, decision maker, and effector are present . Fig. 5 depicts these control functions in a basic control model.In farming systems these are the business processes of the involved actors that transform input material to final products at the end customer’s location. The sensor function measures the actual performance of the object system. The discriminator function compares the measured performance with the norms that specify the desired performance and signals deviations to the decision-making function. Based on a control model of the object system, the decision-making function selects the appropriate intervention to remove the signalled disturbances. Finally, the effector implements the chosen intervention to correct the object system’s performance. Digital Twins allow farmers to decouple the physical flows from information aspects of farm operations . Decoupling of control means that the measurements of the object system’s state are translated into a Digital Twin as visualized in Fig. 6.

The control cycle starts with measuring the object system’s state by the sensor function and with acquiring relevant external data.These data are then translated into a virtual representation of the controlled object system on the basis of a meta model. The Digital Twin includes all information relevant for the supported purposes of usage as specified in a meta model. Dependent on a specific purpose of usage, a virtual view may then filter irrelevant information and present it in such a way that it can be processed optimally by specific users on the basis of a meta model. The next control function is the decision-making function, dutch bucket for tomatoes which compares a virtual view on the object system with a specific control norm. Next, the decision-making function selects appropriate interventions for deviations based on its Decision Support Model, similarly as in conventional control systems. Lastly, the selected intervention is communicated with the effector function, either directly or via the Digital Twin using remote actuator systems. The previous sections have defined distinct control mechanisms of six Digital Twin categories. Fig. 8 incorporates these mechanisms into the control model. This model integrates all six Digital Twin defined categories, but not all elements will be relevant if less categories are applied. The integrated control model especially adds different types of the representation, i.e. imaginary, present, future and past digital objects. Imaginary digital objects represent reference objects that do not yet exist. Present digital objects represent the current state and behaviour of real-life, physical objects. Future digital objects project the expected state and behaviour of objects. Past digital objects represent the historical state and behaviour of real-life objects or objects that no longer exist in the real-world. Furthermore a reference object is added to allow for the representation of conceptual entities that come into existence in the design phase of the product life cycle. Once the conceptual entity is materialized, the real object can be connected to the virtual object.

This conceptual entity remains after the disposal of the real-life object at the end of the lifecycle. Reference object can also be relevant during the usage phase. An example is the usage of imaginary resources for planning purposes, which specify the type of resources and the properties necessary to do the job. Think of, for example, a virtual harvest machine having a certain capacity in specific weather and soil conditions. When the harvesting schedule becomes actual, a physical machine is chosen to do the job for the virtual one . Finally, the interaction between the decision-maker function and the Digital Twins is elaborated. In prescriptive Digital Twins, intervention proposals based on decision support models are transformed into future Digital Twins. As such, the expected object changes of virtual interventions are simulated. The decision maker uses this simulated interventions to decide on the final intervention. Autonomous Digital Twins also translate this intervention decision into planned object changes and subsequently into actuator instructions. Autonomous twins remotely control the effector function that executes these instructions. So far, the concept of Digital Twins and its underlying complexity were defined. The next section will present a technical model that is designed to implement this concept. This section proposes a technical model for the implementation of Digital Twins. A technical architecture describes the components of a system, interactions among components, and the interaction of a system as a whole with its environment . It is usually not drawn in one diagram but separated in multiple so-called architecture views each of which describes an architecture according to specific stakeholders’ concerns . For the purpose of this paper, we focus on visualizing main functionalities that are needed to implement the control model as developed in the previous section. Several technical architectures for Digital Twins are introduced recently. Schleich et al. proposed an abstract reference architecture that addresses some basic modelling principles for ‘twinning’ between the physical and virtual world properties, such as model scalability, interoperability, expansibility, and fidelity.

Alam and Saddik developed a specific a Digital Twin architecture, that analytically describes key properties of cloud-based cyber-physical systems. Redelinghuys et al. designed an architecture for Digital Twins of manufacturing cells comprising six layers, including local data, gateways, cloud-based databases and a layer for emulations and simulations. These authors consider Digital Twins as a next step in IoT-based cyber-physical systems. As a consequence the proposed architectures are similar to reference architectures developed in the IoT domain, in which virtual representation of objects have an important role. Important IoT reference architectures include IoT-A, ITU-T and AIOTI . The Internet of Things—Architecture provides a very in-depth definition of IoT’s information technology aspects . The International Telecoms Unions has developed an IoT Reference Model which provides a high level capability view of an IoT infrastructure . The Alliance for IoT Innovation has defined a High Level IoT Architecture to achieve IoT semantic interoperability . In the present paper we adopted the IoT-A reference architecture because it most explicitly addresses virtual entities as a core element of the architecture. The remainder of this section will introduce the IOT-A reference architecture and how it supports the implementation of Digital Twins. The Architectural Reference Model for the Internet of Things is developed by the European project IoT-A . Besides establishing a common understanding of the IoT domain, IoT-A aimed to provide essential building blocks and design choices for developing interoperable IoT system architectures. The reference model includes five different sub models: an IoT domain model, IoT information model, IoT functional model, IoT communication model and an IoT trust, security and privacy model . The ontological foundation is formed by the IoT Domain Model, which defines main concepts of the Internet of Things like Devices, IoT Services and Virtual Entities , and how these concepts are related. Building upon these concepts, the IoT information model defines the structure of IoT related information in an IoT system on an abstract level. The Functional Model decomposes the main functionalities of IoT-based systems into groups in a layered view. The IoT Communication Model elaborates the technical communication for connecting the different elements of an IoT-based system, including a reference set of communication rules to build interoperable stacks. The sub models are elaborated in very detailed architectural views and accompanied by guidelines. It can be concluded that the IoT-A is a very in-depth and rigorous reference model.

It is beyond the scope of this paper to describe it into detail, but we focus on its IoT functional model . For more details and the further technical implementation we refer to Bauer et al. and Carrez et al. .Basically, a Digital Twin architecture is composed of a physical object in real space, a digital representation of this object in the virtual space and the connection between the virtual and real space for transferring data and information . As argued previously, IoT technologies enable this synchronization of the physical and virtual worlds. The implementation model of our conceptual framework, based on the IoT-A functional model, addresses eight layers . These layers range from a device layer, blueberry grow pot which is attached to physical objects, to an application layer, which includes interaction with Digital Twin Users . The Device layer provides the hardware components that are attached to and directly interact with physical objects such as tags for unique identification, sensors and actuators. Important identification technologies used in agriculture include barcodes and RFID tags . Furthermore, a multitude of different sensors is used to measure dynamic properties of physical things including temperature, crop size, humidity, light, moisture, CO2, ammonia and pH values. Object sensing is also supported by mobile devices such as barcode/RFID readers and smartphones, which enable farmers to perform additional actions such as visual quality inspections. Furthermore, this layer includes remote sensing by satellites, aerial vehicles, and ground based platforms. Small unmanned aerial systems are increasingly used to realize a high spatial and temporal resolution and a high flexibility in image acquisition. Finally, in the device layer actuators are used to remotely operate objects such as tractor implements, climate control, irrigation, coolers, and lights. The Communication layer manages the interactions between different components and enables the communication from the devices to the IoT services. It provides capabilities for networking, connectivity and data transport and enables end-to-end communication that crosses different networking environments. The IoT Service layer contains services and functionalities for discovery, look-up and name resolution of IoT Services. It can be used to get information retrieved from a sensor device or to deliver information to control actuator devices. The Digital Twin Management layer contains functions for interacting with the IoT System on basis of virtual entities. It can give access to all the information about the Digital Twin, from sensor devices, databases or applications. Furthermore, it contains all the functionality needed for managing associations with the physical objects and monitoring their validity.

The IoT Process Management layer provides an environment for the modelling and execution of IoT-aware processes. Deployment of process models to the execution environments is achieved by utilizing IoT Services that are orchestrated in the Service Organisation layer. This layer acts as a communication hub between several other layers by composing and orchestrating services of different levels of abstraction. The Security layer is responsible for the security and privacy of the systems and its users. It includes components like authorization, authentication and identity management. The Management layer is focussed on the configuration of the system. It also reports faults and determines the overall state of the system. Finally, the Application layer provides the intelligence for specific control tasks based on virtual objects. It includes capabilities for usage of Digital Twins across its lifecycle. The different categories of Digital Twins are enabled by diverse technologies, including simulation and optimization tools, statistical forecasting, simulation and machine learning. This layer also includes the user interface for interacting with Digital Twins. The types of user interfaces can vary from 2D graphical user interfaces, as commonly used in personal computers, smartphones and tablets, to advanced 3D interfaces for Virtual and Augmented Reality glasses. The remainder of this paper will illustrate the application of Digital Twins in agriculture by some cases of the IoF2020 project. The framework as presented in the previous sections is applied to five smart farming use cases of the IoF2020 project . It is beyond the scope of this paper to exhaustively deal with the applied models for all cases. Therefore, we provide in Table 4 an overview of the applied control models and in Table 5 the applied implementation models.