Introducing digital twins to agriculture

.


Introduction
Digital twins (DT) are being increasingly adopted by several disciplines, including the manufacturing (Kritzinger et al., 2018), automotive (Caputo et al., 2019) and energy (Sivalingam et al., 2018) sectors, for addressing multidisciplinary problems. DT are digital replicas of actual physical systems (living or not), interweaving solutions of complex systems analysis, decision support and technology integration. DT have gained prominence, partially due to the uptake of Internet of Things technologies, that allow for the monitoring of physical twins at high spatial resolutions, almost in real-time, through both miniature devices and remote sensing, that produce ever-increasing data streams. DT have been useful for converging the physical and virtual spaces , guaranteeing information continuity through the system lifecycle (Haag and Anderl, 2018), system development and validation through simulation (Boschert and Rosen, 2016), and preventing undesirable system states (Grieves and Vickers, 2017).
The DT concept was coined by M. Grieves in a white paper (Grieves, 2014), as a unification of virtual and physical assets in product lifecycle management. Since then, several disciplines have adopted DT, each providing their own definition as there is no generally accepted definition of DT. A working definition for this study considers DT as "a dynamic virtual representation of a physical object or system, usually across multiple stages of its lifecycle, that uses real-world data, simulation, or machine learning models combined with data analysis to enable understanding, learning, and reasoning. DT can be used to answer what-if questions and should be able to present insights in an intuitive way" (Clark et al., 2019).
The benefits of DT applications include reduced production times and costs, hiding the complexity of integrating heterogeneous technologies, creating safer working environments and establishing more environmentally sustainable operations. DT are utilized by several leading companies and organizations, including Siemens (Negri et al., 2017), General Electric, NASA, US Airforce (Mukherjee and DebRoy, 2019), Oracle, ANSYS, SAP, and Altair . Furthermore, the recent availability of commercial software tools to develop DT, like Predix 1 and Simcenter 3d 2 (Negri et al., 2017), is an evidence in itself of increased interest in DT applications.
Information and communication technologies can be leveraged to design and implement the next generation of data, models, and decision support tools for agricultural production systems . Today, technologies like artificial intelligence (Patricio and Rieder, 2018), big data (Wolfert et al., 2017) and Internet of Things (Elijah et al., 2018) find their way in practice, and start to converge. Benefits of this convergence have been demonstrated in DT applications in other disciplines. However, DT are hardly utilized in agricultural applications, and their added value has not yet been discussed extensively. As a result, questions emerge regarding the benefits of DT for agriculture, the characteristics that differentiate them from current practices, and their design and implementation.
The purpose of this work is to investigate the potential added-value of DT in agriculture. To achieve this goal, we will first research the extent to which DT have already been explicitly adopted in agricultural applications, and investigate their reported benefits. Second, we examine the similarities between DT applications in agriculture and other disciplines, to identify opportunities of potential added-value for agricultural DT. Our research questions are formulated as: • RQ1: To what extent have digital twins been applied in agriculture? • RQ2: What is a potential application-based roadmap for the adoption of digital twins in agriculture?
To address these questions, we employed a mixed-method approach, as exploratory research suggested DT have not been extensively used in agriculture. Thus, a literature review alone would not suffice due to the limited number of reported cases in the literature. Our approach consists of a literature review of existing DT in agriculture, and a survey of case studies in other domains, the latter added to compare with the DT adoption level in agriculture and investigate potential future applications. We searched for DT use cases in agriculture, as well as in other disciplines to see how they employ DT. Note that we did not focus on identifying specific DT applications, rather we aimed at generalizing them into abstract, representative use cases. For the use cases identified, we explored the dimensions of maturity, service types and benefits offered. Our methodology is described in detail in Section 2 and the results are presented in Section 3. In Section 4, we discuss our findings concerning the current state of DT in agriculture, the added-value of DT, and we potential areas for future research. Section 5 concludes this work.

Methodology
To answer 'RQ1: To what extent have digital twins been applied in agriculture?', we identified existing DT use cases in agriculture and extracted attributes which helped us assess how advanced these applications were. To identify use cases, we performed a literature review for DT in agriculture and extracted indicators of maturity, service type and benefits. Maturity captures the development stage of the application (e. g., idea, lab, production). Limited use cases of production level DT is an indicator of less widespread use of DT. On the other hand, increased research and deployed applications indicate that DT are still finding their way into agriculture. To describe the purpose of DT on an operational level, we extracted the service type attribute. These services indicate the broader set of operations that DT perform. From the service type, we can understand the complexity of the DT operations, with higher complexity meaning potentially higher added value for the application domain. Also, the service category of DT in agriculture was compared with the service categories found in other disciplines to examine how advanced agricultural DT operations are. Next, to show what is the added-value of DT based on existing applications, we extracted the benefits attribute. Less materialized benefits from the applications indicate limitations for adoption. Below we describe step-bystep how the literature review was performed.
First, we searched in scientific databases and subsequently extended our search to grey literature. We included grey literature because a preliterature search showed that the peer-reviewed corpus covering DT in agriculture is rather limited. By including grey literature, we also cover work in progress and commercial applications that have not been published in scientific literature.
Second, we checked the corpus for relevance. In scientific publications, we read the abstracts to verify that the topic was about agriculture with references to DT. For the grey literature, we scanned the entire articles to see whether they connect DT to agriculture.
Third, we read all the selected articles and extracted use cases of DT applications. References to similar DT applications between multiple articles were considered only once to avoid redundance. We identified each use case with a number, summarized it in a single paragraph describing its functionality, and extracted the reported benefits.
Fourth, we identified the services offered by each DT use case. We used the service classification initially proposed in , and subsequently aggregated in (Cimino et al., 2019). The categories we used for classifying the use cases are presented in Table 1. We categorized the use cases in this way to identify the complexity of operations that DT performed as operation complexity is an indicator of the advancement of DT in agriculture. Also, this categorization helped us compare the types of operation offered by DT in agriculture and other disciplines, and determine any potential gaps to further assess their adoption in agriculture.
Fifth, we categorized the use cases based on their technology readiness level (TRL) to examine whether they are in experimental stage, or if they have been used in production. We partitioned the European Union's TRL scale (European Commission, 2014) into three generic levels shown in Table 2, and used them to tag the use cases. The first level represents DT which were still in a conceptual phase, the second consists of DT that had a working prototype even without the complete planned functionality, and the third level covers mature DT deployments in production.
Sixth, we identified the physical twin, i.e. the physical system that was twinned in each use case. We classified them in the following categories: living plants or trees, animals, agricultural products, i.e. harvested fruits; agricultural fields, farms, landscapes, farm buildings, as barns, greenhouses or other agricultural buildings, agricultural machinery, including equipment and tractor appliances, and food supply chains and logistics.
Finally, we summarized in a table all the identified use cases, their respective descriptions and the extracted three dimensions -service categories, TRL, and physical twin -to depict the breadth of the application of DT in agriculture. Fig. 1 summarizes the methodology for answering RQ1.
To answer the second research question, 'RQ2: What is a potential application-based roadmap for the adoption of digital twins in agriculture?', we searched in literature for use cases aiming to identify the ways in which DT have been successfully applied in other disciplines. Again we aimed at identifying use cases, and extracted indicators of benefits, maturity, discipline, and service type to understand the operations in which DT are most effective and what problems they can solve. First, we searched for peer-reviewed review papers of general DT applications.
1 Predix is a software platform that facilitates data collection, processing and analytics for industrial applications. The product description can be found in https://www.predix.io/.
2 Simcenter 3d is a software environment that integrates 3d modeling, simulation and data management. It includes modules to capture the dynamics of fluids, composites, acoustics and others. The product description can be found in https://www.plm.automation.siemens.com/global/en/products/ simcenter/simcenter-3d.html.
Second, we scanned the full texts for occurrences of the string 'digital twin', to check if the reviews were related to DT. If DT were briefly mentioned and not the main point of the review paper, we considered the reference irrelevant. Third, the remaining articles were examined in alphabetical order based on their title to extract use cases. Repeated mentions of similar use cases were not considered. Fourth, we extracted a short summary of the use cases, the reported benefits that they offered, the discipline, maturity and service categories using the same framework as for research question 1, and the publication and application years. Fifth, we proposed areas of potential application in agriculture, and identified potential benefits based on the use cases in other disciplines. Fig. 2 illustrates the methodology for answering RQ2.

Literature review of digital twins in agriculture
For the literature review of DT in agriculture, we first searched in Web of Science (Web of Science) using the query "digital twin*" AND (agri* OR crop* OR farm* OR aqua* OR animal*). This query returned results which contain DT and derivatives of agri, crop, farm, aqua, or animal, to capture cases of DT in subfields of agriculture. The query returned seven results. After the relevance scan the results were reduced to four (Smith, 2018;Tagliavini et al., 2019;Tsolakis et al., 2019). We then extended the search to Google Scholar (Google Scholar) using the query "digital twin" agriculture. The query returned 947 results. We examined them until five consecutive results were irrelevant (24 results examined), and checked for duplicate results from the previous search in Web of Science, thus reducing the number of results to nine (Tan et al., 1094;Kampker et al., 2019;Moghadam et al., 2020;Qi et al., 2019;Machl et al., 2019;Verdouw and Kruize, 2017;Gomes Alves et al., 2019;Delgado et al., 2019). Extending to the Google search engine (Google), we used the query "digital twin" agriculture which returned 143.000 results. We examined them until five consecutive results were Table 1 The digital twin service categories used to classify the use cases identified by the literature review. The column Typical components lists the components that are usually needed to implement the corresponding services.    irrelevant or referring to previously found applications (38 results examined). We then checked the extracted results for duplicates from our searches in Web of Science and Google Scholar, eventually reducing the results to nine (Monteiro et al., 2018;IBM Research, 2018;R&D WORLD, 2019;Collins, 2019;Barnard, 2019;Chiu et al., 2019;Mokal and Sharma, 2020;Ohnemus, 2020;Wageningen University & Research, 2020). In total our search yielded 22 sources for DT applications in agriculture. From this result-set, we identified 28 use cases. Following the methodology described in Section 2, we summarized each use case and extracted data about the expected benefits, TRL, physical twin, and service category. The results are reported in Table 3. Our search yielded 14 scientific articles. Nine of which were published in journals and five in conference proceedings. Additionally, we identified eight website articles from the grey literature search. Publication outlets, titles and year of publication are summarized in Table 4. We observe that the first published references to DT in agriculture date back to 2017, and most of our sources are from 2019 onward.
The use cases reside in different sub-fields of agriculture: In dairy farming we found DT for the detection of mastitis in cows. Related to apiculture, we found a DT of bee colonies aiming to control their welfare and honey production. In plant production, a DT of tomato crops in a greenhouse aimed to control the growing environment. In agricultural machinery, a DT of tractors was used to emulate their performance prior to purchasing. Other DT included orchards, pig farms and aquaponics production units. We noticed that DT of animals and fields, farms and landscapes are reported with less technical detail. In contrast, DT of agricultural machinery and food supply chains and logistics were often described with more details about their design and operation.
The reported benefits varied, depending on the physical twins. For twins of living systems, like plants and animals, the benefits included early disease identification, production optimization and identification of factors that could degrade their welfare. For agricultural products the benefits were cost savings and improved product quality. Support in crop management decisions allowing for faster action was reported for agricultural fields and farms. Twins of agricultural buildings reported benefits related to growing conditions management and production increase. Lastly, DT in agricultural supply chains and logistics reported benefits included cost savings and more environmentally friendly operations (Table 3).
Physical twins include both non-living subjects, like farm buildings such as farm bins or livestock barns, and living subjects, such as arable farms or individual animals. Most of the DT were found for physical twins of agricultural fields, farms, landscapes and buildings. Fewer were found for living plants and animals or agricultural products and the food supply chain. Fig. 3 illustrates the types of physical twins identified together with the maturity of the use cases as TRL level.
Regarding the TRL, most DT identified in this study were on the conceptual level. In Fig. 3 we observe that DT of agricultural fields, farms and landscapes are mostly on the concept level. We also notice that DT in the food supply chain and agricultural machinery have surpassed the concept level stage. Besides, none of the identified agricultural products DT have reached the deployment stage.
Regarding the service categories, the identified use cases of agricultural DT perform energy consumption analysis, real-time monitoring, system failure analysis and prediction, optimization/update, and technology integration. The majority of the DT perform monitoring and optimization operations (Fig. 4). We do not observe any pattern of the TRL levels across the service categories.

Digital twins in other disciplines
For the examination of DT in other disciplines we searched in Web of Science using the query TS="digital twin*" AND ALL=review. This query returned results that had DT mentioned in the title, abstract, or keywords and had the word review mentioned somewhere in the text. Instead of filtering the type of results to reviews only, we chose to search for the word review because some review papers are not always not explicitly tagged as such in Web of Science, or sometimes they miss the word review from their title. The query returned 37 results. After scanning the articles for relevance, the results were reduced to 23 (Kaewunruen et al., 2018;Patterson and Whelan, 2017;Fraga-Lamas and Fernández-Caramés, 2019;Zheng et al., 2019;Tilbury, 2019;Dewitt et al., 2018;Bolton et al., 2018;Dong et al., 2019;Cohen et al., 2019;Tomiyama et al., 2019;Tao et al., 2019;Lu et al., 2020;Raman and Hassanaly, 2019;Yi Wang et al., 2019;Pizzolato et al., 2019;Mabkhot et al., 2018;Cimino et al., 2019;Gupta and Basu, 2019;Ghobakhloo, 2018;Kim and Kim, 2017;Longo et al., 2019). Following the methodology of Section 2, we identified 68 use cases, and extracted a short summary, benefits, maturity level, discipline, service categories, year of publication and year of application for each case, reported in Table 5.
We observed that DT in other disciplines performed energy consumption analysis, real-time monitoring, system failure analysis and prediction, optimization/update, technology integration and virtual maintenance. Most of them performed monitoring and system failure analysis operations (Fig. 5). The TRL varied by the year. The earliest documented DT application (2011) was that of an aircraft, which was used in production. From 2011 to 2016, new use cases were scarce. After 2016, many DT applications emerged at the concept and prototype levels, as well as some deployed ones. Applications in the concept stage were more frequent than the ones at the prototype and deployed stages. The reported benefits included cost reductions, energy savings, reduced equipment downtime, quantification of system reliability and safer working environments for personnel.

Threats to validity
The results of the literature review for DT in agriculture showed that there are only a few DT use cases reported in scientific literature. Moreover, 13 (Table 3, uc. 11-16, 20-26) out of 28 DT use cases were used in the commercial sector and 7 (Table 3, uc. 20-26) out of those 13 were documented only in non-scientific literature. This may imply that the industry is ahead of academia in the development of DT.
Also, we limited our search to Google Scholar and Google to applications up to 5 consecutive irrelevant or duplicate results. More DT could potentially be found if we examined more results or additional sources.
Another factor that the literature review of this work does not consider is the existence of agricultural applications which are not defined as DT in literature. There are potentially applications that are used as DT but for unknown reasons they were not tagged as such and as a result they were not included in our results.
Besides, in our literature review we included conceptual level DT applications, which means that they are not established applications, but work in progress.

Current state of DT in agriculture
In this section, we investigate the state of DT in agriculture by comparing it to the state of DT in other disciplines. The results of the literature review in agriculture show that the available literature is limited. Considering the year of publication, DT have been discussed in other disciplines since 2011 (Table 5, uc. 54), while in agriculture the first references occurred in 2017. Our interpretation for this delay to investigate DT, is that agricultural researchers are more risk-averse than in other disciplines. A reason may be that in agricultural applications, firms are often small and medium farms. Such farms can bear less risk than bigger companies in other industries, who can afford to experiment and innovate, and thus pioneered DT. Also, DT in other domains are mostly concerned with non-living physical twins, as complex industrial Table 3 The use cases of agricultural DT. Use cases are referred as "uc" and their corresponding numbers in the text. The numbering of the use cases continues for the use cases in other disciplines.        and manufacturing applications. In agriculture, even the non-living physical twins, as those of agricultural buildings, still indirectly interact with plants or animals. The direct or indirect interactions with living systems introduce more challenges for DT in agriculture. We identified only two overlapping use cases between our searches for agricultural DT, and DT in other disciplines. Use cases (uc. 83, 84) correspond to (uc. 6, 3). We expected a larger number of overlapping use cases, especially as our search of DT in other disciplines did not exclude agriculture-related use cases. This may be an indication that agricultural DT have not been adopted extensively, as they are not selected as representative use cases in DT reviews.
The benefits of the applications mentioned in the agricultural use cases include cost reductions (uc. 6), more detailed information (uc. 3), catastrophe prevention (uc. 15), positive economic impacts (uc. 7), aid in decision making (uc. 4) and more efficient management operations (uc. 12). Looking at the benefits of DT in other disciplines, we observe that they have a broader range. They also include safer human-machine interaction (uc. 58), building cost and energy efficiency estimation (uc. 35), and insights into complex multidisciplinary systems (uc. 94). DT in agriculture have not yet reached the point to demonstrate similar benefits.
Regarding the TRL, we were initially surprised to see that all levels are approximately equally represented. This large number of field-deployed or production-level DT could indicate a high adoption level in agriculture. However, upon closer inspection, we noticed that 6 out of 8 deployed DT were extracted from a single article (Verdouw and Kruize, 2017), reporting on the results of the FIWARE Accelerator Programme (FIWARE Foundation, 2020), whose purpose was to create applications using the FIWARE platform. 3 Apart from the DT deployed by the FIWARE program, we observe that there has been little progress in advancing DT beyond the concept and prototype levels to the production level, where they can be used in real-world conditions. A reason for this may be that in other disciplines there are greater financial incentives, and larger research capacity to try out new technologies, or report their findings at earlier stages. Also, some applications on the conceptual level were described abstractly without any detailed technical design reporting, i.e. uc. 1, 3, 9. To our knowledge, Wageningen University and Research has recently introduced an investment theme on Digital Twins, developing twins of tomato crops and arable and dairy farms, but they are still on a conceptual stage (uc. 27, 28).
Another interesting finding from Fig. 3 was that the supply chain and logistics and agricultural machinery twins were the only ones that did not have any use cases on the conceptual level. While this could be circumstantial, it may also indicate that agricultural DT targeting these sectors are more mature than others. As DT of agricultural supply chains and logistics build upon relatively similar deployments in other supply chains and manufacturing, this could explain their relatively higher level of maturity. However, we did not check thoroughly to what extent DT of agricultural supply chains are concerned with perishables. This argument also pinpoints a significant challenge of DT in agriculture: Most agricultural operations have to do with living subjects, like animals and plants or perishable products, and creating DT for such systems is harder than for non-living human-made systems. Another reason why most DT are on concept and prototype level might be that agriculture is a slow adopter of technology, partly due to the growing complexity of information technology (Delgado et al., 2019). To successfully develop DT, the community must become familiar with a variety of related technologies including Internet of Things, machine learning and big data. Most of these technologies are still considered new fields of experimentation in agriculture (Basso and Antle, 2020), and once the community gains confidence around them and adopt best practices for their application, we are likely to see more DT emerging in prototype and deployed levels.
Considering the service categories, most of the agricultural DT offer monitoring and optimization services. Other service categories reported Table 4 The source, article type and publication year of the use cases for the literature review in agriculture.  were related to energy consumption analysis, and a few of the DT acted as technology integration tools. In other disciplines, we also came across the virtual maintenance category which was absent in agricultural DT. A reason for this gap could be that implementing an advanced technology like DT with more complex operations can be expensive (Delgado et al., 2019), at least in the early experimental phase of its adoption. Applications of DT performing virtual maintenance could be useful for determining the optimal repair/maintenance strategy of agricultural machinery before laying hands on it, similar to repairing subsea equipment in (uc. 75).
Regarding the variety of the applications, from Fig. 3 we observe that a variety of applications like livestock farming (uc. 6), cropping (uc. 4) and apiculture (uc. 16) are encompassed. Yet, we believe that there is more room for DT to grow in each subfield. In our view, one of the reasons for not having a wider range of applications is the added complexity of the systems that DT pursuit to digitize, especially as this domain is lagging in digitization. Many agricultural systems are living systems, comprising of complex processes, which are harder to model than DT of products or human-made systems. This is in agreement with our findings related to DT in healthcare, another domain that also has to do with living physical twins: Only two use cases were identified related to healthcare (uc. 22, 46). Challenges related to living physical twins include capturing underlying processes that are still not wellunderstood, and accurately monitoring certain processes, for example   The use cases of DT in all disciplines. Use cases are referred as "uc" and their corresponding numbers in the text. The numbering of the use cases continues from the use cases in agriculture.        nitrogen leaching in crop systems. In agricultural systems, it is also common that certain processes are not digitized because there are no financial incentives for doing so. Another aspect affecting the adoption of DT in agriculture is that the community has to build trust in the interplay of the DT components for its correctness. This trust is essential to create DT that can accurately represent the inner workings of a system, propose maintenance strategies and alternative ways of management. Yet, building this trust in agriculture is difficult, because many decisions affect living systems where, unlike in other disciplines, consequences can be hard to reverse.
The lack of data culture also slows the adoption of DT in agriculture. DT require large amounts of data to operate, and the expected benefits are not eminent in small-scale deployments. In this respect, the lack of a data culture  and compartmentalization of agricultural systems understanding inhibits DT development and decreases potential for adoption. As a last note, integrating DT components and updating them in real-time can be daunting. For a community that is highly interdisciplinary and less information technology-oriented (Brown et al., 2019), this is a major turnoff.

The added-value of digital twins
This review identified few applications of DT in agriculture, with several of them being only superficially described in the corresponding articles. This suggests that DT benefits have not been clearly communicated to the agricultural community yet. Consequently, the community has not yet had the chance to investigate how they could utilize them and include them in their current practices. In this section, we pinpoint in the form of characteristics the benefits that DT can bring to agriculture. The characteristics can be seen in Fig. 6.
The vision behind DT is to offer personalized curation of complex systems. This means that DT can account for local system idiosyncrasies, that are often too complex to be accounted for in a generic model. DT adapt to local conditions in each individual physical twin, by fusing data and learning from them. DT are customized to mimic the individual characteristics of each system instance and deployment, and expose the system under different perspectives like system health, operation effectiveness, and profitability.
Streamlining of operations is another characteristic of DT. They offer an automated pipeline of operations like data acquisition from sensors, performing simulations, creating reports and controlling actuators. These operations are executed continuously, without requiring the attention, time and expertise of the users. DT bring together operations that previously were offered by a range of tools, hide their complexity, save time and remove context switching obstacles for the users. In this way, DT democratize technology and make it available to a wider range of stakeholders.
A key aspect of DT is information fusion, as they integrate and enrich information originating from several heterogeneous sources. DT observe physical twins from different perspectives by using multiple sources of data and assessing possible outcomes of actions. Information fusion combined with the continuous nature of operations depicts the complete picture of the past and current state of the system, and allows to estimate future states.
Uncertainty quantification is another characteristic of DT. DT can take into account the cumulative effect of the involved uncertainties since they observe systems from different angles. This information can then be customized and communicated to the stakeholders according to their expertise.
DT often embed permission level controls. The type of reports and controlling mechanisms can vary, based on the user of the application. This makes it possible to create different levels of transparency, depending on the sensitivity of the handled data and the importance of the operations taking place.
Finally, DT may demonstrate human-centered intelligence to control mechanisms for aspects that were neglected in the past, like human-machine interaction for safer working environments.

The future of digital twins in agriculture
The added value of DT has not yet materialized in agricultural applications. DT could be used pervasively, on different spatial and temporal scales and with varying levels of complexity, depending on their  components and the desired functionality. We expect that the future of DT will evolve from simpler cases, exhibiting fewer components, to more sophisticated ones. We propose a roadmap for the development of DT in agriculture, starting from simple DT applications, with fewer components and simpler functionality, gradually adding components and functionality, to demonstrate the full potential of DT.
On a fundamental level, a DT will include monitoring, user interface and analytic components. These components are the first step towards empowering a DT to monitor and analyze agricultural systems and offer a continuous stream of operations. An example DT with these components could be deployed to monitor the microclimate of a greenhouse and provide insights for its management. In this case, the DT would monitor environmental conditions, like solar radiation, humidity and CO 2 , analyze them according to user-defined thresholds and report its findings, similar to the use case (uc. 21).
A slightly enhanced DT could include actuator components to control fans and windows in a greenhouse. The monitoring and control operations would be performed continuously, notifying different stakeholders with information that is relevant to them. For instance, in the case of consecutive stormy days, the DT would notify the farmer that it closed the windows because the temperature dropped, and notify the supply chain stakeholders that the production will be delayed because the plants cannot grow fast enough with the current weather. Also, the DT will report which indicators surpassed certain thresholds, thus taking specific actions using its actuators, and consequently assuring the stakeholders of its correct operation. Similar twins could be deployed to food silos (uc. 12) to keep track of their stock and autonomously organize their proactive replenishment, notifying the supply chain stakeholders and farmers respectively, and to livestock farms to keep track of environmental indicators that are known to affect animal welfare (uc.

25).
Further enhancing DT with simulation components is necessary for them to support decision-making based on past and future predicted states of the physical twin. A dairy farm DT could use simulation to forecast the occurrence of mastitis due to intensive milking for each individual cow. Utilizing this DT, a farmer could evaluate multiple milking scenarios and choose the one that strains the cow the least (uc. 1). Data analysis and simulation would happen in local or guaranteed cloud infrastructure to ensure data privacy. More advanced, simulations could investigate factors that have already lead to the appearance of mastitis, and result in improved breeding decisions. On an agricultural farm, DT of fields could use simulation to approximate the behavior of equipment in local conditions (uc. 4). Utilizing such a DT, a farmer could test a harvester, before purchasing it, on her local field with different weather scenarios to measure fuel consumption and plant damage.
Incorporating a learning component brings agricultural DT to the next level. A learning component may allow DT to assist in management operations for systems where the underlying mechanisms are unclear. In the case of a livestock farm, a DT with learning capabilities would be able to find patterns in real-time and in historical environmental data that could facilitate the onset and spread of diseases like swine fever. This would help stakeholders to take proactive measures to prevent not only the spread but also the appearance of diseases (uc. 6). Additionally, the DT would identify the most important variables shaping these patterns, estimate related risks, and clearly communicate the involved uncertainties, by presenting probability metrics for example.
Towards Digital Earth (Goodchild et al., 2012), a large-scale DT of an agricultural landscape, may consist of multiple DT of individual farms, each with several learning components. Such a DT will be able to consider the inter-field dynamics regarding water flow, fertilizer dispersion and nutrient leaching. It would provide variable fertilizer rates, based on site-specific intelligence, for example what amount can be absorbed by each field without being dispersed to other fields, and how much each field should be irrigated considering groundwater levels, and the availability of irrigation infrastructure. This would happen by learning from historical data about how the amount of fertilizer and irrigation affected the crop yield and depleted the nutrients of each field in the past. Ultimately, the DT would constantly improve itself in defining the acceptable fertilizer amounts and irrigation through continuous learning, also learning from the past decisions of the individual farmer. Besides, capitalizing on this information would lead to the creation of better cropping patterns, using different constraints like weather, profitability and field nutrient replenishment rate.
Further improving agricultural twins with a human-machine interface component would allow the establishment of safer working environments. A DT of a harvester with a human-machine interface component could trace the position of the workers and their actions to ensure that the machine is distant enough to avoid injuries (uc. 33). Also, a DT of grain bins could detect human presence inside the bin with cameras, and stop the procedures that cause grain movement to prevent entrapment. This is crucial as a large number of injuries occur every year with agricultural equipment due to the lack of safety measures (Jadhav et al., 2016).
Overall, DT can be applied to several agricultural subfields like plant and animal breeding, aquaponics, vertical farming, cropping systems and livestock farming. Adopting DT can start with simple setups, that can be gradually enhanced with more components to make them more intelligent and autonomous.

Considerations regarding the application of DT in agriculture
The application of DT in agriculture also involves potential pitfalls. As mentioned in (Smith, 2018), controlling physical twins through their virtual counterparts may lead to a lack of attention to the real-world systems. In agriculture, such neglect could cause irreversible damage, as DT are applied to living physical twins, among other things.
There are also cases where DT are not yet feasible, due to the large amount of resources they require to be developed, and the high complexity of the physical twins (West and Blackburn, 2017). This could be the case of some agricultural system interactions that cannot be accurately quantified yet. There are also concerns about the technology skills required to create DT (Lohtander et al., 2018). DT development requires specialized knowledge from several technology domains, which can be a serious threat in an already multidisciplinary domain like agriculture.
Synchronization between the physical and virtual twins is another target that is difficult to achieve (Talkhestani et al., 2018). In agriculture, human-made systems like agricultural equipment could be easier to synchronize with the virtual system, unlike natural systems such as animals or land parcels.
Also, the integration of DT components can be difficult (Kurth et al., 2019). In agriculture, this could be the case for combining the simulation and monitoring components for crops, as they rely on different infrastructures, software and end-users.
Last but not least, the widespread success of DT in agricultural applications does not only depend on technology, skills, or data infrastructures and availability but the involved business aspect. As with any new technology that is to be introduced in a farm, DT need to demonstrate their added value and the return on investment.

Conclusion
Returning to our first research question, we found that there are already a few applications of DT in agriculture. However, they are in primary stages and are not designed thoroughly enough to offer the benefits that other disciplines enjoy. Exceptions included some deployed applications that were part of a European Union-funded program. We believe that there is still a long way to go before the agricultural community can fully seize the benefits of DT. Agricultural researchers and stakeholders should make an effort to stay up-to-date with technological advancements and seek to find links between agricultural problems, and problems that are solved with DT in other disciplines.
Regarding the second research question, we proposed a roadmap of applications, starting from DT with simpler functionality, incrementally adding components to gradually demonstrate the benefits that are already present in other disciplines. As for the twins themselves, we foresee that there will be some confusion in the coming years about what a DT is and when a technology can be considered a DT. Research has been done to classify technologies based on how close they are to becoming DT (Kritzinger et al., 2018), but it is still difficult to identify when a system can be called a DT. For the needs of most agricultural applications, we suggest that a DT should have at least the monitoring, interface and analytic components.
We identified two distinctive characteristics of DT in agriculture while reviewing the use cases and proposing our application roadmap. The first difference is that many agricultural DT involve directly or indirectly living systems and perishable products. While DT are ideal to provide insights into such complex systems and incorporate nondeterministic processes, their integration with the physical twin can be difficult. This is further amplified due to the idiosyncrasies of living physical twins. The second difference lies in the spatio-temporal dimension of their operation. DT in other disciplines range between the size of an airplane to that of a factory. Agricultural DT range from individual plants and animals to twins of land parcels, farms, or regions. As such, one may need to consider effects across these scales. On the temporal dimension, agricultural DT differ due to the slower response rates of their physical twins. Agricultural processes like the growing of plants tend to evolve relatively slow, so at least initially there is no need for high-frequency interactions between physical and digital twins. These two characteristics of agricultural DT need to be considered when developing DT inspired by DT in other disciplines.
As a final note, given the potential for the adoption and the benefits of applying DT in agriculture, we strongly believe that they have the prospect to bring a technological breakthrough in the near future.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.