Smart

.


Background
Agricultural businesses face a vast number of simultaneous challenges.Shrinking marginals, complicated pan-European regulations and external, as well as internal, demands to mitigate their environmental footprint are all examples of requirements to be met.As a response, several different techniques are proposed to meet the needs of farmers.Even though farming has been developing technologically for centuries, the 21st century offers a wide range of technological possibilities that could deeply affect the future of farming.One of them is Artificial Intelligence (AI).
AI is to a large extent responsible for the smartness in smart farming.The term 'smart farming' constitutes a wide scope and demanding expectations.In this paper, smart farming is defined as the system of data driven tools for decision support in one or several parts of a farm's production, not restricted to nor limited by the agricultural sector they belong to.Smart farming could enable increased yield volumes, mitigate the workload for farmers, contribute to climate change adaptation and future-proof farming for the coming centuries.With this in mind, smart farming is expected to affect several areas within the agricultural sector.To mention a few, some trained AI models are implemented to predict the optimal time for planting and harvesting crops, prevent nutrient deficiencies and the spread of diseases, and guarantee food safety [6] .
Contrary to most earlier research, this study investigates technical aspects as well as non-technical aspects of smart farming.Several earlier studies have scrutinized technical aspects such as optimal remote sensing picture resolution and important cybersecurity aspects to sensor systems.Here, those aspects are considered but other essential, practical aspects such as data ownership and data sharing are also analyzed.Furthermore, non-technical aspects to smart farming, for instance trust and profitability, are discussed.By this interdisciplinary approach, new insights into the possible application of AI in agriculture are provided.Additionally, the wide scope allows for a comparison between three different agricultural sectors: arable farming, milk production and beef production.The study took place in Sweden, but a large part of the findings may be relevant in an international context as well.

Table 1
Respondents of the interview study distributed over categories.On the horizontal axis the respondents are distributed based on their orientation towards meat, dairy production, arable farming or 'Interdisciplinary' for those organisations without any such specified focus.On the vertical axis the respondents are distributed over the categories of which occupation they have.

Literature review
To begin with, a comprehensive literature review explores the agricultural sector, its agricultural technology initiatives, and former research about smart farming.This was done by a structured literature review as described by Berrang-Ford et al. [2] .The literature review was conducted in the first half of 2021 using the Web of Science database using the following search string: (("machine learning" OR "deep learning" OR "artificial intelligence" OR smart OR AI) AND ("precision farming" OR "precision agriculture" OR "precision livestock farming" OR farm * OR agricult * ) AND (Sweden OR Scandinavia OR Europe)).Next, we selected those articles that met all inclusion criteria: publication date between 2015 and 2021, directly or indirectly linked to the agricultural sector and focused on individual farms or groups of farms rather than macro-perspectives on countries or regions.Out of the 87 results in the search, 32 articles met the inclusion criteria and were fully reviewed.

Interview study
Thereafter, qualitative data was collected through a semi-structured interview study examining how different agricultural stakeholders regard smart farming technology.All interviews took place in the first half of 2021.In total, 21 respondents were interviewed from different parts of the agricultural sector in Sweden.Table 1 shows an overview of the respondents.The respondents were grouped by their occupation, where ten are individual farmers (F1-F10), seven are people working at commercial enterprises and cooperatives in the agricultural sector (C1-C7), two are researchers at Swedish research institutes (R1-R2) and the remaining two work at a governmental agency (A1-A2).Out of the seven respondents from commercial enterprises and cooperatives, four come from organisations that are deemed to have some economic interest in the agricultural sector, although they also have a cooperative function.Two other respondents in the same categorization are strictly cooperatives, one with ties to the public sector.The final is a certification organ.The interviewed researchers are hired by a Swedish university with agricultural focus, and the respondents from the governmental agency are also tied to different functions within a Swedish authority focused on agriculture.The interview questions are included in the Supplementary Materials.The lists of questions were created prior to the interviews and have not been altered.Different lists of questions were used for farmers (Suppl.S1) and for organizations, companies, governmental authorities and researchers (Suppl.S2).For farmers, a different set of questions was used for those farmers familiar with the concept of smart farming (Suppl.S1 A), vs. those farmers not familiar with the concept of smart farming (Suppl.S1 B).For organizations, companies, governmental authorities and researchers, only one list of questions was used, since all were selected based on their experience with smart farming.All interviews were held through digital meeting platforms and all but two were recorded with the permission of each respondent.The codes mentioned in Table 1 are used throughout this article to refer to the respective interview respondents.

Results from literature review
The Web of Science database returned 87 articles matching our search string.From these articles, 32 studies met all inclusion criteria and were fully reviewed.This section presents the main results from the literature review.We divided our findings into two main topics which are represented in the sections below: (1) techniques for data gathering within smart farming and (2) smart farming technologies applied within agricultural sectors.

Techniques for data gathering within smart farming
Essentially, there are two types of data gathering techniques that are used in smart farming: remote sensing and Internet-of-Things (IoT).The techniques may be implemented in any type of farming, depending on the activity for which the data is fit to function as decision support.

Remote sensing
Remote sensing technology enables detection and monitoring of physical characteristics of the earth's surface.Remote sensing data is collected from a distance, commonly from satellites and drones.The three most common properties of remote sensing data are spatial, spectral, and temporal resolutions [ 13 , 19 ].Spatial resolution is the pixel size of an image, a property that affects the ability to detect objects through imagery.Differently, spectral resolution refers to the spectral sampling intervals size and number which affect the ability of the sensors to detect objects in electromagnetic regions.The temporal resolution regards the frequency of acquired data [13] .
The availability and economics of using remote sensing data collection is addressed by Khanal et al. [13] , which present remote sensing technology alternatives both open-accessed (e.g., NASA's Landsat, the European Space Agency's Sentinel satellite series) and for some cost (e.g., RapidEye, GeoEye-1) [13] .However, the resolution of the data varies, where the trend is that medium-resolution data ( ≥ 10 m pixel size) is free whereas the prices for high resolution ( ≤ 5 m) and very high resolution ( ≤ 1 m) data increase in proportion to their increasing quality [13] .
Regarding data resolution, Meier et al. [19] opine that site-specific smart farming depends on high resolution, as detection of anomalies are impossible or insipid with too large pixel sizes.Of course, depending on what kind of analysis the data aims to contribute to, the need for resolution varies.For example, predicting the crop yield within a field can accomplish a high accuracy despite a coarse resolution [13] while detection of plant diseases through hyperspectral imaging requires a detailed resolution [20] .

Internet-of-Things (IoT)
Internet-of-Things (IoT) is a collective concept for objects with incorporated electronics and connections that enable remote control and information sharing.In agriculture, IoT is mainly used for collecting data through different types of sensors.By further data analysis, valuable information can be derived as decision support, e.g. for farmers [1] .Kamienski et al. [10] define four main challenges for IoT development in smart farming.First, the IoT system must have a high level of adaptability.Since the needs of farmers often significantly vary, the IoT system must be customizable to local circumstances but still not increase the required work for the farmer.Secondly, the IoT deployment must be efficient.As Kamienski et al. [10] write, "there is no 'one size fits all' in IoT systems ".Thus, each system needs to be configured, the Internet connection and farm infrastructure must be reliable, and the farmer must deploy enough human and economic resources into this process.Furthermore, the scalability is affected by the previous factors but also depends on whether the system, and the models learned, are supposed to work for just one farm or entire agricultural consortiums.Lastly, the complexity of the IoT system can be interpreted as a trade-off between making the middleware broker complex and the software application simple, or the reverse [10] .
Another aspect to IoT in smart farming is security.Since the data often is valuable for the farmer and is regarded as a business secret, Kleinschmidt et al. [14] describe the need for end-to-end encrypted communication from the sensor to the application.In practice, this means that the IoT sensor network must have a synced security strategy to the cloud database and the potential fog computing network [14] .By ensuring security, the probability that the farmer trusts the IoT system increases.Still, trust in IoT systems does not just depend on security but also on the precision of the sensors.Without ensuring that there are no systematic measurement errors in the sensors, few farmers would trust the learned model or the real-time data [21] .

Smart farming technologies applied within agricultural sectors
Agriculture is a heterogenous industry with sectors in need of a diverse set of technologies.Since the gathered data differs between the sectors, certain data types are more common in some sectors than others.Also, the activities performed in the different sectors call for different use cases when applying AI to agricultural businesses.

Smart farming in animal husbandry
The potential of smart farming in animal husbandry, such as dairy-, beef-and fur production, is largely constituted by increasing productivity and profitability by streamlining and automating tasks and information [4] .Much research consists of ways to monitor and look after the animals automatically or semi-automated [3][4][5] .These articles suggest that devices, both wearable and non-wearable, may be incorporated in the animal stable and that these devices can gather data that can give indicators on the health of the animals.The data that may be gathered through these devices vary, but the wearable devices can measure heat, hormone levels, rut etc.The non-wearable devices more typically are 3D cameras for body condition scoring and infrared imaging, sensors that monitor environment and weather as well as automatic weighing scales and gates [ 4 , 5 ].

Smart arable farming
In arable farming, an important feature of smart farming is to be able to calculate the vegetation index of fields or areas to be able to monitor when it is time for harvest and other activities.This may be done by both remote sensing and IoT solutions.Viljanen et al. [22] train a machine learning model aimed to optimize the "balance between the highest possible yield quantity and an adequately high digestibility for feeding ".By using an inexpensive drone system that can get multispectral data from an RGB camera and an infrared camera, traditional physical tools for predicting ley yield can be replaced by smart machine-learned models with higher accuracy [22] .Furthermore, the research of predicting yield and quality of silage can also be accomplished through satellite data, as presented by Griffiths et al. [8] .The study shows that it is possible to detect mowing events of grasslands, and therefore characterize the landuse intensity by looking at satellite imagery [8] .
In terms of yield prediction, Feng et al. [6] stress the importance of incorporating biophysical characteristics of the crop in machine learning algorithms.This means that to learn a model with high precision, it is important to simulate the growing process of the crop to ensure that the model learns the crop characteristics in different stages of the growing process.Furthermore, Matos-Moreira et al. [17] uses manual soil samples to further improve their model.By including manual sampling and analysis with a variety of existing data sources one may learn a model to predict the quality of a crop or the concentration of some matter at a given place and time.
Another application of precision farming is to detect sickness or pests among crops.Torai et al. [20] study how diseases can be detected in crops by classifying, or labeling, areas in pictures as "healthy ", "infected ", "diseased " or "aged ".Thereafter, methods such as hyperspectral imaging, Bayesian networks, and an analysis through probabilistic latent semantics are applied to detect the diseases [20] .This study is a good example of a remote sensing technology applied to agriculture which needs a very high resolution of data, preferably on a scale of centimeters.
One dilemma when applying artificial intelligence to arable challenges is how to use the different types of available data.Kerkow et al. [11] use fuzzy mathematical modeling to solve this problem.This approach allows for mixing machine learned climate models with wind data and expert knowledge of the landscapes to build precise models [11] .

Implementation of smart farming
The literature review also brings up some interesting aspects regarding the implementation of AI in agriculture.Medvedev and Molodyakov [18] highlight both theoretical and practical knowledge of smart farming as requirements for successful implementation.Unfortunately, seldom farmers have either the economic resources or the time to attend longer educations within the subject.To meet the lack of technical education within smart farming, Medvedev and Molodyakov [18] propose smaller model-based courses that should cover technical, economic and management aspects to smart farming.A crucial part of the education is that the courses are on-demand, so that busy farmers can access it whenever it suits them [18] .
Both business cases and clear driving forces are named as critical components to spreading the use of smart farming technologies in society ( [ 7 , 15 , 16 ]).Barriers that hinder the drive towards smart farming are categorized as economic, institutional behavioral, and organizational [15] as well as market [7] .Furthermore, they identify social and moral drivers to play a key role in terms of creating a societal demand for smart farming.Without the support from society at large, innovations will not be adopted by key actors, they conclude [15] .
Other research aims to map the barriers to implementing and diffusing smart farming technologies.Kernecker et al., [12] describe that farmers approach smart farming technologies differently given how much smart farming technologies the farmers have already adopted.The so-called adopters perceive the barriers to adopt smart farming technology as high investment costs, a difficulty in interpreting data, a lack of interoperability or precision in devices, that farmers cannot see the added value of the new technology or the relative advantage of the system, as well as a lack of neutral advice from advisors and other actors.The non-adopters also perceive high investment costs and unclear added value as barriers.Additionally, they regard too demanding complexity of use, that the technology is not appropriate for their context or farm size, as well as a lack of access to proof of concept from a neutral point of view, as obstacles [12] .
Finally, the literature review highlights the importance of data presentation and visualization, both in arable farming and livestock farming.Beside identifying possible applications of the technology in agriculture, several research groups [ 13 , 21 ] argue that methods within machine learning and AI require decision support tools that visualize the data in comprehensive ways.

Results from interview study
This section highlights the main findings from the interview study.The results are divided into technical and non-technical aspects of implementing AI in farming.Finally, we present the results summarized by respondent group (farmers, companies/corporations, research institutes and governmental authorities).

Technical aspects of implementing AI
Applying AI to the agricultural sector is not a homogenous challenge since the sector varies substantially.One pattern, stated by a farmer respondent, is that farmers of different agricultural sectors almost always believe that the implementation of smart farming technologies has come further in other sectors than in their own (int.F4).The agricultural sector that most farmers highlight as currently the most technologically advanced is the milk production.Milk robots were introduced to the commercial market decades ago, and with the milk robots the fodder of an individual cow can be customized, increasing its health status and production capacity.Due to the milk robots, the dairy industry is regarded notably data driven (int.C4).
One important aspect to consider when evaluating the success of the milk robots is the short feedback loop.Since cows are both fed and milked daily, the machines can adjust quickly depending on the latest input (int.C4).Furthermore, Swedish dairy farmers have a long history of collecting data by being part of the so-called Kokontrollen, a cow data collection application owned by Växa Sverige.Even if Kokontrollen today is web-based, Swedish dairy farmers have been reporting to it for more than 100 years.Previously, all data was collected manually but today almost all data connected to milk production is automatically gathered by the milking robots (int.C6).
Contrasting to milk production, arable farming is diverse with different crops requiring distinct machines and technologies.Hence, a single successful machine is difficult to implement for the entire arable farming sector, making its technological development more complex.However, it is possible to create effective technology for specific crops.As a rule of thumb, crops with high manual work, such as vegetables, use lots of technology since they operate on small, more controlled areas (int.C3).In such environments, such as green houses, the feedback loop is faster and there are less uncontrollable factors, such as weather or wild hogs, which makes the application of new technology and AI easier (int.C3, F3).
Of the three agricultural sectors compared in this study, beef production is considered by the respondents as the least technologically developed.Nevertheless, one respondent at a major company believes that meat production will have a central role in the development of the Swedish primary food production (int.C4).The list of possible innovations includes making the value chain digital by automatically transferring information to the slaughterhouses regarding characteristics of the animals they will receive.By mandatory RFID tags for all cattle, the respondent argues there is an enormous potential, since the development of the animals could be followed in real time throughout the value chain.With such a system, the slaughterhouse could plan far in advance for incoming meat quality and volume.Simultaneously, a grocery store could send data to the farmers regarding the current popularity of different kinds of meat, enabling the farmers to adjust their production to the current consumer behavior (int.C4).Furthermore, if one could autonomously and automatically weigh the cattle, their growth curves can be predicted which would enable optimization of the timing for sending animals to slaughter.By this optimization, one could avoid having full-grown animals that both drain economical resources and emit environmentally damaging methane gas (int.C4).

Data collection within smart farming
Regarding data and the activity of collecting data, the responses from the interviews reflect different realities within the agricultural sector.
On the one hand, some respondents say that farmers generally are positive towards gathering data on their farm (int.C3, F6, F8).On the other hand, some responding farmers state that they collect almost no data on their farms, although they say that they understand that data could add value to them (int.F1, F3, F5).In-between is a spectrum of attitudes towards data gathering and implementation of technology in the farms.Some respondents from the larger companies and cooperatives suggest that the attitudes might be affected by the perceived inconvenience that data gathering causes.They all believe that more farmers would have a positive view on it if it was made easier for them to collect it.However, there is also a sense that the data is not used optimally, partly because it is saved in different databases that are not interconnected (int.C3, C4, C5).The responses from the respondents indicate that data is being gathered differently depending on the agricultural sector.For instance, many respondents in the dairy section state that there is a lot of data gathered, to a high degree on an individual level, on the farm animals (int.C6, F2).In contrast, arable farmers also collect data on almost all farms, but that data is not always as detailed.An arable farmer may collect remote sensing satellite data on its farm, but sometimes not with a resolution of square meters, but rather on a field or even farm level.The inputs, i.e. the resources added to the soil, are what would be interesting for the farmer to get decision support on, if one could see a beneficial correlation between input and output (int.F4).
One responding farmer with previous experience from the tech industry, believes that the problem with applying AI to arable farming is the lacking volume of interconnected data (int.F10).The whole data chain is not connected today, he states.In practice, the input data taken during, for example, arable seeding is not properly connected to the output of the harvest.Additionally, the insights from the harvest are not used as a decision basis for the next seeding.Thus, the data loop is not closed, which it would need to be for AI to be efficient (int.F10).This data gap combined with the large amount of uncertainty factors, such as unpredictable weather, is a technical hindrance to the learning of AI models.

Generalizability vs. precision
In the field of AI and machine learning, there is an important tradeoff between bias and variance.In the interviews, the respondents had different opinions on the matter.The concept was discussed with the respondents as 'generalizability' and 'precision' instead of their technical terms.Some respondents say that precision is extremely important since a technical solution that only predicts or detects something half of the time is useless (int.A1, C6, R2).At the same time, other respondents say that as long as the predictions are slightly better than human predictions or detections then the model can be as general as one wants.In fact, many respondents claim that there is a much larger market for standardized models than the ones that are too adapted after local needs (int.C4, F4, F5, F8).There is a tendency among arable farmers and corporations that they tolerate a higher degree of generalizability while livestock farmers need more precision (int.C1, C4, C6, F4, F5, F8).A respondent in the livestock farming sector claims that a farm would never really benefit from a technical solution that could only detect rut among the animals one out of three times.The respondent further underscores the need for precise models dealing with biology and living animals (int.C6).Of course, many respondents bring up that there is a need for balance between generalizability and precision, and that it would be optimal if there was some degree of customizability in that aspect so that each solution can fit each farm (int.C3, C5, F4, F5, F8).

Automatization and decision support
In terms of customizability, identifying how farmers want their technical tools and machines to be designed, and how they want them to assist on the farm, is key.Among the farmer respondents, only one respondent specifically mentions a preference for smart farming technologies to completely automate activities on the farm (int.F10).Instead, a large share of farmers is prone to either having technologies that recommend a course of action out of the collected data (int.F1, F5, F6), or that the technology only presents data so that the farmer may make decisions on their own (int.F2, F4, F9).Decision support and assessments based on compiled data already exist to a certain degree but may be used more often and efficiently should the data become more accessible and of better quality.A responding dairy farmer says that although the system provides some recommendations from the continuous and aggregated data, the decisions on the farm are still only based on the raw data (int.F2).

Data sharing
One key concern for the development of smart farming technologies is ownership of the data.Most smart farming systems are created as closed technological ecosystems, with limited possibilities of sharing data in between each other.This technological segregation hinders the systems to share data with each other and is thereby an obstacle to the interconnection between systems.Descending from the rivalry between the major transnational agricultural technology companies, including the quest to both pin the users to their specific technological ecosystems and avoid giving their rivals a chance to create competitive technology, this structure is difficult to change (int.C3, F10).With that said, two respondents (int.C4, F10) note a tendency for transnational agricultural technology companies to move away from technology that ensnares the user to their ecosystem, to more open data flow.Such open data flow is believed to create more value for the businesses and their users.Consequently, a higher degree of data is expected to be on open standards (int.C4, F10).
Even if the companies providing the technology make some progress towards open data sharing, a couple of projects are created to facilitate the data sharing compatibility.GigaCow, a research project by the agricultural university SLU on data for dairy farms, aims to enable data sharing by automatically exporting the data from different milk robots over time.Such initiatives are welcome to most farmers.However, this is a third-party work-around solution and not as straight-forward as if all machines would automatically be open for data sharing (int.R2).

Cybersecurity
Some respondents lift the potential threat towards online IT systems as a risk when implementing new smart farming technology (int.A1, C4, C5, F1, F2, F8).The risk of being hacked poses a threat both to farmers and to society at large.Focusing on society at large, a respondent from a governmental agency describes cybersecurity as a particularly important aspect of digitalization in agriculture (int.A2).This respondent believes that such a data platform probably would be classified with an extremely high security and secrecy label and be managed by the Swedish Security Service SÄPO.Therefore, this could be regarded as a clear barrier for the development process of a common data platform.Nevertheless, the respondent adds that in case of potential cyber-threats it would be better to have the data stored on a common platform than with individual farmers, since people would be managing and looking after the platform to a much higher degree than farmers currently are securing their data (int.A2).Even though these issues are mostly raised by the larger organizations and authorities, the threat is also acknowledged by some farmers.They believe that connected data platforms with weak security make the farm quite vulnerable to threats (int.F1, F2, F8).However, one farmer commented that "it is not worse than having all money in a bank account, and that I trust today." (int.F1).Other respondents, both governmental agencies and farmers, recognize the IT systems as possibly vulnerable but are not necessarily worried.Instead, they reject the belief that lacking cybersecurity would pose a greater threat to agriculture than to any other sector in society (int.A2, F5, F10).

Non-technical aspects of implementing AI
AI and smart farming technologies play an important role on the strategic agenda of most interviewed agricultural organizations and co-operatives (int.A1, C1, C2, C4, C5).For one major Swedish agricultural company, AI is considered one of the most important tools to achieve a more sustainable farming while increasing the yield (int. C1).Another respondent from a Swedish agricultural enterprise states that their take on digitalization of the agricultural sector will be to invest in software integration of different tools to facilitate smart farming technologies for farmers, so that they can monitor their farm and make correct decisions (int.C2).Digitalization and optimizing the data usage are prioritized over investing in new hardware, such as server halls.By prioritizing software, they believe they benefit the farmers more directly (int.C2, C4).

Political interest in smart farming
When it comes to digitalization of such a fundamental societal system such as the agricultural sector, many strategic decisions are of nationwide interest.Some of the interviewed respondents from larger organizations and authorities believe that there is a wide interest that the agricultural sector becomes smarter (int.A1, A2, C7).However, farmers are themselves accountable for making this technological transition.Two respondents argue that there is a lack of initiatives from the state or from the large organizations to drive the propagation of digitalization forward in a structured manner (int.A1, C7).One respondent, working at a governmental authority, addresses the topic of nationwide interest in digitalizing the agricultural sector (int.A2), stating that AI in agriculture is a natural step moving forward.The respondent says that there are a lot of internal discussions in governmental agencies regarding if and how they should take a more active leadership role in the digitalization of Swedish agriculture.The governmental official thinks that Sweden is behind with its digital development compared to other countries with weaker economic conditions and budgets for agriculture (int.A2).A natural first step, according to this respondent, is to create a common national data platform for all agricultural data to be compiled on.Still, this respondent sees no clear political ambition driving this change, while this could speed up the digital transition tremendously.Although there is no wish to 'force' farmers into using agricultural technology and digitalizing their businesses, it is a likely progress if there is a nationwide and political interest in going in that direction (int.A2).

Economic incentives
As in any other industry, the agricultural sector is driven by the quest for increased profit.Money is a motivator, not only for larger agricultural enterprises but also for farmers (int.F3, F5).Therefore, the general low profitability in agriculture is a major problem for farmers.Optimization plays an important role for the often unprofitable Swedish agricultural farms to be competitive on the world market (int.A2, F3).Even though there are lots of subsidies connected to food in the European agricultural system, no farmer respondents recognize any subsidies for investments in new technologies at a farm-level.Instead, the technological transition that is supposed to lead to more sustainable food production or larger output is financed by the individual farmer (int.R2).
Different farmers have distinct economic incentives to implement smart farming technologies in their work.Generally, there is one group of farmers that have less reason to care about implementing new technologies since they will have structures in place to reach their revenue in any case.This group often owns their own property and farmland.On the other hand, there are farmers that lease their farmland and therefore constantly must become more and more effective (int.C3, F4, F5).It is not only a matter of farm ownership though, also the size of the farm affects the probability that smart farming technologies will increase profitability.With a small farm, farmer respondents believe it is difficult to profit from smart farming techniques (int.F1, F5).A farmer with a small farm describes that he cannot afford buying new equipment, such as a new tractor, himself.Upgrading the machine park is necessary for smart farming technologies to gather enough useful data (int.F1).This can be linked to the major macro trend of consolidation of farms.Basically, this means that smaller farms cannot afford to compete with the larger ones that can use their competitive advantages of being larger.There is simply not enough profit in managing most small farms, a problem which forces many farmers to merge with neighboring farms (int.C7, F4).

Software as a service
Another trend that impacts the agricultural sector is how technologies are sold and distributed.Today, most technology is bought as a hardware which is often a huge expense for the farmer.However, slowly things are changing.There is a transition happening towards services being bought as 'Software as a Service' (SaaS) solutions.This allows for business models in which the sold hardware is much cheaper than today or even provided at no cost, while the farmer pays a fee to subscribe for using the set of hardware and software.One respondent from an agricultural cooperative foresees that this change will have major implications and wonders whether, in ten years from now, tractors will be sold solely as a rental service instead of as a product.To enable this, an enormous amount of data will be needed (int.C4).

Dependency and trust in technology
One communicated and discussed concern about implementation of smart farming technologies is the dependency it might create towards technology.Dependency on technology refers to a system that relies on automated or semi-automated activities based on often incomprehensible software, a constant power supply or Internet-access.The system itself is not problematic to any of the respondents.However, there are some concerns regarding the cases when this type of system fails.One respondent, from an organization, states that the usefulness of the system would be compromised if the communication infrastructure would somehow break (int.C6).The concern is expressed in different ways and with different urgency.Livestock farmers express their concern about this since their activities revolve around living beings, whose comfort and health rely on the technological systems continuing to operate (int.C5, F2).Also, when it comes to dependency on technology, another aspect that several respondents mention is that some practical knowledge among farmers and advisors might be forgotten (int.A1, F1).One responding farmer believes that if he applies too much technology to his farm he would risk losing some of the local, tacit knowledge of the farm.Particularly, some local variations of the farmland he finds difficult to represent correctly with data.Since there are a vast number of connected parameters affecting how a crop at a specific place will grow, he fears that a program could miss some critical aspects (int.F1).This may be linked to a certain expressed mistrust towards technology, that it needs to be double checked to make sure it is doing the right thing while working autonomously (int.F1, F4).

Working conditions
In general, there is a positive attitude towards smart farming and what it could mean, to the agricultural sector as a whole and to farmers specifically.Incorporating smart farming technologies could mean that time and costs for activities, such as irrigating and fertilizing, are reduced.Therefore, farmers can better manage their time when using well-functioning new technology (int.C3, C5, C6, F7, F8, R2).One positive side effect of this is an improved work environment for the employees (int.F8).With that in mind, researcher respondent R2 states that farmers are generally bad at valuing their time spent compared to the economic return.Agricultural farms are extremely pressed to get a rewarding return on their investments which leads to, at times of the year with high workload, farmers not getting much sleep at all (int.R2).This is confirmed by a farmer who says that since he works so much, some hours are nearly unpaid (int.F7).Implementing AI in agriculture could potentially mitigate these intense periods of large workloads somewhat, which would give social values back to the farmers.
Another dimension of investments and implementation of new technology in agriculture, is that investments in smart farming are not always viewed as necessary by farmers but rather something neat and trendy.Thus, such investments are described to be paid by the "amusement account " (int.C3, F5).This is confirmed by a farmer that states that most of the technological investments made on his farm are motivated by his interest and fascination with technology (int.F8).Respondent C4 says that a lot of farmers gladly spend money on new and exciting tools and machines, for instance new tractors.From this, it seems like many farmers think that the charm of running an agricultural business is to be able to tailor and adapt the farm according to one's liking.While some respondents like doing things very manually (int.F9) others like to develop their way of working consistently with new types of technology (int.F10).

Market demands and academia
Similar to the need to involve users in the design process, one hurdle for initiatives and research projects to reach the market is that their concepts are not based on agricultural market demands (int.C6, R1, R2).Three respondents argue that this might be a result of researchers and developers not being in contact with end users and not doing proper market research to identify for which demands there are needs to develop solutions.Instead, they believe that agricultural technology is often developed for the sake of technology development itself or, contrastingly, as a biproduct of agricultural research (int.C6, R2, A1).One researcher describes that, for several years, Swedish academia actively focused on other parts of agriculture than technology (int.R2).However, during the last decade, Swedish universities have started to understand the importance of new technologies in farming and allocated

Summary of results by respondent group
To summarize the results of this interview study, the themes and topics are divided into what appears to be the demands or opportunities for AI in agriculture, as well as the barriers or hurdles that hinder the use of it.Furthermore, based on the contrastive responses and views of different groups of respondents, the demands and barriers are differentiated by the respondent groups that all have distinct roles in the agricultural sector.Table 2 shows an overview of the most important points from the interviews, divided over the different respondent groups.
To begin with, the responses from farmer respondents show that there are many opportunities linked to the usage of AI and smart farming technologies in agriculture.Most importantly, according to them, new smart farming technologies have the potential of increasing their profitability, either by contributing to higher revenues or freeing time spent on some tedious tasks.On the other hand, the large initial costs to set up the technologies are identified as a barrier.However, if economical means allow for investing in such solutions, farmers believe that the investments will pay off in terms of profitability and competitiveness.Other factors that act as demands for smart farming technologies are their potential to be more sustainable and that they make farming more fun.Further barriers according to farmers are the complex solutions and lack of interoperability, as well as the poor prerequisites and opportunities of continuous education regarding technology in agriculture.Also, the fickle market makes smart farming risky to invest in for farmers.
From a commercial enterprise point of view, there are many opportunities connected to smart farming, but also some critical barriers to overcome.The respondents of this group see potential in increased cooperation between companies as well as with farmers, business cases in providing Software as a Service and additionally to streamline logistics connected to agriculture.Nevertheless, data sharing and cybersecurity are seen as large hurdles to the use of these technologies.
Respondents from research institutes also express a positive view on accelerated use of AI in agriculture.They believe such a development would result in more data collected by the farmers, which would decrease the time researchers themselves spend on gathering data.This would, according to the researcher respondents, lead to a faster and better research on agriculture.However, data sharing hinders, once again, the scientific development since high-paced research is hard to conduct without proper access to data from different sources.An additional identified barrier for smart farming is the mistrust from farmers that the scientifically developed solutions mirror a real agricultural demand and are not just developed for the sake of technology.Finally, the respondents from governmental agencies claim that there is a great interest and demand for propagating smart farming technologies for national competitiveness as well as other economic reasons.Still, they are not sure how to position themselves in this transition, which slows down the process of digitizing the agricultural sector.This respondent group also views cybersecurity and data sharing as critical barriers to overcome.

Discussion & conclusions
This paper provides a review of the main opportunities and hurdles for applying AI to agricultural businesses.By conducting a structured literature review and an interview study with 21 respondents from various parts of the agricultural industry, data has been gathered to get a holistic view on the use of smart technology in agriculture.The scope of the thesis is deliberately wide, focusing on three agricultural sectors: arable farming, milk production and beef production.Furthermore, the respondents are categorized by their role in the sector, ranging from governmental authorities, commercial enterprises, researchers as well as farmers.This broad view allows to acquire knowledge that ranges over several production sectors, as well as over several kinds of organizations with different views on the agricultural sector.
Most stakeholders interviewed in this study share a demand for applying AI to agriculture.Driving the farmers towards smart farming technologies are the needs for increased profitability, reduced workload and often a genuine curiosity for new technology.Surprisingly, all these aspects are not completely captured in the literature review.For example, there are studies about the impact smart farming can have on the relation between humans and animals on a farm [9] , but they did not show in the literature review search.On the contrary, some expected drives for smart farming were not expressed by the respondents, such as the advantageous impact that smart farming can have on the environment through less nutrient loss.Instead, profitability stands out as the most influential factor which makes a clear business case an essential requirement connected to the propagation of smart farming technologies.Since more and more agricultural products become available in the form of SaaS, allowing for sharing and renting equipment, the business case is changing for both farmers and machine producers, opening new possibilities.Nevertheless, for smart farming to really transform the agricultural sector, governmental agencies and commercial enterprises might need to take a more active role in the transition.Such aspirations are especially important to ensure that the governmental and societal demand for reduced emissions and increased sustainability is met in the technological shift.
For the transformation to be successful, it is essential that the structures, allowing farmers to apply the smart farming technologies, are modern.One key requirement is that farmers have continuous and easy ways to acquire up-to-date knowledge of how to apply smart farming.Therefore, there is a need to ensure technical, agricultural education which is easily accessible through for example flexible, on-demand courses.Additionally, the smart farming techniques need to be modifiable to match the varying transparency and adaptability demands that different farmers have.
Regarding how implementation and propagation of AI in agriculture might be hindered, this study identifies some factors that act as barriers.The most prominent one is how data is managed, which can be further specified to data sharing and ownership as well as cybersecurity.This is a complex question that as of now does not have a clear solution, neither technically nor legally.Here lays an important role for research institutes as well as authorities.However, there is a consensus among respondents that to transition the agricultural sector into a more data-driven and digital environment, the technical infrastructure must be secure.The solution must be able to guarantee that sensitive data is not available for intruders while at the same time guaranteeing access for the intended users.Furthermore, for the end users to be able to benefit from the digitalizing transition of the sector, the data models require a high degree of flexibility.This stems from the wide variety of machinery at farms as well as the varying level of technological interest and knowledge among the farmers.
Moreover, an important aspect that slows down the process of implementing smart farming technologies and AI in agriculture is the economical dimension expressed by the respondents.A large part of this are of course the high investment costs, but other economic aspects also play a part in this barrier.For example, the fickle market demands, the general low profitability in agriculture as well as the trend towards consolidation of farms all contribute to making investments full of risk.Other identified barriers that hinder the spread of AI in agriculture are some social factors, for example the concerns about technological over-dependency and insufficient end user trust towards technology.The lacking trust seems to stem from over-selling from developers of technology as well as a gap between the technology that is developed and the real market demands.
As for the technical solutions that could potentially solve the demand for AI and smart farming technologies, there are many possible ways.In this study, findings show that a lot of the data and sensors types already exist.The problem that remains to be solved is to connect the input data to the output data by developing the datasets, and thereby closing the data cycle.Today, the dairy sector generally holds a closed and elaborate data cycle whereas generally the meat and arable sector have less developed data gathering and therefore less precise decision support tools.This is highlighted in both the interviews and the literature review, as high-resolution data allows for more precise and detailed decision support.Although, after a thorough process of data gathering from input to output, one can build models and evaluate which one of them performs best with some specified evaluating metrics.Additionally, a general problem and difficulty in building machine learning models is that models tend to take too many variables at the same time.The results show the importance of 'starting small' when building the models, i.e. using few input variables to begin with and then tune the model adding only one more variable at a time.
It is also found that all possible use cases and technical solutions demand a high precision for classification model output as well as low prediction errors for regression models.Decision support in agriculture manages and affects core parts of the agricultural business, and therefore it is important that estimations and predictions are accurate.Interestingly, respondents from the arable sector express that they, as of now, accept higher levels of total error in the model.However, for future purposes and solutions with increased complexity, the total error must decrease which is likely to affect the bias-variance trade-off.A requirement for achieving precise supervised machine learning models, adapted to the local farm, will be easy pre-processing of the data.Thus, the data labeling process must either be simplified by developers or offered to the farmers as a service by consultants.
Technologically, the agricultural sector has developed for decades, but the shift towards smart farming techniques and data-driven agriculture might be one of the greatest transitions.Applied AI in agriculture has the potential to optimize and streamline agricultural activities in all sectors in agriculture.By data-driven decision support, and even tasks performed completely automatically, farmers hope to improve their output both in terms of quantity and quality, mitigate carbon emissions, decrease work time, and increase profits.For commercial enterprises and governmental agencies, the transition allows for updated supply chains and planning models, improving the agricultural industry on a macro-level.Still, several challenges remain unsolved, jeopardizing the speed of the transition.Here, there are important tasks for companies, authorities and research institutes.Nevertheless, with such strong incentives, the long-term trend towards increased usage of AI in agriculture is clear.The question is no longer if smart farming will continue to develop, but how the hurdles will be resolved, and which stakeholders will benefit from its radical transformative effects.

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.

Table 2
Overview of the main results from the interview study, divided by different respondent groups