8. The human impact of data bias and the digital

119

8. The human impact of data bias and the digital agricultural revolutionKelly Bronson, Sarah Rotz and Adrian D’Alessandro

INTRODUCTION

We write this chapter in the midst of a global COVID-19 pandemic, which has left millions of lives upended for an indeterminate time as we await what the future will look like. In this context, we suggest that COVID-19 is both a crisis and a govern-ment leadership opportunity: to imagine an alternative future, specifically for the food system. Since the latter 20th century, governments all over the world have taken a back-seat position, letting businesses and the goal of profit maximization drive social policy decisions. This is true of the global food system, which conse-quently has moved toward fewer and larger farms that specialize and capitalize their operations (Hendrickson and James 2005; Holtslander 2015; Rotz and Fraser 2015). Mazucatto (2020: n.p.) suggests that in the midst of COVID we ought to ‘rethink what governments are for: rather than simply fixing market failures when they arise, they should move towards actively shaping and creating markets that deliver sustain-able and inclusive growth. They should also ensure that partnerships with business involving government funds are driven by public interest, not profit.’ This chapter directly responds to Mazucatto’s call and explores the extent to which current part-nerships and innovations between governments and industry that are aiming to drive a digital ‘revolution’ in agriculture can support inclusive techno-social change—that which contributes to the good of the many instead of the few.

Governments around the world are investing heavily in the realization of the future ‘smart’ farm, or ‘farm 4.0’ (Rotz et al. 2019a). On the smart farm, machinery, such as tractors, are fitted with sensors for collecting data into big datasets, which are then ‘mined’ by computer programs, or algorithms, that can produce farm-management advice (Janssen et al. 2017). For example, data collected by harvesting machinery are algorithmically manipulated to produce color-coded yield maps displaying real-time harvest information to the farmer. Proponents of smart farming technologies (such as Bayer/Monsanto who sell a yield-mapping tool) claim that this technology allows for the efficient use of resources, thus maximizing profitability and sustainability in agriculture (Adamchuck et al. 2010). On the promise that smart farming will contribute to global food security as well as climate mitigation and adaptation (Bucci et al. 2018), significant public money is being funneled toward digital farm technologies (see GOC 2017). In Canada, for example, the federal government has given millions of dollars of support to a consortium of industries called the Protein

120 Handbook on the human impact of agriculture

Industries Supercluster as a means to ‘promote wide-scale adoption of data analytics and artificial intelligence for improved nutrient and crop management’ (GOC 2019).

We are motivated by the COVID-19 pandemic to pause and reflect on what con-stitutes sustainability in digital agriculture as well as the potential human impact of these developments. Preliminary work (see Bronson 2019) suggests that smaller and diverse farms are not supported by current commercial smart farming tools, even though these farms are known to be important for food security and environmental sustainability (IPES 2016). As such, we ask: Which farmers and food system actors do scientific decisions surrounding digital agricultural technologies serve? We develop a combined critical social science and computer science method to investi-gate key issues in agricultural big data and algorithms. In particular, we draw on 22 unstructured qualitative interviews with designers of these technologies working in the private (14) and public sectors (8) in Canada and the U.S. which were conducted by the first author. We also draw on the computer science expertise of the third author to analyze (a) two proprietary corporate big data analytics tools (Farmers Edge N-application and soil texture classification); and (b) one public agricultural big data and machine learning initiative spearheaded by computer visualization researchers. Ultimately, we argue that there exists an unhealthy bias in agricultural big datasets and models with which they are used and we argue that governments ought to incentivize innovations which serve the goals of food system diversity and inclusion and thus maximize the human and environmental benefit from current techno-social shifts.

BACKGROUND

Many actors from the farm to the international level are investing in what appears to be an inevitable ‘revolution’ in agriculture founded upon the collection of large volumes of data. As early as 2002, Sidney Cox declared that advances in sensing technologies and a subsequent information technology revolution would allow for better management of environmental crises related to agriculture. Cox (2002: 94) elaborates:

In the 21st century we face a set of common environmental problems … To combat these problems we must look for international collaboration on an increasing scale, exploiting our ever-deepening understanding of the physical world, aided by an increasing array of tools for exploration of that world. Many of those tools can be considered under the subject heading, Information Technology (IT) since, by definition, IT is concerned with the acqui-sition, recording and communication of information.

Today, data collection sensors are embedded right into farm machinery called ‘precision’ farm equipment. This machinery is ‘precision’ because, when operated in conjunction with GPS, the data collected can be used to point to specific areas of a field in need of scarce or harmful inputs such as water or chemicals (Poppe et al. 2015; Sonka 2015). Data from farm machinery and also remote data collection (such

The human impact of data bias and the digital agricultural revolution 121

as satellites) are ‘fed’ into algorithms, which can ‘mine’ them for predictions about yield and input needs in the future. In some cases, these algorithms are self-learning (also described as machine learning). It is through a reduction in inputs that precision agriculture is thought to present sustainability gains. Indeed, because ‘precision agriculture substitutes environmental information and knowledge for physical inputs’ (Bongiovanni and Lowenberg-Deboer 2004: 359), it is widely believed to be a ‘paradigm shift’ or an altogether different approach to agriculture (see Rossel and Bouma 2016).

Despite the promise of the digital approach to help address grand challenges like agriculture’s contribution to climate change, it is possible that these tools raise challenges of their own. We know from previous innovation-led agricultural revolutions—notably the Green Revolution and biotechnology—that agricultural innovations bring benefits to some people and socio-cultural, economic and political costs to others (Bronson 2015; De Schutter 2015; Moseley 2017). Both hybrid seeds and genetically modified organisms (GMOs) are a now taken for granted as part of the global food system (Kloppenburg 2005), but their expense and legal infra-structure have exacerbated power inequities between small and large farmers, and between farmers and powerful agribusinesses (Laforge et al. 2018; Qualman 2011). These power dynamics have translated into social impacts, such as a ‘cost-price squeeze’ on farmers: oligopoly corporations have increased machinery and chemical input costs beyond increases to commodity prices (Galt 2013). This economic pres-sure is felt acutely by small farms, who receive little support to weather financial risk (Martinson and Campbell 1980).

Smart farming innovations beyond GPS guidance on tractors have not received widespread or consistent adoption (see Gardezi and Bronson 2019); hence, their implications are not well known. While there has been tremendous attention paid to the analysis of big data and algorithms in other sectors, such as medicine and retailing (boyd and Crawford 2012; O’Neil 2016), relatively few social scientists have turned a critical eye toward big data in the context of agriculture.1 We can make inferences about likely categories of human impact raised by big data in the context of agriculture by drawing on insights from other sectors, such as health and social media. Scholars have detailed the implications of the collection of personal data on privacy, civil liberties and on human rights (Kitchen 2014). Some scholars have focused on the implications of biased data collection, which has reproduced societal disadvantages especially when they are used to ‘feed’ or ‘teach’ algorithmic decision tools thus ‘driving discrimination upstream’ (Pasquale 2015: 35). Notable here is Eubanks’ (2018) Automating Inequality, which directs attention to the role of digital technologies in social institutions like medical insurance provision decisions and affordable housing policy. Through detailed ethnography, Eubanks shows that data-based systems categorize people for a range of non-technical reasons that map onto historic racial and socio-economic biases, which result in the rejection of applications or cessation of benefits (see also Noble 2018). Similarly, O’Neil’s (2016) book with a clever title—Weapons of Math Destruction—lays out impacts of personal data use and automated decision-making in socially consequential domains,


from decisions about the hiring and firing of public-school teachers, the awarding of insurance benefits, or in legal decisions about whether to grant parole. Like Eubanks, O’Neill shows how partial datasets and models used to manipulate them have dire implications for the already vulnerable (see also Pasquale 2015).2 So far, no compa-rably detailed analysis of bias in agricultural big data and algorithms exists (see also Bronson 2019).3

While we might borrow from the insights developed by analyses of big data and AI in other sectors, there are likely interesting and unique aspects specific to agri-culture. For example, it is worth considering the basic premises of machine learning. Broadly speaking, machine learning refers to algorithms that can improve on their completion of a given task by learning to perform that task drawing on big datasets. This is important in the smart farming context because agricultural data is sparse and difficult to collect and, historically, agricultural decision support has been guided by research and experimentation related to genotype, environment, and management (GEM). Even a superficial reflection on the GEM relationship makes it obvious why agriculture may pose a unique set of circumstances to the machine learning and big data science communities: there are hundreds, if not thousands, of agriculturally relevant crops, and potentially hundreds of cultivars and varieties of each crop. Moreover, there is an enormous range of environmental conditions and management practices within which these crops can be contextualized. The amount of high-quality data necessary for machine learning algorithms to learn and produce high quality (accurate, for example) decision support models for each and every specific crop is immense. Given this, any progress in these methods will come from prioritizing very specific crops under very specific management practices in very specific conditions. These are human decisions which deserve attention and oversight.

METHODS

By focusing on agricultural big data and algorithms that are being used to ‘drive’ farm decision-making, we extend research in food studies on agriculture and innova-tion, as well as research in critical data studies on data bias and harm. To do this, we develop a unique combined social science and computer science method premised upon a theoretical commitment to seeing data as harboring, in the words of Gitelman and Jackson (2013: 3), ‘the interpretive structures of their own imagining’ (see also Bowker 2006). Someone determines from the outset what counts as a possible var-iable that will be used to measure a particular phenomenon. Additionally, someone decides how to collect data on that variable, and they make a number of decisions about how to organize and analyze those data. In the case of an automated decision support system for farmers, humans have, for example, chosen the ‘inputs’ (the host of data points on variables like soil and seed); if the system ‘learns,’ then a human has decided which data will ‘train’ it. Someone has to produce a matrix of weights that will be used by the classifying machine to determine the classification for new or future input data. Each stage of this process is an epistemic event that involves


categorical judgement (Bowker 2006). These scientific judgements do not happen in a vacuum but are influenced by a complex entanglement of factors from the scientist’s own life experience and values, to the technical or political and economic context in which they work.

Specifically, we draw on the computer science expertise of the third author in our analysis of the cases that follow. We attempt to detail which kinds of data are being collected as well as analyze, where possible, the coding behind the machine intelligence using these big data, or the models these data are being ‘fed’ into. One complication is that corporate datasets and algorithms are a pernicious black box, full of assumptions that cannot be questioned or challenged by social scientists on the outside of the corporation. The way the law has developed around personal data has allowed copyright protection around the original selection and arrangement of databases, and trade secrecy allows companies to recover damages from anyone (for example a rogue employee) who discloses confidential data and code (Pasquale 2015).4 Corporations (and lawyers) justify these data control practices not only as trade secrecy but also as part of their attempt to protect consumer data security. In some cases, therefore, without access to code or ‘raw’ data, we could comment only on what Bogost (2015) calls the ‘procedural rhetoric’ or the implicit or explicit argu-ment a computer model or algorithmic decision makes.

We analyzed original datasets and models where available, but also patent appli-cations and technical webinars describing the decision platforms. We combine our analysis of the technical material with qualitative interview data. The first author conducted interviews with designers of smart farming technologies working in the private and public sectors in Canada and the U.S. Scientists were asked open-ended questions about their practice as well as their motivations and goals in relation to agriculture and technology. Participants were computer scientists, computational biologists, statisticians, geo-position specialists and agricultural engineers; all were recruited via purposive sampling that was ‘snowballed’ from a few initial partici-pants selected from existing networks. Interviews were almost entirely conducted by phone between January 2016 and June 2018 and they lasted anywhere from 60 minutes to three hours. Transcribed recordings of the interviews were coded induc-tively, organizing key themes using the software program NVivo.

We were guided by the following questions in our analysis: What farm sizes and strategies do the data collection and computer programming decisions target? What problems are these agricultural big datasets and algorithms or machine learning being enlisted to solve? What or who is being left out of this scientific decision-making?

RESULTS

Case 1: Private Sector Agricultural Big Data and Analytics

Farmers Edge was founded in Manitoba, Canada, by Wade Barnes, who grew up on a several thousand-acre grain farm, an average size for the Midwest. According


to their Smart Farming website,5 Farmers Edge focuses on providing farmers with digitally driven advice that will allow them to ‘manage risks’ and ‘maximize profita-bility.’ What they offer is a customized data collection service using a private system of weather stations (installed on one’s own farm), proprietary ‘telematic devices,’ or plug-in sensors that can be installed on farm machinery, and what they describe as highly ‘reliable’ remotely collected environmental data from a private constellation of satellites (via a partnership with Planet Labs). According to Barnes, ‘reliable’ data is that which is ‘up-to-date data exactly when farmers need it most.’6 Timing is of crucial importance in farming, and this is increasingly the case given weather variabilities presented by global climate change. Farmers need to be able to perform tasks like seeding or harvesting quickly—when the weather and other conditions are just right. This is especially necessary on farms with large acreages, because even fractions of a second saved on tasks like seeding can make a big difference over thousands of acres.

While Wade Barnes identifies the origins of his company’s advantage and success as fully privatized activities of data collection, storage and in-house data expertise, there is actually a more complex relationship between private and public, especially in relation to remote collected environmental data. ‘Anthony,’ an analytics officer in the government of Canada who conducts environmental and crop mapping, described this relationship and its implications at length. A primary characteristic of this relationship is that innovation is often unequally shared between public and private entities. He explained that while public institutions contribute a great deal of data and innovation to private enterprise, they are often barred from accessing certain advantages that private data offer, such as higher resolution data (for instance, via Plant Labs satellite data). Anthony said that scientists like himself working in public contexts face prohibitive costs that prevent them from accessing and using such data. He argued that, at the same time, however, public sector innovation—for example, historic data collection from government sensors and satellites stored and accessed from North American Space Agency (NASA) and U.S. Geological Survey data repositories—and contemporary data sets, such as the maps he creates and openly publishes, are essential for research and development in private contexts. While the data Anthony creates is publicly available, he is also aware of structural biases and inequities surrounding its utilization. ‘These maps … they’re interesting and people will use them and they get distributed around, but if you really want to use [our data] for something neat and something meaningful then you have to take the data and you have to use the data, download the data, and have the experience and software to do something with it. We know that industry is using these maps and data.’

The ‘industry’ to which Anthony is engaged is largely composed of companies like Farmers Edge. How then, and for whom, does Farmers Edge use this data? Our analysis shows that they draw from their private data collection and centralized data storage and management alongside algorithms and modeling to provide what they call ‘decision support’ in the form of recommendations that are delivered to farmers via a FarmCommand application about how, when and where in the field to perform tasks. Thus far, FarmCommand developers have targeted data collection only on


a small selection of major agronomic commodity crops. Because of their low per-unit value, agronomic crops such as corn, canola, or soy are typically planted on extremely large acreages. Moreover, for a host of reasons (land cost and environmental varia-tion) there is a significant fragmentation in land acreage across North America and, subsequently, large disparities in the market for agricultural machinery. Prairie farms are large, with 26 percent cultivating more than 3,525 acres (Statistics Canada 2016: Table 32-10-0156-01). For example, the average farm in Saskatchewan, Canada, is 1,700 acres, versus 261 acres in Nova Scotia (Statistics Canada 2016). Because of the lower per-unit value of commodity crops, larger farm size does not directly correlate with high income earnings for many prairie farms. Indeed, 57 percent, or 70,245 farms, report an annual farm operator income under Can$49,000 (Statistics Canada 2016: Table 32-10-0027-01). However, large holdings over 1,000 acres and flat prairie environments are well suited to expensive pieces of equipment for several reasons. This is because these farms often have more capital available for investment than on smaller farms. Additionally, large acreage production systems thrive on the scale and scope of efficiencies that large tractors with data collection sensors can offer. Finally, these tractors have been specifically designed for large acreage agri-culture, where turnaround at the edges of fields is easier and there is little variation of topography. In effect, Farmers Edge digital innovations are targeting mid-western North American farms that follow a ‘productivist’ and ‘industrial’ farming strategy (Buttel 2003; Friedman and McMichael 1987). One scientist working for the corpo-ration explained that it is precisely these operators who are the company’s ‘bread and butter.’ ‘Edna,’ an engineer working for one of the world’s largest agribusinesses, said that her corporation simply cannot cater to smaller, diverse farmers, because ‘you would have to look at how many small farmers have to work full time to pay for farming … So … hmmm … how could the corporation produce these for smaller farmers who have no money to pay for the technology?’ Instead, she said, this corpo-ration is supporting the modernization of conventional agriculture:

We want to showcase modern conventional agriculture and this includes the reasonable usage of pesticides and GMOs and technology to make better decisions, to be more effi-cient … it’s about not putting more fertilizers in a certain area without needing to. In some cases the software can help you calculate which areas of your field are not profitable at all … it’s more efficient.

Beyond a bias toward a particular set of crops and the farmers that grow them, addi-tional biases emerge by looking at the patent application for one specific decision support tool under FarmCommand called N-Manager. An analytics tool, N-Manager is designed to ‘analyze’ different environmental conditions and management prac-tices and then to provide nitrogen application recommendations to farmers. The basic premise is that data-driven machine learning can optimize fertilizer use to increase yield and reduce chemical waste and byproducts. A bias in this application is that it only works for farms using chemical inputs (that is, not organic farms) since the stated intent behind the development of N-Manager is to help these farmers reduce


such inputs. Moreover, much of the specifics of the modeling used for N-Manager are outlined in a U.S. patent application (Folle et al. 2020), which reveals that the application relies on public crop models such as DSSAT (Hoogenboom et al. 2019a, 2019b; Jones et al. 2003).

DSSAT (Decision Support System for Agrotechnology Transfer) is a public crop-modeling platform that allows growers to input information relating to weather, soil, genetics, management practices and pests, and gives information on the treatment of different varietals. Every crop model within DSSAT was developed through extensive field trials and other experimentation largely conducted by the public sector agricultural research community. This kind of agronomic science is very costly, because it requires, as one scientist we interviewed said, ‘boots on the ground.’ Yet this work gets funded via individual research labs that apply to govern-ment granting agencies. DSSAT handles genotype (genetic makeup) information. The DSSAT system either explicitly contains experimentally obtained parameters describing a plant species, ecotype or cultivar, or it has methods for performing sta-tistical approximations of a unique cultivar borrowing analogously from parameters of similar plant models. This is where biases become apparent.

The DSSAT system has a variety of models for many types of plant species. However, the available genotype information for each plant species is skewed towards broadacre commodity crops, such as those which serve as feedstock. Corn is by far the most modeled with 170 cultivars, but the list also includes 55 cultivars of rice, 54 of millet and 51 of soybean. In contrast, genotyping of non-commodity crops is less common (with the number of cultivars in parentheses): alfalfa (1), bahia (1), bermudagrass (1), greenbean (1), sugarbeet (2), cabbage (2), faba (2), tomato (4) and pepper (5). Completely absent are models for cultivars of fruiting trees (pears, apples or avocados), leafy greens (kale, arugula), as well as most types of squash. The Farmers Edge N-Manager thus relies on DSSAT which represents a partial supply of models. However, the patent application (and much of the Farmers Edge promotional material) refers specifically to maize. Stated differently, Farmers Edge has chosen to focus their algorithmic design efforts related to the building of N-Manager on corn, and has thus designed N-Manager to support the large-scale and input intensive farm operations growing commodity corn. This ‘productivist’ bias is compounded by an additional factor: making good on the advice given by N-application can only happen by using variable rates of nitrogen within a field, which requires extremely expensive ‘variable rate’ equipment. Such equipment is not only unaffordable for most small farmers, but it’s also largely irrelevant to their needs and production systems.

For those seeking to advance research and innovation for small-scale and organic farming systems, data biases and access constraints work against them. Megan, a public sector computer scientist working at a U.S. land grant institution to develop digital tools for organic farms, explained that she is working to use digital sensors and imaging software in order to develop a greater variety of plant models and more in-depth information on field performance in diverse environments. However, her work is strongly limited due to the lack of access to sophisticated computing:


The whole system of, like, the varietal descriptions right now, are based off of a not very deep understanding … right now the only standard for what you’re told about the variety is the germination percentage, which doesn’t really tell you very much about how that seed will perform in the field. So, that’s why I’m using that imaging software. There is much more well-developed imaging software that’s available, but I don’t have that available to me. So, I’m trying to use that with kind of limited amount of technology I have.

Another smart farm offering from Farmers Edge is their soil texture prediction tech-nology, which is part of their ‘Laboratory Services.’7 The soil texture application is a way of classifying the types of particles in a sample of soil and their relative material composition (such as clay, sand, silt). Quantification of soil texture is an incredibly useful way to characterize the suitability of soil for growing certain types of crops. In fact, soil-related measurements are an important input to crop-modeling. A recent patent application by Farmers Edge outlines a strategy that uses near-infrared spec-troscopy and machine learning algorithms to quickly and inexpensively make soil texture predictions. All farmers test soil, though in Canada this has historically been conducted by private enterprises and accredited by provincial ministries.8 Farmers interested in testing their soil for moisture and organic and non-organic compounds must send samples by mail to an accredited facility. Even the university laboratories (such as at University of Guelph and Dalhousie University) are public only in the sense that they are housed within publicly owned and managed infrastructure and prioritize tests for public agencies such as the Canadian Food Inspection agency over the commercial testing. Viewing this existing system as inexpedient, Farmers Edge executives decided to build their own soil-testing infrastructure, which includes this machine intelligence. Key here is that the machine learning algorithm requires data to learn from. While data collection is always a selection process, this is arguably even more so with soil data. Soil texture differs not only globally but also within the same local environment and collecting enough variety in soil data to generate complete or comprehensive ‘training’ dataset for machine learning is infeasible. The patent application does not explicitly state how the data was (or will be) collected. However, the patent does specifically mention a dataset collected using samples from ‘the Canadian prairies.’ Farmers Edge’s soil texture tool is thus partial, with both the dataset and the machine intelligence learning from it skewed toward farms operating within a narrow range of soil conditions.

Additionally, as farmers use Farmers Edge decision support systems, Farmers Edge continues to collect soil samples, ‘crowdsourced’ from use. In growing their data, the corporation continues to improve the algorithm over which they have intellectual property. While benefits from this improvement arguably accrue to the farmer, the predictive analytics advance as more data is collected, providing incom-mensurate gains for the corporation. If the method for quantifying soil texture is sound, and the data is of reasonable quality, then as data is accumulated by Farmers Edge (which is data provided by growers), the value of their offering also increases and their ability to dominate sections of a market improves. We see here a pattern seen in other sectors, where those who start collecting quality data early on have an advantage over other offerings in the market. We know this story well in the context


of social media and health data, where companies like Google and Amazon that have collected personal data for over a decade have an advantage in AI development. Indeed, ‘Rick,’ who works for a leading Canadian-based precision agricultural cor-poration, put it this way: ‘our biggest competitor is no longer Monsanto, it’s Google.’

Case 2: Public Computer Science Research Data and Machine Intelligence

Problems related to agriculture have gradually come to the attention of the broader artificial intelligence research community, and many public research datasets have been made available for research purposes. It is important to recognize that within the artificial intelligence research community, data is a critical component for making any scientific progress. As such, control over these datasets is very impor-tant, and any biases in public research datasets will shape the direction of research. The example we give here of data bias comes from plant agriculture research that is being conducted by computer vision scientists, and it centers around public research datasets focused on plant phenotyping. The dataset we explore is called the Agriculture-Vision dataset (Chiu et al. 2020), which has been made publicly availa-ble for a ‘Vision for Agriculture Workshop’ that will take place during CVPR 2020, described further below.

The dataset was formed as part of an industry/academia collaboration with indus-try partner, Intelinair, a company involved in smart agriculture using aerial imaging. Public and private sector research partnership in digital agricultural innovation is much like it is in the ‘tech’ sector at large, with fundamental research happening in the public sector while feeding talent and patented ideas into ‘start-ups,’ which then get acquired by large corporations. ‘Edna’ put it this way: ‘What [a large corporation] does is identify a technology that’s very good, and instead of spending money trying to do it in-house, they buy it, and in some cases they leave the branding as it is, and in others they do their own branding … in the case of John Deere it gets painted green!’ Governments are actively fostering this kind of private–public collaboration. ‘Rob,’ who at the time of the interview was developing a smart test farm in Canada, described the consortia of industry and government interests involved in the project this way:

The original proposal that we put forward was to work as part of a supercluster. We were looking for solutions for agriculture in the technology sector. The announcement came out for superclusters and we decided to take a shot at it and work with a group in Alberta. It sort of become an open innovation area where companies are coming together and maximizing their expertise in different fields that intersect with agriculture—so, remote sensing, auton-omous vehicles, data analytics, agronomy … the government funding helps bring those together … hopefully we can get some momentum going.

The Conference on Computer Vision Pattern Recognition (CVPR) 2020 is an important annual conference within the computer vision research community, which includes both private and public scientists. According to the conference website it is, ‘the premier annual computer vision event.’9 To underscore the importance of


this conference: CVPR accepted 1,294 of 5,160 paper submissions and had 9,227 registered attendees in 2019. Within the conference, dozens of workshops are often held. These workshops typically focus on niche areas within the computer vision community. The Vision for Agriculture Workshop is a new competition this year, which provides a dataset for computer vision practitioners interested in agricultural problems. The award for developing a model that performs well on this dataset is at least USD$5,000 for first place.

Exploring the details of this dataset, which attracts a significant amount of attention within this research community, reveals several important biases. The Agriculture-Vision dataset contains 94,986 aerial images of farmlands (presently the largest annotated public research dataset available for aerial agriculture imaging problems). Each image within the dataset is paired with expert annotated labels for what is called a ‘semantic segmentation problem.’ There are three important features to consider when analyzing bias within this dataset: crop identity, method of collec-tion and semantic categories. The crops within the fields which were imaged to create this dataset were predominantly soybean and corn. The images were collected using an unmanned aerial vehicle (UAV) or drone. The challenge requires that a model segment an image into the following seven categories: background, cloud shadow, double plant, planter skip, standing water, waterway, and weed cluster. Important among these semantic categories are the double plant and planter skip categories. A double plant is a planting error caused by an operator of a planting machine which occurs when a crop is double planted. Likewise, a planter skip is a planting error where a patch of agricultural land is missing crops. Given this information, the bias in the dataset becomes obvious—like the commercial decision platforms and the data which feed them, this dataset prioritizes corn and soybean farming, and in particular, farming operations that can afford expensive planting machinery and UAVs for analyzing their fields. Similar to what was outlined earlier with respect to FarmCommand, this dataset has the potential to escalate power disparities within the industry, favoring a particular type of food system which focus on large-scale operations selling commodity crops.

DISCUSSION

Scientific decision-making on smart farm technologies occurring in both private and public contexts favors the needs of farmers who follow a productivist strategy—growing commodity crops, on large acreages, and using expensive machinery. For example, the vast majority of agricultural big datasets that are being ‘fed’ into predic-tive analytics are collected only on major agronomic crops such as corn (versus hor-ticultural crops such as vegetables and fruits). This is a bias which also occurred with genetically engineered seed system development (Welsch and Glenna 2006). Our analysis shows that it is built into the long history of public sector plant modeling, which provides a foundation for current digital agricultural technologies. Moreover, big data-based biases have the potential to increase over time by ‘teaching’ artificial


intelligence systems from information generated on those agricultural operations deploying the decision platform/analytics tools. In the case of Farmers Edge soil detection machine intelligence, this means it becomes a tool which learns on farm data that is region-specific (to the Midwest part of North America).

First, from a purely economic perspective, this bias in the innovation system makes good sense. Prairie farms growing the bulk of grains, oilseeds, pulse, and corn feedstock on large acreages have the money to pay for expensive technologies, and agricultural corporations are driven by the goal of maximizing profit (Wolf and Wood 1997). But this bias presents several social and environmental problems, such as water pollution and soil degradation (Kibblewhite et al. 2008; Matson et al. 1997), biodiversity loss (Lynch 2009; Tilman et al. 2001), rising greenhouse gas emissions (Camargo et al. 2013; Lynch et al. 2011), exploitative labour practices (Basok et al. 2013; Preibisch 2010), human health concerns (Barański et al. 2014; Guthman 2011; Winson 2013), land grabbing (Akram-Lodhi 2015; Li 2011), and corporate concentration (Hendrickson and James 2005; Howard 2016). More specifically, climate adaptation literature emphasizes the value of small and diverse farms for food system resilience (IPES 2016) and achieving sustainable development goals (FAO 2018), yet dominant agricultural innovation trajectories appear to disadvantage these very farming strategies and important social actors (Gliessman 2013). And crucially, although small and diverse farms may not contribute as significantly as large ones according to aggregate economic measures, they are still, even in Canada and the U.S., the most numerous and thus crucial for society (see Machum 2005; Statistics Canada 2016). Said differently, the mid-western farms from which Farmers Edge soil datasets are grown and whose problems motivate agricultural engineers represent only 46 percent of Canada’s total farms (Statistics Canada 2016). That means more than one half of Canadian farms and the farmers they support are left out of the benefits that big data may generate. One empirical study with American agro-ecological and organic farmers reveals that they feel constrained by poor access to data appropriate to the diversity and local conditions of their farms (Ferguson and Lovell 2015). Moreover, the biases in big data and modeling we uncovered appear not only to result from but also reinforce the privatization of agricultural technology and information transfer (Piesse and Thirtle 2010), not least because agricultural data themselves represent an emergent form of capital. Given the potential economic value in agricultural data, corporations are rushing to collect and to control farm data. As our interviews show, not only farmers but also data scientists working on those issues relevant to organic and agro-ecological operations are limited by access to data and data infrastructure.

Second, the biases in agricultural big data and analytics feed into the long-standing shift in the relationship between capital and labor. Where once the laborer had a degree of control over the means of production and creativity, skill and the use of tools and knowledge, this control is further eroded within today’s smart farming context. Commercial platforms like FarmCommand are marketed to farmers as a way to let data drive farm decision-making, optimizing the limitations of a human’s ability to store and process information. Moreover, the ‘machine intelligence’ of data-driven


farming itself, as we have seen, currently encodes a very narrow expertise that is being built into machines and data systems. For example, much of the experientially derived knowledge of diverse farm environments and horticultural varieties is not being accounted for in the development of smart farming agricultural technologies. Proponents of smart farming boast that automated decision systems no longer require the brains or presence of laborers; yet the farm manager and/or laborer now not only becomes a ‘machine minder,’ but also helps to grow datasets without commensurate gain as information gets transferred away from labor and into capital (Harvey 2020). Data bias is wrapped up in a larger shift where knowledge, capacity and skill move from labor into fixed capital external to labor; science and technology is the entry for capital into the production process. This strategy works better for those food system players—notably historically powerful agribusinesses and powerful farmers—who have control over these technologies (including data as a technology) over others (Fine et al. 1994).

Third, although a logic of privatization (including limited data access) dominates this innovation landscape, corporations served by this logic are supported by public money and science, as well as farmer-collected data (Fielke et al. 2019).10 In Canada, for example, commercial digital agricultural tools are supported by public sector funding like the supercluster initiative, with digitization thus following the historic pattern of government privileging the goals of only a narrow band of agricultural stakeholders (Warner 2009). Given that much of the research and funding made available for digital agricultural innovation—agri-tech startups, for example—is publicly supported, our finding about bias presents a fundamental challenge to the democratic and inclusivity promise of the digital agricultural revolution.

Our analysis reveals something of additional value for scholars theorizing the role of political economy in food: Despite the obvious role of corporate interest on the smart farming and wider food system trajectory, there is a complexity that monolithic theories of privatization cannot account for on their own. The movement toward farm 4.0 is not simply a consequence of the dominance of private interest and power but rather a particular logic about innovation and farming that weaves together private and public R&D. For example, the patent application for N-Manager reflects biases toward productivist farming that exist within public sector plant modeling, which is then used by computer scientists designing analytics tools for predicting fertilizer needs. Further, the interviews show close associations between public and private sector research in the smart farming innovation pipeline. Harvey (2020) describes how technological innovation is both dependent upon and feeds back into capitalist interest, with governments having taken on the role of subsidizing industry research and development. To stay competitive, capitalist enterprises from agribusinesses to farms must use technological innovation and capture technical advancements prior to their competitors (and limit competitor access to such advantages). This helps to explain the commercial approach to private data collection and restricted data access as well as what appears to be a ‘rush’ to collect data on crops of economic importance.


Significantly, internal R&D is not the only way to achieve this competitiveness. Increasingly, companies deploy the strategy of technological capture through merger and acquisition as well as partnerships that allow for the capitalization of the fruits of public education. The impetus to ‘capture’ and win the technological race is not solely confined to corporate activity, but this technological dynamism is built into wider society; Governments are consistently seeking to improve technology, and to reward those who have the most advanced technology who are seen to ‘lead the pack,’ so to speak. Technological advancement and application across industries and uses becomes a business in and of itself. It is within this context that we can under-stand support systems like the Canadian superclusters. Government has not abdicated from funding innovation but rather has shifted its role to subsidizing what might be called the business of technology across political economic contexts. As the Minister of Innovation who launched the superclusters put it: ‘Superclusters are dense areas of business activity. Where people with diverse skills work together. Where companies large and small work with universities, colleges and technical schools … We want to speed things up. So we’re investing nearly a billion dollars to turn more ideas into solutions. More skills into jobs. And more companies into global successes’ (GOC 2018: n.p.).

CONCLUSION

We do not yet have a clear picture of the human impact of digitization in agriculture, but our findings indicate that the application of big data and AI to food production will likely further concentrate power in the hands of a few large farms and a few even more powerful agribusinesses. We now know that the application of previous agricultural technologies—from hybrid seeds to GMOs—has been beneficial for these same food system actors but deleterious for numerous smaller and ‘alternative’ producers. Farming aimed at producing commodities for export markets requires fewer farmers to meet its goals; over the past 100 years we have seen a steady and unwavering increase in large-scale farms with revenues greater than USD$250,000 per year, while we have seen a decline of smaller farms, whose operators have been unable to compete by taking on the economic risks associated with increasing pro-ductivity. This same pattern of consolidation has been followed by agribusinesses. According to the prevailing doctrine, which looks from the aggregate perspective and evaluates via market logic, this pattern is simply the natural work of forces like competition weeding out economically ‘irrational’ food system actors. But from the human perspective of many farmers this is a ‘crisis’: farmers commit suicide at higher rates than any other occupation in the U.S. (Weingarten 2018)—many work upwards of 60 hours per week to grow food but they make so little money that they cannot afford to buy groceries for their own families without also working second jobs off the farm. And the plight of these marginalized farmers and farm workers is inseparable from the fate of farming communities: The existence of at least some farms which grow food for more local markets is key to food system resilience, and


farming organically or agro-ecologically contributes enormous value to ecosystem health and the long-term fate of us all.

Taking up the COVID challenge we presented at the outset of the chapter: What would it mean for governments to drive innovation according to the alternative value system of sustainable, equitable, and inclusive growth and ensure that partnerships with business involving government funds are driven by public interest over eco-nomic gain? Perhaps instead of funding large firms (who buy up small ones) that design data-based analytics for crops like corn, governments could develop digital tools for a diversity of crops that are meaningful to smaller farmers engaging in prac-tices like crop and environmental diversification—which produce vital ecosystem services. Put simply, governments need to divest from large, commercially driven innovations as a means to solve complex socio-technical problems, as this will likely only exaggerate economic inequity across the food system.

The vast majority of government funding in agriculture is designated for expand-ing and commercializing the conventional farming industry. There is currently little research and data development for place-based and bioregional efforts to enhance socio-ecological diversity. But, in order to meaningfully address systemic inequities (including labor exploitation, racism and colonialism) within and beyond agriculture, government funds must be directed toward programs and tools that prioritize social and ecological health, as these aspects are consistently externalized by economic markets. Indeed, datasets could be built to capture and encourage more complex ecosystem variables and outputs for non-commercial farmers. Efficiency as a bench-mark of success for technical development could be replaced by complexity which might mean labor-supportive rather than labor-displacing end products. These prod-ucts should be publicly-driven and developed using priorities of worker health and well-being alongside agro-ecological diversity.

More specifically, decision-makers and data developers must ask: How can agri-cultural technology and data be used to move away from market-based agricultural land and property regimes and toward community designed and led governance processes, to build alternatives to intensification and commercialization for all farmers, and to develop support and resources for currently marginalized farmers and aspiring farmers (primarily migrant agricultural workers, farmers of color, and indigenous farmers). Now is the time for publicly-funded resource management and conservation programming alongside public investment in non-industrial farmers and community-based and -led innovation.

NOTES

1. Notable exceptions include Bronson 2018; Bronson and Knezevic 2016; Carbonell 2016; Carolan 2017; Eastwood et al. 2019; Higgins et al. 2017; Rose and Chilvers 2018; Rotz et al. 2019a; Rotz et al. 2019b; Wolf and Wood 1997.

2. Other scholars in this space look at less politicized uses of big datasets and algorithms, such as Eil Pariser’s (2011) work on Google Search.


3. Some attention has been paid to the potential that agricultural big data will exacerbate agribusiness power by given corporations like Monsanto ‘unique insights into what farmers are doing around the clock, on a field-by-field, crop-by-crop basis into what is currently a third or more of the US farmland’ (Carbonell 2016: 2; see also ETC Group 2016).

4. Also, the algorithms are often enormously complex and sometimes embody artificial intelligence that even their inventors have a difficult time fully understanding or explain-ing. Such cyberdrift might be even more disturbing than deliberately manipulated results.

5. See https:// www .farmersedge .ca.6. See https:// youtu .be/ RHhe8QS3bP0.7. See https:// www .farmersedge .ca/ farmers -edge -laboratory/ .8. See, for instance, http:// www .omafra .gov .on .ca/ english/ crops/ resource/ soillabs .htm.9. See http:// cvpr2020 .thecvf .com.10. The issue of lack of farmer access to agricultural equipment and data is another issue

beyond the scope of this chapter, but it is an issue. As one UN-funded report puts it: ‘despite the fact that a significant amount of existing agricultural data concerns the farmer, very little data flows to them. This issue is magnified where data about small-holder farmers is concerned’ (The Engine Room n.d.).

REFERENCES

Adamchuck, S., Viacheslav I. and Gebbers, R. 2010. Precision agriculture and food security. Science 327: 828–31.

Akram-Lodhi, A.H. 2015. Land grabs, the agrarian question and the corporate food regime. Canadian Food Studies/La Revue Canadienne des études sur l’alimentation 2(2), 233.

Barański, M., Średnicka-Tober, D., Volakakis, N., Seal, C., Sanderson, R., … Leifert, C. 2014. Higher antioxidant and lower cadmium concentrations and lower incidence of pesticide res-idues in organically grown crops: a systematic literature review and meta-analyses. British Journal of Nutrition 112(5): 794–811.

Basok, T., Bélanger, D. and Rivas, E. 2013. Reproducing deportability: migrant agricul-tural workers in south-western Ontario. Journal of Ethnic and Migration Studies 40(9): 1394–1413.

Bogost, I. 2015. Persuasive games: the expressive power of videogames. Cambridge, MA: MIT Press.

Bongiovanni, R. and Lowenberg-Deboer, J. (2004). Precision agriculture and sustainability. Journal of Precision Agriculture 5(4): 359–87.

Bowker, G. 2006. Memory Practices in the Sciences. Cambridge, MA: MIT Press.boyd, d. and Crawford, K. 2012. Critical questions for big data. Information, Communication

and Society 15: 662–79.Bronson, K. 2015. Responsible to whom? Seed innovations and the corporatization of agricul-

ture. Journal of Responsible Innovation 2: 62–77.Bronson, K. 2018. Smart farming: including rights holders for responsible agricultural innova-

tion. Technology Innovation Management Review 8: 7–14.Bronson, K. 2019. Looking through a responsible innovation lens at uneven engagements with

digital farming. Wageningen Journal of Life Sciences 90–91: 100294.Bronson, K. and Knezevic, I. 2016. Big data in food and agriculture. Big Data and Society 3: 1–5.Bucci, G., Bentivoglio, D. and Finco, A. 2018. Precision agriculture as a driver for sustainable

farming systems: state of art in literature and research. Calitatea 19: 114–21.Buttel, F. 2003. Environmental sociology and the explanation of the environmental reform.

Organization and Environment 16: 306–44.


Camargo, G.G.T., Ryan, M.R. and Richard, T.L. 2013. Energy use and greenhouse gas emis-sions from crop production using the farm energy analysis tool. Bioscience 63(4): 263–73.

Carbonell, I. 2016. The ethics of big data in big agriculture. Internet Policy Review 5: 1–13.Carolan, M. 2017. Publicising food: big data, precision agriculture, and co-experimental tech-

niques of addition. Sociologia Ruralis 57: 135–55.Chiu, M.T., Xu, X., Wei, Y., Huang, Z., Schwing, A., … Shi, H. 2020. Agriculture-vision:

a large aerial image database for agricultural pattern analysis, https:// arxiv .org/ abs/ 2001 .01306 (accessed 23 June 2020).

Cox, S. 2002. Information technology: the global key to precision agriculture and sustainabil-ity. Computers and Electronics in Agriculture 36(2): 93–111.

De Schutter, O. 2015. Don’t let food be the problem: producing too much food is what starves the planet. Financial Post (July/August), http:// foreignpolicy .com/ 2015/ 07/ 20/ starving -for -answers -food -water -united -nations/ (accessed 23 June 2020).

Eastwood, C., Klerkx, L., Ayre, M. and Dela Rue, B. 2019. Managing socio-ethical challenges in the development of smart farming: from a fragmented to a comprehensive approach for responsible research and innovation. Journal of Agricultural and Environmental Ethics 32: 741–68.

ETC Group. 2018. Year end status of the agribusiness mega-mergers, software v. hardware, Deere and Co. is becoming Monsanto in a box. Annual Report, December 2016, http:// www .etcgroup .org/ content/ deere -cobecoming -monsanto -box (accessed 1 September 2018).

Eubanks, V. 2018. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. New York: St. Martin’s Press.

Ferguson, R. and Lovell, S.T. 2015. Permaculture for agroecology design, movement, practice and worldview: a review. Agronomy for Sustainable Development 34: 251–74.

Fielke, S., Garrard, R., Jakku, E., Fleming, A., Wiseman, L. and Taylor, B.M. 2019. Conceptualising the DAIS: implications of the ‘Digitalisation of Agricultural Innovation Systems’ on technology and policy at multiple levels. Wageningen Journal of Life Sciences 90–91: 100296.

Fine, B., Goodman, D. and Redclift, M. 1994. Towards a political economy of food. Review of International Political Economy 1(3): 547–52.

Folle, S.M., Hernandez, J.A. and Hasan, M. 2020. Method and system for estimating effective crop nitrogen applications. US Patent Application 2020/0068797.

FAO [Food and Agriculture Organization]. 2018. Family farmers must remain central to agroecology scale-up, http:// www .fao .org/ news/ story/ en/ item/ 1114099/ icode/ (accessed 23 June 2020).

Friedmann, H. and McMichael, P. 1987. Agriculture and the state system: the rise and fall of national agricultures, 1870 to the present. Sociologia Ruralis 29(2): 93–117.

Galt, R.E. 2013. Placing food systems in first world political ecology: a review and research agenda. Geography Compass 7: 637–58.

Gardezi, M. and Bronson, K. 2019. Examining the social and biophysical determinants of US Midwestern corn farmers’ adoption of precision agriculture. Precision Agriculture 21: 549–68.

Gitelman, L. and Jackson, V. 2013. ‘Raw Data’ Is an Oxymoron. Cambridge, MA: MIT Press.Gliessman, S. 2013. Agroecology and going beyond organic. Agroecology and Sustainable

Food Systems 37: 147–8.GOC [Government of Canada]. 2017. Budget 2017, skills and innovation, http:// www .budget

.gc .ca/ 2017/ docs/ bb/ brief -bref -en .html #section1 (accessed 15 September 2019).GOC [Government of Canada]. 2018. What is a supercluster? https:// www .ic .gc .ca/ eic/ site/

064 .nsf/ eng/ 08126 .html (accessed 23 June 2020).GOC [Government of Canada]. 2019. Protein industries supercluster. https:// www .ic .gc .ca/

eic/ site/ 093 .nsf/ eng/ 00012 .html (accessed 15 September 2019).


Guthman, J. 2011. Excess consumption or over-production? US farm policy, global warming, and the bizarre attribution of obesity. In R. Peet, P. Robbins, and M. Watts, eds., Global Political Ecology, 51–66. New York: Routledge.

Harvey, D. 2020. Anti-capitalist chronicles. Collective form of action. Podcast. April 16.Hendrickson, M. and James, H.S. 2005. The ethics of constrained choice: how the industri-

alization of agriculture impacts farming and farmer behavior. Journal of Agricultural and Environmental Ethics 18(3): 269–91.

Higgins, V., Bryant, M., Howell, A. and Battersby, J. 2017. ‘Ordering adoption’: materiality, knowledge and farmer engagement with precision agriculture technologies. Journal of Rural Studies 55: 193–202.

Holtslander, C. 2015. Losing our grip: 2015 update. National Farmers Union 36, https:// studylib .net/ doc/ 18899290/ losing -our -grip - - -2015 -update.

Hoogenboom, G., Porter, C.H., Boote, K.J., Shelia, V., Wilkens, P.W., … Jones, J.W. 2019a. The DSSAT crop modeling ecosystem. In K.J. Boote, ed., Advances in Crop Modeling for a Sustainable Agriculture, 173–216. Cambridge: Burleigh Dodds Science Publishing.

Hoogenboom, G., Porter, C.H., Shelia, V., Boote, K.J., Singh, U., … Jones, J.W. 2019b. Decision Support System for Agrotechnology Transfer (DSSAT) Version 4.7.5 (https:// DSSAT .net). Gainesville, FL: DSSAT Foundation.

Howard, P.H. 2016. Concentration and Power in the Food System: Who Controls What We Eat? New York: Bloomsbury Publishing.

IPES [International Panel of Experts on Sustainable Food Systems]. 2016. Report to the United Nations, June, http:// www .ipes -food .org/ images/ Reports/ U niformityT oDiversity _FullReport .pdf (accessed 1 February 2019).

Janssen, M., van der Voort, H. and Wahyudi, A. 2017. Factors influencing big data decision-making quality. Journal of Business Research 70: 338–45.

Jones, J.W., Hoogenboom, G., Porter, C.H., Boote, K.J., Batchelor, W.D., … Ritchie, J.T. 2003. DSSAT cropping system model. European Journal of Agronomy 18: 235–65.

Kibblewhite, M.G., Ritz, K. and Swift, M.J. 2008. Soil health in agricultural systems. Philosophical Transactions of the Royal Society B, Biological Sciences 363: 685–701.

Kitchen, R. 2014. The Data Revolution: Big Data, Open Data, Data Infrastructures and their Consequences. London: Sage.

Kloppenburg, J.R. 2005. First the Seed: The Political Economy of Plant Biotechnology. Madison, WI: University of Wisconsin Press.

Laforge, J., Fenton, A., Lavalée-Picard, V. and McLachlan, S. 2018. New farmers and food policies in Canada. Canadian Food Studies/La Revue canadienne des études sur l’alimen-tation 5: 128–52.

Li, T.M. 2011. Centering labor in the land grab debate. Journal of Peasant Studies 38(2): 281–98.

Lynch, D. 2009. Environmental impacts of organic agriculture: a Canadian perspective. Canadian Journal of Plant Science 89: 621–8.

Lynch, D., MacRae, R. and Martin, R.C. 2011. The carbon and global warming potential impacts of organic farming: does it have a significant role in an energy constrained world? Sustainability 3(12): 322–62.

Machum, S. 2005. The persistence of family farming in the wake of agribusiness: a New Brunswick, Canada case study. Journal of Comparative Family Studies 36(3): 377–90.

Martinson, O.B. and Campbell, G.R. 1980. Betwixt and between: farmers and the marketing of agricultural inputs and outputs. In F.H. Buttel and H. Newby, eds., The Rural Sociology of the Advanced Societies, 215–53. Montclair, NJ: Allanheld, Osmun.

Matson, P.A., Parton, W.J., Power, A.G. and Swift, M.J. 1997. Agricultural intensification and ecosystem properties. Science 277(5325): 504–9.


Mazucatto, M. 2020. The Covid-19 crisis is a chance to do capitalism differently. The Guardian, 18 March, https:// www .theguardian .com/ commentisfree/ 2020/ mar/ 18/ the -covid -19 -crisis -is -a -chance -to -do -capitalism -differently (accessed 24 June 2020).

Moseley, W.G. 2017. The limits of the new green revolution for Africa: a political ecology critique. Brown Journal of World Affairs 23: 2.

Noble, S.U. 2018. Algorithms of Oppression: How Search Engines Reinforce Racism. New York: New York University Press.

O’Neill, K. 2016. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. New York: Crown Books.

Pariser, E. 2011. The Filter Bubble. New York: Penguin Press.Pasquale, F. 2015. The Black Box Society. Cambridge, MA: Harvard University Press.Piesse, J. and C. Thirtle. 2010. Agricultural R&D, technology and productivity. Philosophical

Transactions of the Royal Society of London 365: 3035–47.Poppe, K., Wolfert, J., Verdouw, C. and Renwick, A. 2015. A European perspective on the

economics of big data. Farm Policy Journal 12: 11–19.Preibisch, K. 2010. Pick-your-own labor: migrant workers and flexibility in Canadian agricul-

ture. International Migration Review 44(2): 404–41.Qualman, D. 2011. Advancing agriculture by destroying farms? The state of agriculture

in Canada. In H. Wittman, A.A. Desmarais and N. Wiebe, eds., Food Sovereignty in Canada: Creating Just and Sustainable Food Systems, 22–42. Halifax, Canada: Fernwood Publishing.

Rose, D.C. and Chilvers, J. 2018. Agriculture 4.0: broadening responsible innovation in an era of smart farming. Frontiers, https:// www .frontiersin .org/ articles/ 10 .3389/ fsufs .2018 .00087/ full (accessed 1 February 2019).

Rossel, V. and Bouma, J. 2016. Soil sensing: a new paradigm for agriculture. Journal of Agricultural Systems 148: 71–4.

Rotz, S. and Fraser, E.D.G. 2015. Resilience and the industrial food system: analyzing the impacts of agricultural industrialization on food system vulnerability. Journal of Environmental Studies and Sciences 5(3): 459–73.

Rotz, S., Gravely, E., Mosby, I., Duncan, E., Finnis, E., … Fraser, E.D.G. 2019a. Automated pastures and the digital divide: how agricultural technologies are shaping labour and rural communities. Journal of Rural Studies 68: 112–22.

Rotz, S., Duncan, E., Small, M., Botschner, J., Dara, R., … Fraser, E.D.G. 2019b. The politics of digital agricultural technologies: a preliminary review. Sociologia Ruralis 59(2): 203–29.

Sonka, S. 2015. Big data: from hype to agricultural tool. Farm Policy Journal 12: 1–19.Statistics Canada. 2016. Census of agriculture, https:// www150 .statcan .gc .ca/ n1/ pub/ 95 -640

-x/ 95 -640 -x2016001 -eng .htm (accessed 1 February 2019).The Engine Room. n.d. Responsibly opening agriculture data, https:// www .theengineroom

.org/ responsibly -opening -agriculture -data (accessed 23 June 2020).Tilman, D., Reich, P.B., Knops, J., Wedin, D., Mielke, T. and Lehman, C. 2001. Diversity and

productivity in a long-term grassland experiment. Science 294(5543): 843–5.Warner, M. 2009. Civic government or market-based governance? The limits of privatization

for rural local governments. Agriculture and Human Values 26: 133–200.Weingarten, D. (2018). Why are America’s farmers killing themselves? The Guardian, 11

December, https:// www .theguardian .com/ us -news/ 2017/ dec/ 06/ why -are -americas -farmers -killing -themselves -in -record -numbers (accessed 23 June 2020).

Welsh, R. and Glenna, L. 2006. Considering the role of the university in conducting research on agri-biotechnologies. Social Studies of Science 36: 929–42.

Wolf, S. and Wood, S. 1997. Precision farming: environmental legitimation, commodification of information and industrial coordination. Rural Sociology 62: 180–206.

Winson, A. 2013. The Industrial Diet: The Degradation of Food and the Struggle for Healthy Eating. Vancouver, BC: UBC Press.

Documents

8. The human impact of data bias and the digital