Chapter One: Agriculture Under Compound Pressure

Part One — Why Does Agricultural Transformation Matter?

The question of this part: what problem is artificial intelligence supposed to solve in agriculture?

Discourse about digital agriculture often begins from the development pathways of its products, and from the technical capabilities available, and then goes looking for use cases in agriculture. This book takes the road from the other direction: it begins with the agricultural problem, with the decision and its consequence, and with the human being who may pay the price of an error or reap the fruit of a right choice; then it examines the technology, not as an answer that precedes the question, but as a possibility that must prove by evidence that it fits the problem, is feasible in reality, and is worthy of becoming part of the agricultural decision.

The first three chapters establish the ground on which the whole book stands. The first begins from agriculture under the weight of its compound pressures; the second asks when change becomes a genuine transformation that touches the decision and the outcome, rather than merely an old procedure that has moved from paper to screen; and the third places in the reader's hands a scale for evidence by which to distinguish a fact that has earned its right to trust, a claim repeated until it came to look like a fact, and a promise that still stands in need of proof.

These three chapters are the key to everything that follows. Whoever reads them will be able to read any applied chapter in this book critically; whoever skips them may find themselves confronted with numbers they do not know how to weigh.

Chapter One: Agriculture Under Compound Pressure

The Chapter's Message

Agriculture does not face a single problem that can be solved with a single tool. It operates under a compound pressure in which water scarcity, soil degradation, weather variability, energy, labour, prices, finance, and infrastructure all interact. The value of artificial intelligence does not begin from its capacity to predict, but from its capacity to make these relationships visible while the decision is still possible.

1. The Field as a Living and an Economic System at Once

When a digital system sees a patch of altered colour in a satellite image, it may describe it as a decline in a vegetation index. The farmer, however, asks: is the cause thirst, or salinity, or disease, or a difference in variety, or the shadow of a cloud, or a fault in the irrigation? The answer changes the action, the cost, and the risk. This distance between detecting a signal and interpreting it agronomically is the distance in which a system either succeeds or fails.

Recent reviews reveal the breadth of what artificial intelligence applications are able to see, predict, and automate, from field monitoring to yield estimation and the improvement of food quality. But they reveal, at the same time, a less glamorous and more decisive limit: technical capability does not turn into agricultural benefit unless it finds data worthy of trust, infrastructure able to carry it, human skills that know how to use it, and performance that holds up when it moves from one crop to another, from one season to another, and from laboratory to field [SRC001][SRC002][SRC003]. For this reason, the phrase "feeding a growing world population" may be a legitimate humanitarian goal, but it is not an automatic pass for every sensor or robot. Before the tool comes the specific pressure; before prediction comes the decision that can be changed; and before the promise come the window for action, the baseline, and the outcome that will say, with evidence, that the technology added real value.

Agriculture is a biological system; the plant does not stop growing because the network went down. It is an economic system; a farmer may know the best treatment yet lack the liquidity to buy it. It is an institutional system; a pesticide may be effective in a published study yet not be registered for that crop in the relevant jurisdiction. And it is a social system; a correct recommendation may be rejected because it does not accord with the user's language, or experience, or the distribution of labour within the household. Any design that deletes one of these dimensions produces an incomplete answer.

2. A Map of Agricultural Challenges and Their Transmission Pathways

The preceding table makes clear that a challenge does not turn into a benefit merely by adding "AI." Between the two lies a causal chain that must be tested: a valid measurement, an appropriate interpretation, an actionable option, a response at the right time, and follow-up on the outcome. If a single link breaks, the system may achieve no effect at all, even if the prediction model itself is accurate.

3. Climate and Rainfall: From Reading Averages to Engineering Decisions

The value of climate is not measured by what averages alone say, but by the decisions they enable in the right place and at the right time. A national average for rainfall may be statistically correct, yet of limited value when applied to a particular field. Likewise, a short-range forecast may help postpone irrigation or spraying, but it does not, on its own, provide a sufficient basis for planning a long-term investment in a perennial crop. Hence, a clear distinction must be drawn among three time horizons, each with its own questions, tools, and decisions:

• Operational weather: supports immediate decisions, such as the timing of spraying or harvest.
• The season: weeks to months, guiding the choice of variety, the determination of planting date, and the estimation of input requirements.
• The climate: years and decades, testing a crop's long-term viability, the efficiency of water infrastructure over the long run, and the feasibility of investment.

The error begins when a system treats these horizons as one and the same; information that is correct at one temporal scale may become misleading when transferred to another. The cost of this confusion rises in arid regions, where the marginal value of every unit of water is higher, and where the requirements of pumping, desalination, or cooling may shift the burden from the water resource to energy without actually eliminating it. For this reason, it is not enough to describe a production system as "more efficient per kilogram." Relative efficiency alone does not reveal the magnitude of total pressure on resources. The decisive question is: did efficiency gains lead to an expansion of cultivated area or an increase in production cycles, such that total water withdrawal rose despite improved performance per unit of product? The separation between unit efficiency and total consumption is not an accounting detail, but a fundamental condition for evaluating real impact. This matter will return as an analytical axis in the chapters on irrigation, greenhouses, and sustainability. Studies conducted in some regions of the world indicate that water scarcity, dependence on external supplies, and cold chains are not independent files, but interlinked links in a single system of food and resource security. Nevertheless, many figures circulated in those studies differ according to the definition of the market, the time period, and the scope of measurement; they are therefore not suitable for constructing a final regional figure that can be generalised with confidence. What can be retained scientifically is the structure of the decision itself: a rigorous comparison should balance domestic production, crop type, water and energy intensity, the robustness of imports, and the capacity of stocks to absorb shocks. From this perspective, not every increase in protected agriculture counts as an automatic gain for sustainability; its real value is determined by what it achieves at the level of the entire system, not by what isolated efficiency indicators suggest.

Strategic conclusion: climate does not become actionable knowledge except when it is read at the correct scale, and efficiency does not become sustainability except when it reduces the overall pressure on resources, rather than redistributing or expanding it in another form.

4. Soil: A Slow Challenge That Is Hard to See

Not every challenge is as loud as a flood; some creep in silently and accumulate slowly, as happens when soil structure degrades. In this context, digital systems tend to give priority to what is easy to measure repeatedly, while properties that change more slowly, or that cost more to analyse in a laboratory, may remain more decisive for understanding reality and making decisions. The repetition of soil moisture measurements, however dense, does not compensate for the absence of knowledge about its texture, or its salinity, or its effective depth. Likewise, a map of colour or spectral reflectance does not become a reliable analysis of fertility unless it rests on precise calibration and a representative sample. Here arises an editorial and technical risk of great significance: that the density of data should turn into an illusion of knowledge. An abundance of measurements does not necessarily mean completeness of understanding, just as precision of display does not guarantee soundness of inference. The sound method begins by building clear layers of certainty; the observation is recorded as it is, its source, unit, and timestamp are preserved, and then the inference is placed alongside it as an interpretation of it, not a substitute for it. And when the data are not sufficient, the system's recommendation should express this deficiency clearly: gather an additional piece of information, or carry out an on-site check, or complete the analysis with a representative sample. A trustworthy system does not fill gaps in knowledge with imagined numbers; it converts uncertainty into a practical step towards more precise knowledge and a more firmly grounded decision.

5. Labour: From the Spectre of Replacement to the Reality of Task Migration

Discussion of automation and the entry of artificial intelligence into the operational sphere is often reduced to a single anxious question: how many jobs will disappear? Yet this measure, for all its importance, does not reveal the deeper transformation taking place within occupations themselves. Automation does not always eliminate work so much as it redistributes its tasks, changes the skills required, and shifts responsibilities from one party to another. The task of initial visual inspection may pass from the worker to a camera, but this transfer creates other tasks in return: installing the device, cleaning it, verifying the validity of its alerts, repairing its faults, and documenting the exceptional cases it cannot interpret. The machine may lighten the burden of dangerous or repetitive work, yet it may at the same time create a new dependence on a specialised technician or a distant supplier. From this standpoint, the Food and Agriculture Organization's report on automation places the principles of inclusion and context-appropriateness at the heart of the transformation, warning that the effects of automation differ according to farm size, the resources available, and the institutional structure surrounding it [SRC031]. Accordingly, the evaluation of automation should not be confined to the job title, but must penetrate to the level of the task itself. This evaluation begins with a series of decisive questions:

• What task has changed, and how has it changed?
• Who performed it before automation, and who supervises it afterwards?
• What new skills has its completion come to require?
• Who bears responsibility when the machine breaks down or fails?
• Do physical risks actually decrease, or do they turn into exhausting burdens of mental monitoring?
• Do workers have a clear pathway for training, qualification, and transition into the new roles?

Responsible automation does not assume that the labour market will rearrange itself automatically, nor does it treat workers as a side effect of technical progress. It regards the transformation as an integrated institutional and human process, in which the movement of tasks, skills, and responsibilities is managed with the same care given to the introduction of machines.

6. Prices and Finance: When the Information Is Correct and Not Actionable

The farmer does not plant inside a bare "statistical average"; he makes his decisions in a specific season, within a limited budget, and under the weight of existing obligations. Therefore, the arithmetical correctness of a recommendation is not enough to guarantee its validity on the ground. An analytical tool may recommend a variety with a higher expected return, yet its need for early liquidity, or a limited capacity for storage, may make another variety more suitable and more viable. Likewise, investment in a particular crop may look profitable according to an analysis spanning five years, yet it remains out of reach for someone who cannot bear the down payment or survive a single season of loss. Hence, the economic challenge cannot be reduced to the price of the input alone; it is the product of an interlinked system of costs, risks, and constraints, including:

• price volatility and instability;
• the timing of payment and the availability of liquidity;
• the cost of borrowing and the terms of finance;
• risks that cannot be insured against;
• the time required for training and qualification;
• periods of interruption and the losses that follow from them;
• the availability of spare parts and the cost of their maintenance;
• the farmer's right to transfer his data when changing supplier.

These relationships reveal why a good technology may spread at a slow pace, even when its benefits are proven. Slow adoption does not necessarily mean that farmers are "resistant to innovation"; it may be a rational response to a promising technology that does not yet fit the reality of liquidity, or the structure of risk, or the conditions of operation. The real value of any recommendation is not measured by its theoretical correctness alone, but by the farmer's ability to implement it and to continue with it without exposing his farm to a financial drain he cannot withstand.

7. Holding Structure and Inequality in Access to Agricultural Technologies

Farms differ greatly in area, financial resources, the nature of production, and the capacity to use modern technologies. Studies concerning farm structure and land concentration indicate that holding sizes and their roles in production are marked by clear diversity [SRC037]. Nevertheless, judgements should not be built on farm size alone; small farms are not alike in their needs and capacities, just as large farms are not always the most able to adopt every technology or to benefit from it. But the size of the holding remains an influential factor in the economic feasibility of technology. The large farm can, in most cases, spread the cost of purchasing a device or a digital system over a wider production area, and may be better able to employ a specialised technician, absorb the losses of an unsuccessful trial, and negotiate with suppliers over prices and terms of data use. As for the small farm, the technology may be useful to it technically, yet not economically affordable if it is obliged to bear the full cost on its own.

An illustrative example: a shared service instead of individual purchase Suppose a group of farmers needs to use a drone to monitor crops and detect stress zones within their fields. Purchasing a separate drone, training an operator, and maintaining the device may not be an economical option for every farmer, especially if use is seasonal or limited. This can be addressed through a cooperative or a local extension service that owns the drone and operates it on behalf of a number of farmers. The operating body collects the data according to a clear protocol and provides each farmer with results specific to his holding, while guaranteeing his right to know how his data are used, retained, and transferred. By this mechanism, the cost of the device and its operation is distributed across a larger number of beneficiaries, without depriving the small farmer of the benefit the technology provides. Achieving fairness of access is not confined to the manner of owning devices; it also includes the design of the means of service delivery. Short text messages, or telephone calls, or recourse to an agricultural extension agent may be more suitable than a complex digital application that requires an advanced handset, permanent internet connection, and high technical skills. Therefore, agricultural modernity should not be measured by the number of devices in use or by their degree of complexity, but by their capacity to improve the decision, reduce risks, and reach different categories of farmers on fair terms. Successful agricultural technology is not necessarily the most advanced, but the most suited to users' needs and the most able to make its benefit available without making the size of the holding or the limitation of resources an obstacle to benefiting from it.

8. Historical Projections: When a Projection Survives and Loses Its Context

Some numbers acquire an authority that exceeds the documents in which they were born. They recur in reports and presentations until they seem to be a fixed truth untouched by time. Among the most prominent of these figures is the claim that the world needs to increase food production by 70% by 2050. Yet understanding this number requires returning to its original context: What was it measuring? From what point did it begin? And on what assumptions did it rest? In 2009, the Food and Agriculture Organization of the United Nations presented two interrelated pictures of the global food future, not a single number combining population growth and food production into one meaning. The first picture rested on the United Nations' medium-variant population projection in its 2007 Revision, and estimated a rise in world population from about 6.8 billion people at the time the document was prepared to 9.1 billion people in 2050, an increase of approximately 34%. The second picture concerned the volume of production required to meet projected demand in a world more populous, more urbanized, and with higher incomes, and under changing dietary patterns. According to this scenario, meeting that demand could require increasing global food production by about 70% by 2050, compared with the 2005–2007 average [SRC014].

What does the 70% figure measure? This figure does not mean that the world's population will grow by 70%, nor does it mean that the yield of every hectare must rise by the same proportion. It is the outcome of a long-term scenario that combined a set of interrelated variables, chief among them:

• population growth;
• rising income levels;
• expanding urbanization;
• changing patterns of food consumption;
• assumptions concerning resources, yields, and inputs;
• technological development;
• trade flows.

It is also essential to note that the document did not use a single reference period across all its indicators. While the production scenario adopted the 2005–2007 average as a general basis for comparison, certain calculations relating to calories and cereals drew on other reference periods. Reducing the entire document to a general phrase such as "the 2005–2007 baseline" therefore obscures the divergence among indicators and weakens methodological precision.

Why did the figure change from 70% to 60%? The Food and Agriculture Organization's 2012 revision reveals clearly that this figure was a historical projection open to revision, not a final scientific constant. The organization re-estimated the increase required over the same period at about 60% instead of 70%. The principal reason was that updated data showed base-year production to have been higher than earlier calculations had estimated.

The global food challenge did not suddenly recede by ten percentage points; rather, what changed, to a considerable degree, was the measurement from which the calculation began. When the base-year value rises, the percentage increase needed to reach the same future value, or one close to it, falls. This is a fundamental methodological point: a percentage cannot be interpreted, or compared with others, without knowing the starting point from which it was calculated. Subsequent work by the organization likewise used other estimates that proceeded from different base years, ranges, and scenarios. Direct comparison among these numbers therefore becomes misleading unless the definitions, limits, reference period, and underlying assumptions of each estimate are clearly stated.

Historical value, not a permanent description of the present The 2009 projection does not lose its value merely because it was revised. Its true value is historical and methodological: it reveals how the food security challenge was framed at a specific scientific and political moment, and how expectations about population, income, urbanization, and consumption were converted into a long-term quantitative estimate. But on its own it is not suited to describing the world in 2026, nor is it sufficient to justify the adoption of any particular technology. More importantly, the document itself did not reduce food security to increasing supply. Producing more food does not by itself guarantee the elimination of hunger; food may be available in markets yet remain out of reach for people who cannot afford it or who lack the means of access. Food security, then, is not understood through the volume of production alone, but also through the actual capacity to obtain food.

A disciplined number begins with its context The danger lies not in a number's remaining in memory, but in its remaining alive after its context has died. Disciplined scientific use of any projection therefore requires clarifying:

• the year in which it was issued;
• the time horizon it targets;
• the variable it measures;
• the baseline used;
• the most important assumptions on which it was built.

When describing present reality, one should turn to the most recent official source available, rather than recycling a historical projection as though it were a diagnosis of the present. A dated number is an important witness to the thinking of its era; a number stripped of its date may be more forceful in rhetoric, but it is less valuable in producing knowledge. For this reason, the phrase "increase production by 70%" should not become a rhetorical bridge across which one leaps directly from a global challenge to a sensor, an algorithm, or a robot. The need for technology is not established by an abstract global figure; it begins with a specific agricultural decision, evidence that can be tested, and an effect that can be measured.

9. What Does Artificial Intelligence Actually Add to the Agricultural Decision?

The real value of artificial intelligence does not begin when it tries to replace agricultural expertise, but when it widens the field of vision and gives decision-makers a better chance to understand and to intervene at the right time. The algorithm does not know the field as the farmer or the specialist knows it, but it can detect signals difficult to track manually, gather scattered evidence, and convert it into estimates open to examination and practical use. This value takes shape in four interrelated functions:

1. Observation: widening the scope of vision

Artificial intelligence can capture subtle changes or patterns that may escape the eye, or that are difficult to detect manually at the same speed and frequency across wide areas. Such signals may include:

• the early onset of water stress;
• an unusual change in plant colour;
• a disturbance in animal behaviour;
• a decline or change in the performance of an agricultural machine.

The value lies not in collecting images and readings for their own sake, but in converting observation into an early alert that draws attention to a place worth examining before the problem widens or the window for intervention is lost.

2. Estimation: moving from measurement to a specific prediction

The estimation function begins when scattered measurements are converted into a probability or a prediction clear in subject, place, and time period. A system might estimate the risk of a disease outbreak over the coming days, or predict water requirements in the next irrigation cycle, or define a likely production range. But a professional estimate does not present the result as though it were a confirmed fact. Its value is complete only when it makes clear:

• the degree of uncertainty attached to it;
• the conditions under which its use is valid;
• the limits beyond which it may lose accuracy;
• the period and place to which it applies.

A useful forecast, then, is not a bare number, but a number surrounded by the conditions of its interpretation and the limits of confidence in it.

3. Coordination: gathering evidence within a single decision context

Agricultural information is usually distributed across multiple sources: field sensors, prior records, weather forecasts, agronomic rules, available resources, and economic and regulatory constraints. Artificial intelligence can gather these elements and present them within a single context that helps the person responsible see the full picture. The value here does not stem from the abundance of data, but from ordering and connecting it well. A decision-maker needs to understand the relationship among three essential matters:

1. What is happening in the field?
2. What do the resources and constraints permit?
3. What action can be carried out in the time available?

When information is organized in this way, data is transformed from an accumulating archive into a practical basis for making decisions.

4. Feedback and operational learning: turning outcomes into improvable knowledge

The system's task does not end with issuing a recommendation. The more important step is comparing what it observed, predicted, or recommended with what actually happened in the field, then using the discrepancy to improve both the model and agricultural practice. This does not mean allowing the system to learn or alter its behaviour without oversight. What is intended is establishing a disciplined review cycle that reveals:

• points of error in the prediction or the recommendation;
• potential sources of bias;
• changes that have occurred in operating conditions;
• cases requiring an update to the model;
• agricultural procedures in need of adjustment.

Through this mechanism, error becomes not a mere technical failure but information that helps improve subsequent performance, within clear human oversight and defined responsibilities. From prediction to decision: where does human responsibility remain? However accurate these functions become, they do not replace agricultural understanding, nor do they absolve the human being of responsibility for the decision. An algorithm may indicate a heightened probability of disease, but it cannot on its own determine whether the sign warrants an urgent field inspection, or whether the cost of delay exceeds the cost of a false alarm, or whether the proposed intervention is lawful, safe, and available on the farm in question. An algorithm may recognize a pattern in the data, but the meaning of that pattern is not determined within the data alone. It is shaped in the context of soil, climate, crop growth stage, and field history, and is influenced by the farmer's capacity to respond and by the time, resources, and executable options at their disposal. For this reason, a system's maturity is not measured by its ability to conceal the distance between prediction and decision, but by its ability to make that distance clear and manageable. A mature system makes plain:

• what it observed;
• how it arrived at its estimate;
• its degree of confidence in the result;
• the limits of its knowledge and the validity of the model;
• the consequences of error, whether a false alarm or a failure to detect a genuine risk.

At that point the algorithm becomes a partner in seeing, not a substitute for agricultural judgment. It alerts before the window for intervention narrows, and it gathers evidence scattered across multiple sources, but it leaves the authority of choice and the responsibility for it to those who understand that the field is not merely a set of numbers, but a living system with its own context, limits, and consequences.

10. A Checklist: How Is the Agricultural Problem Framed Before Buying the Technology?

A successful technical project begins by identifying the decision that needs improving, not by selecting the device or the algorithm. Before purchasing a product, contracting with a vendor, or funding a project, clear written answers should be prepared to the following questions:

First: The Decision and Expected Value

• What agricultural decision do we want to improve? And who currently makes it?
• What loss are we seeking to reduce, or what opportunity do we want to capture? And can it be measured within a defined time period?
• What baseline and reference benchmark will we use to compare the technology's results? Is it current practice, an expert estimate, a simple rule, or non-intervention?

Second: Data, Its Quality, and Its Ownership

• What data is available for making the decision? And is it sufficient and fit for purpose?
• Who owns this data, and who has the right to access and use it?
• What known errors or gaps does it contain?
• Can the user export their data when changing the system or the vendor?

Third: Errors and Risks

• What is the consequence of a false alarm, when the system indicates a problem that does not exist?
• What is the consequence of a false negative, when the system fails to detect an existing problem?
• What harm might arise from a delayed alert or intervention?
• Who bears responsibility for following up on results and handling error?

Fourth: The Actionability of the Recommendation

• Is the action the technology proposes actually available?
• Is it lawful and compliant with regulatory requirements?
• Can it be funded and carried out at the appropriate time?
• Are the labour, skills, and resources needed to apply it available?

A recommendation that is correct in computational terms has no value if carrying it out is impossible in practical or financial terms.

Fifth: Follow-up and Safe Reversal

• Who follows up on what happened after the recommendation was carried out?
• How is the outcome measured and compared with the baseline?
• What is the mechanism for reversal if the system fails or the results reveal unexpected harm?
• Can one safely return to the previous procedure, or does the system create a dependence that is hard to undo?

Sixth: Operational Continuity and User Rights

• Does the workflow continue when the internet connection is cut?
• Can it be operated in the vendor's absence, or when technical support cannot be reached?
• Can the user understand the basis on which the decision was built?
• Can they object to the recommendation or override it when it conflicts with field reality?

If clear answers to these questions are absent, the problem is not that the model or product has not yet been selected; it is that the decision project itself has not been designed. Technology does not automatically repair a problem's vagueness; it may merely give it a shinier interface and a higher cost. The professional starting point is to define the decision, measure its value, and understand its risks, and then to select the technology that serves it—not to reshape the problem to fit what the technology offers.

Chapter Summary

Compound pressure does not necessarily justify compound technology. The better intervention may be improving the irrigation record, training the worker, repairing a cold chain, or making an official source available in comprehensible language. Artificial intelligence may be the link that binds these tasks together and improves their timing and precision. Judgment is not issued from the name of the technology, but from its effect on a specific decision chain and from its ability to remain safe and economically feasible in the real world.

Evidence Notes

The mapping of the field draws on the reviews [SRC001][SRC002][SRC003], on the official report on comprehensive automation [SRC031], and on the analysis of farm structure [SRC037]. The projection [SRC014] has been retained solely as a historical document. The regional files were used to extract the structure of the problems and questions; no market or food-security figures were carried over from them that were not closed in the source register.