Digital twins are coming to agriculture. We should be careful what we mean by that.
Lately I’ve noticed a number of posts suggesting that digital twins may be the next big thing in agricultural technology. It’s been quite the topic at conferences too.
You can see why the idea is appealing. If you can connect together machinery telemetry, farm management records, sensor feeds, satellite imagery, weather data and operational history, then you can assemble it into a digital representation of the farm. Once that representation exists, you can simulate scenarios, run predictions and support better decisions.
It’s a good story. Perhaps a little too good to be true.
Agricultural technology has moved through a sequence of these waves over the past decade. First came the promise of better farm software. Then data platforms. Then connected sensors. Then Blockchain provenance. Now AI and perhaps, digital twins.
Each wave arrives with its own evangelists and language. Each suggests that agriculture is on the verge of becoming fully legible to software.
And yet farm businesses have a habit of resisting the good stuff.
The concept of a digital twin comes from industries where systems behave in mostly predictable ways.
A jet engine, for example, is a tightly bounded physical system. It is governed by physics that engineers understand well. It produces a dense stream of telemetry. Sensors are calibrated, monitored and trusted. A digital model can be built and continuously updated with real data, allowing engineers to simulate behaviour, predict failures and optimise performance.
Some parts of agriculture do look similar to this.
I think modern farm machinery is the most obvious example. Tractors, sprayers and harvesters produce large volumes of operational data about engine performance, machine state and usage. They also collect a shed load of operational data like yield, chemical usage etc. Companies like John Deere have spent years building connected equipment platforms centred around operations centre that already resemble aspects of a digital twin.
Controlled environments such as greenhouses and some aquaculture systems are another example. These are closer to industrial processes than open biological systems. Sensors are dense, conditions are controlled, and modelling becomes more feasible.
In these domains the digital twin concept fits quite nicely and I think could prove interesting.
Once we move from machinery to the farm itself, things become a lot more complicated.
A farm is not a single engineered system. It is a biological one. Soil chemistry, rainfall patterns, plant physiology, pest pressure, genetics, markets and human decisions all interact across time.
Even something that sounds straightforward like understanding crop performance in a paddock, quickly becomes a large web of causes.
What happened before sowing. When rainfall arrived relative to fertiliser. Whether nitrogen was available when the plant needed it. Whether disease pressure arrived at the wrong moment.
Agronomic models can approximate parts of this system, but and despite many very smart people dedicating their careers to understanding them, they are always approximations. Many of the variables that matter most are either poorly measured or in many cases not measured at all.
A digital twin implies a level of fidelity that is difficult to achieve in systems like these.
A lot of the digital twin narrative assumes farms are already producing dense, reliable and structured data.
Sometimes they are. More often they are not.
In some situations the problem is fragmentation. Information is spread across machinery platforms, livestock systems, agronomy tools, spreadsheets and accounting software. In others the problem is more basic: the measurements needed to understand the system simply do not exist, or they exist at low quality. Sensors drift out of calibration. Observations are recorded late or not recorded at all. Important context lives only in someone’s head.
This matters because a twin built on incomplete signals is not really a twin. It is a partial reconstruction.
That may still be useful, but it is worth being honest about what kind of thing being built.
At this point it is fair to point out that farm management software already records much of the operational history people say is missing.
And that is true.
Good farm management systems can tell you what happened in a paddock last year, what chemicals were applied, where animals were moved and how much feed was allocated. These systems are valuable, and many farms rely on them every day.
But they rarely capture the entire context of what happened around those activities.
The rainfall event that preceded a spray pass. The machinery telemetry from the operation itself. Changes in livestock weight during the same period. A paddock observation recorded during a walk. A satellite signal indicating plant stress.
Pieces of this story live across multiple systems. Some are structured. Some are not. Some never get captured at all.
The result is that the operational history of a farm is usually fragmented.
However, when you think about it, there is another digital twin on every farm already.
It lives in the head of the farmer.
Experienced producers constantly simulate their systems mentally. They remember past seasons. They notice small changes. They run comparisons against earlier years. They recognise patterns long before a model could. They have gut feel.
This mental model is built from observation, memory and judgement accumulated over decades.
Any digital system that claims to replicate or replace it needs to clear a very high bar.
There is also a practical question that tends to be missing from digital twin discussions.
What decision becomes sufficiently better to justify the cost?
Building and maintaining a true digital twin requires sensors, infrastructure, integration and constant calibration. Data pipelines must be maintained. Models must be validated. Systems must interoperate.
Agriculture has made progress here, but interoperability and standards are still patchy and to many they’re boring..
Technology that succeeds in farm businesses usually does so because it solves a very specific problem clearly enough that the benefit is obvious. Large infrastructure layers struggle when the value to producers is indirect or delayed.
I’ve been trying to think a bit differently about this problem.
Rather than beginning with the idea of a perfect digital representation of the farm, we could start with something more grounded: understanding what actually happened.
What sequence of events led to this outcome. What conditions existed at the time decisions were made. How different parts of the system interacted across the season.
In practice that means focusing on capturing and connecting operational events across the farm things like machinery activity, livestock measurements, environmental signals, management actions and observations etc and keeping them in a form that preserves their context over time.
Once that operational history or context exists in a coherent way, other tools become far more useful. Models can run against it. Analytics can explore it. And increasingly, AI systems can reason across it to surface patterns and explanations that would otherwise remain hidden.
The goal is not to build a perfect mirror of the farm.
It is to make the farm system more understandable.
Digital twins may well become useful in certain pockets of agriculture. Machinery systems already show how that might work. Controlled environments are another obvious candidate. Some aspects of livestock monitoring may eventually move in that direction too.
But agriculture has always rewarded technologies that begin with the practical realities of farming rather than the vocabulary of the technology sector.
The most valuable digital tools will probably not be the ones that promise a virtual copy of the farm.
They will be the ones that help producers understand their systems more clearly, learn from what has already happened and make fewer mistakes the next time around.
Sometimes thinking ahead of the hype cycle is simply a matter of asking a better question.
What problem are we actually trying to solve?