For generations, agronomic knowledge has been produced on research stations like AU Flakkebjerg, where fields are designed to answer specific scientific questions. This is where agriculture became measurable, comparable and reproducible. But a quiet shift is underway. Today, experiments are escaping the confines of the research station, turning ordinary farm fields into laboratories and farmers into active contributors to agricultural research.
“For a long time, agronomic experiments lived on research stations,” says Takashi Tanaka, Tenure Track Assistant Professor at the Department of Agroecology at Aarhus University. “But farmers have always experimented. They just didn’t call it research.”
Today, those everyday experiments, changing fertiliser rates, seed density, or management strategies across a field, are increasingly generating vast amounts of data. Yield monitors, GPS-guided machinery, variable-rate technology and even smartphones now record what happens on real farms, under real conditions. The challenge is no longer collecting data, Takashi Tanaka argues. It is making sense of it.
That challenge is the focus of Takashi Tanaka’s latest invited review, “ Advanced Data Analytics for On-Farm Experimentation ”, which synthesises more than a hundred scientific studies on how modern statistics, machine learning and simulation can unlock the potential of on-farm experimentation (OFE).
At stake is nothing less than how future agronomic knowledge is produced, who it is produced for, and for which purpose.
Classical agronomy is built on small, carefully designed plot trials. For nearly a century, the principles introduced by Ronald Fisher, randomization, replication and blocking, have underpinned agricultural science. These experiments, conducted under controlled conditions, are powerful tools for isolating cause and effect.
But they also have limitations.
“Small-plot experiments are excellent for understanding mechanisms,” Takashi Tanaka explains. “But they struggle with external validity. Farmers don’t farm on research stations.”
Real fields are messier. Soil properties change gradually across space. Weather varies from year to year. Machinery follows tramlines. Treatments are applied in long strips rather than neat squares. As a result, data from on-farm experiments often differs from the assumptions behind classical statistics.
If one fertiliser treatment happens to coincide with better soil, was it the treatment or the soil that increased yield?
This question has long made researchers cautious about on-farm data. Yet, as precision agriculture technologies have spread, on-farm experimentation has become increasingly common. Farmers now routinely test practices directly in their own fields, generating data at spatial scales and resolutions that were unthinkable a few decades ago.
“What has changed,” Takashi Tanaka says, “is that we now have tools that can handle this complexity, if we use them carefully.”
One of the central messages of Takashi Tanaka’s review is that method matters. Using the wrong analytical approach can produce confident-looking results that are fundamentally misleading.
Linear mixed models, for example, have emerged as a key tool for analysing spatially heterogeneous field data. By explicitly modelling spatial correlation, these models can to some extent separate treatment effects from background variability.
“Mixed models are not magic,” Takashi Tanaka cautions. “They don’t fix a bad experiment. But combined with sensible design, like replicated strip trials, they can greatly reduce bias.”
Bayesian approaches go a step further by quantifying uncertainty directly. Instead of asking whether one treatment is significantly better than another, Bayesian models estimate probabilities: What is the chance that a given practice will increase yield or profit under these conditions?
This probabilistic framing, Takashi Tanaka argues, is often more intuitive for farmers and advisors than classical p-values.
“In farming, uncertainty is unavoidable,” he says. “Weather, prices, pests, everything changes. Bayesian methods allow us to incorporate that uncertainty instead of pretending it doesn’t exist.”
Alongside advances in statistics, machine learning (ML) has rapidly entered agronomic research. Random forest models, neural networks and other algorithms excel at capturing nonlinear relationships in large, complex datasets. In on-farm experimentation, they offer the tantalising possibility of optimising inputs like nitrogen fertiliser at fine spatial scales.
But here, Takashi Tanaka also urges caution.
“Most machine learning models are very good at prediction,” he explains. “But agronomic decisions are about causation.”
A model may predict yield accurately while completely misunderstanding why yield varies. If fertiliser rate happens to correlate with soil quality across farms, a machine learning model might attribute yield differences to fertiliser, even if fertiliser has little causal effect.
This distinction matters when farmers use models to decide how much fertiliser to apply. Overestimating response to inputs can lead to wasted inputs and environmental harm.
Recent research, reviewed by Takashi Tanaka and colleagues, shows that so-called causal machine learning approaches can improve estimation of treatment effects. Yet these methods are still developing, and no single algorithm works best in all contexts.
“The message is not ‘don’t use machine learning,’” Takashi Tanaka says. “It’s ‘don’t use it blindly.’”
One field, one season, one experiment on its own, this is rarely enough to generate robust recommendations. Weather variability alone can overwhelm treatment effects.
That is why Takashi Tanaka sees cross-site synthesis as one of the most important frontiers in on-farm experimentation.
By combining results from many farms and years, researchers can identify patterns that are invisible at the individual-field level. Meta-analysis, Bayesian hierarchical models and regional machine learning approaches all offer ways to integrate dispersed experiments into coherent evidence.
“These methods allow us to move from anecdotes to evidence,” Takashi Tanaka says.
Importantly, cross-site learning also has social implications. When farmers see their experiments contributing to a larger knowledge base, and receiving feedback in return, participation tends to increase. Networks form, and learning accelerates.
“In that sense,” Takashi Tanaka adds, “on-farm experimentation is not just a technical system. It’s a social one.”
Much of the precision agriculture literature focuses on large, highly mechanised farms in Europe, North America and Australia. But globally, most farmers operate on a very different scale.
In Japan, for example, fields are often smaller than half a hectare. In parts of Africa and Asia, mechanisation is limited and yield monitors are rare. Does on-farm experimentation make sense in these contexts?
Takashi Tanaka believes it does, if the tools are adapted.
“Smallholder systems actually have a huge potential for learning,” he says. “But the technologies need to be affordable and accessible.”
Here, smartphones play a surprisingly important role. Recent studies show that yield can be estimated using smartphone images, computer vision and lightweight sensors. Farmers can collect data without expensive machinery, and experiments can be designed around local realities.
“These tools won’t replace all measurements,” Takashi Tanaka notes. “But they lower the barrier to participation.”
The result could be a more inclusive form of agronomic research, one that extends beyond well-funded research stations and large commercial farms.
One of the more counterintuitive elements of modern agronomic research is the use of synthetic data. When real-world experiments cannot reveal the true yield respons, because it is unobservable, simulation offers a way forward.
By generating artificial datasets with known causal relationships, researchers can test whether analytical methods actually recover the correct answers. Process-based crop models, such as APSIM or WOFOST, simulate plant growth under varying conditions and can produce realistic, spatially heterogeneous yield responses.
“These simulations are like flight simulators for agronomy,” Takashi Tanaka explains. “They allow us to test methods safely before applying them in real fields.”
Hybrid approaches that combine simulated data with real on-farm data are particularly promising. They can compensate for limited sample sizes and help models generalise across space and time.
Taken together, the developments reviewed by Takashi Tanaka point toward a quiet transformation in agronomic science. Knowledge is no longer produced only in controlled experiments and transferred to farmers. Instead, it increasingly emerges from iterative learning cycles that involve farmers, advisors, data scientists and researchers.
Yet this transformation also raises difficult questions.
How much complexity is useful before models become opaque? How should uncertainty be communicated? Who owns the data generated on farms? And how can recommendations remain trustworthy when analytical pipelines become ever more sophisticated?
“There is a risk of overconfidence,” Takashi Tanaka admits. “Advanced analytics can give very precise-looking answers. But precision is not the same as truth.”
For him, the solution lies not in choosing a single “best” method, but in combining approaches, comparing results, and grounding analysis in agronomic understanding and farmer experience.
“In the end,” he says, “the goal is not perfect models. It’s better decisions.”
Collaborators: Department of Agroecology at Aarhus University, University of Sao Paulo, and University of Nebraska-Lincoln.
Funding: This work was financially supported by the JST FOREST Programme under Grant Number JPMJFR221C.
Conflict of interest: None
Read more: The publication “ Advanced data analytics for on-farm experimentation: methods, challenges, and opportunities for modern agronomic decision making ” is published in Agronomy & Crop Ecology. It is written by Takashi Tanaka, André F. Colaco, Taro Mieno, and Simon S. Riley.
Contact: Tenure Track Assistant Professor Takashi Tanaka , Department of Agroecology, Aarhus University. Mail: takashi@agro.au.dk
Communications Advisor Camilla Brodam Galacho , Department of Agroecology, Aarhus University. Tel.: +45 9352 2136 or mail: brodam@agro.au.dk
Plant Production Science
Advanced data analytics for on-farm experimentation: methods, challenges, and opportunities for modern agronomic decision making
30-Apr-2026
No potential conflict of interest was reported by the author(s).