The newest frontier in potato storage is not simply better sensing. It is learning to recognise the faint, interconnected signals that precede rot, quality decline and potentially devastating commercial losses.
By Lukie Pieterse, Editor and Publisher, Potato News Today
For generations, successful potato storage has depended on a combination of engineering, crop knowledge and human attentiveness.
An experienced storage manager learns to recognise the warning signs: an unusual smell, a stubborn area of condensation, a temperature that does not settle as expected, a pile that seems reluctant to respond to ventilation, or a processing sample whose colour begins moving in the wrong direction.
This accumulated human knowledge remains indispensable. But it has always faced one serious limitation: by the time a problem can be smelled, seen or confirmed through routine sampling, biological deterioration may already be well established inside the pile.
A newly launched Canadian data initiative points towards a different model—one in which the potato store begins warning its operator before the damage becomes obvious.
The idea is deceptively simple. Historical storage failures contain information. If the environmental and operational data recorded during the days and weeks preceding those failures can be collected, organised and analysed, recurring patterns may emerge. Those patterns could eventually enable machine-learning systems to recognise rising risk in an operating store.
This is the premise behind the Rotten Potato Data Project, launched in July 2026 by Canadian agricultural technology company Cellar Insights.
The project is inviting growers and storage operators to contribute digital records associated with significant historical storage incidents. These may include rot, hot spots, physiological breakdown and other forms of quality deterioration. The records will be converted into a de-identified library of real storage events and matched with the sensor and controller data that preceded them.
The immediate objective is not to automate the storage manager out of the building. It is to provide that manager with an earlier and more focused warning.
Why historical failures matter
Artificial intelligence is often discussed as though an algorithm can be pointed at a problem and instructed to solve it. In agriculture, reality is much less convenient.
A machine-learning model must first be shown enough meaningful examples. It needs to see not only normal storage behaviour, but also verified examples of what happened before a real problem developed.
This is where historical spoilage records become valuable.
Temperature, relative humidity, carbon dioxide, ventilation activity and other measurements may each tell part of the story. A single elevated reading, however, does not necessarily mean that a pile is deteriorating. Conditions naturally fluctuate as the crop moves through wound healing, pull-down, holding and reconditioning.
The more important information may lie in the relationship between variables:
- Was carbon dioxide rising while a particular zone remained unusually humid?
- Did temperature decline more slowly in one part of the store than elsewhere?
- Did a ventilation adjustment temporarily suppress a warning without correcting the underlying biological activity?
- Was an unusual gas signature detected before temperature changed noticeably?
- Did the same combination of signals occur before confirmed spoilage in another store?
A properly trained model can examine thousands—or eventually millions—of these relationships more consistently than a person reviewing separate graphs.
That does not make the machine more knowledgeable than the operator. It makes it capable of identifying complex or easily overlooked patterns and directing human attention towards the place where it may be needed most.
The pile is a living biological system
The commercial value of predictive storage becomes clearer when we remember that a stored potato crop is not an inert commodity.
Tubers continue to respire. They release heat, moisture and carbon dioxide. They heal wounds, consume stored carbohydrates and gradually advance in physiological age. Disease organisms may remain relatively quiet before favourable conditions allow them to develop rapidly.
The crop entering storage is also highly variable. It may contain differences in maturity, skin set, pulp temperature, bruising, soil load, disease exposure and field stress.
Agriculture and Agri-Food Canada advises that potato pulp temperature at harvest should generally fall between 10°C and 18°C. Colder tubers are more susceptible to bruising, while potatoes harvested above 18°C face greater danger from disorders including bacterial soft rot and Pythium leak.
That initial condition follows the crop through the storage door. A building cannot completely undo the effects of heat, drought, disease, immature skin or rough handling.
Storage management must therefore interpret the biological consequences of the growing and harvesting season—not merely control the temperature of the surrounding air.
From averages to zones
One of the weaknesses in conventional monitoring is the temptation to rely on averages.
A store may report an acceptable average temperature while concealing substantial differences within the pile. A small area of wet, damaged or diseased potatoes can behave very differently from the remainder of the crop.
The average may look reassuring precisely because most of the store is still healthy.
Predictive monitoring is potentially valuable because it can examine changes by zone and across time. Instead of asking whether the building’s average conditions remain within the target range, it can ask whether one area is beginning to deviate from its own expected behaviour.
This is an important distinction.
The earliest warning of trouble may not be a reading that exceeds a fixed threshold. It may be a gradual divergence—a section of the pile that cools more slowly, retains moisture longer or responds differently whenever fans operate.
These are the quiet anomalies an attentive operator tries to notice. Data analysis can help make them more visible.
Gas sensing adds another layer
Temperature and humidity are fundamental storage measurements, but they are not always the earliest indicators of biological deterioration.
Cellar Insights has been developing systems that combine wireless monitoring with measurements of carbon dioxide and volatile compounds associated with potato breakdown. Its machine-learning approach examines these signals alongside conventional storage data.
This matters because decomposing potatoes can release characteristic gases before the problem becomes detectable through human smell or visible inspection.
Gas sensing should not be interpreted as a magical detector capable of locating every defective tuber. Air movement, sensor position, pile density and the nature of the disorder all influence what reaches a sensor.
But when gas information is combined with temperature, humidity, carbon dioxide and controller history, it can strengthen the overall picture. Several modest changes occurring together may offer a more credible warning than any one sensor could provide independently.
The value of an earlier warning
An alert has no value unless there is still time to make a decision.
If a deterioration problem is identified early enough, an operator may be able to:
- Inspect a particular bin or section of the store.
- Alter ventilation or air-distribution strategies.
- Investigate moisture accumulation or airflow restriction.
- Increase the frequency of crop sampling.
- Separate or move a vulnerable lot.
- Adjust the planned delivery schedule.
- Consult a storage specialist before the problem escalates.
- Document whether corrective action is improving or worsening conditions.
Not every spoilage event can be prevented. Some problems entering storage are already too advanced, while others may be difficult to reach or manage without disturbing a substantial volume of crop.
Even then, an earlier warning can support a better commercial decision. Knowing that a lot is unlikely to hold until April may allow it to be moved in January while it still meets market or processing requirements.
The economic benefit may therefore come from avoiding a total loss, protecting grade, preserving fry colour, reducing shrink or preventing a deteriorating lot from compromising adjacent potatoes.
AI should advise—not pretend to know
The enthusiasm surrounding artificial intelligence should not obscure its limitations.
A predictive system does not understand potatoes in the way an experienced grower or storage manager does. It recognises statistical patterns in the information available to it.
Its reliability will depend on several factors:
- The quality and calibration of the sensors.
- The completeness of historical controller records.
- Accurate identification of when and where a problem occurred.
- Representation of different varieties, regions, storage designs and crop conditions.
- The ability to distinguish correlation from a genuinely useful warning signal.
- Independent testing under commercial conditions.
- Clear communication of uncertainty and false-alarm rates.
A model trained predominantly on one storage design, variety or production region may not perform equally well elsewhere. An anomaly may indicate genuine biological deterioration, but it could also reflect a damaged sensor, a changed control setting or a temporary airflow condition.
The system should consequently be treated as decision support, not as an unquestionable authority.
The storage manager still needs to examine the crop, understand its history and judge whether the recommended response makes biological and commercial sense.
Data ownership and grower trust
The project also raises a broader issue that agriculture can no longer avoid: who benefits from farm-generated data?
Storage records may reveal sensitive information about crop losses, management decisions, infrastructure performance and commercial vulnerability. Growers will understandably be cautious about sharing them.
Cellar Insights says submitted information will be treated confidentially and de-identified before being used to develop and evaluate its models. The company also says individual farms, storages and incidents will not be publicly disclosed.
That commitment is essential, but the longer-term industry discussion must go further.
Growers contributing valuable operational data should receive clear answers about consent, ownership, retention, secondary use and commercial benefit. They should know whether their information may be combined with other datasets, whether it could be used to develop future products, and what protections remain if the business owning the platform changes hands.
Trust is not a footnote to agricultural data systems. It is part of the infrastructure on which those systems depend.
Turning private misfortune into shared intelligence
There is something unusually practical about learning from failed storage seasons.
Growers naturally prefer to forget a cellar that developed serious rot or a crop that had to be moved early. Such events can carry financial loss, frustration and occasionally a sense of personal failure.
Yet a storage failure is rarely meaningless. It leaves behind a trail of environmental and operational information. If those records can be handled responsibly, they may help reveal the early patterns that were invisible at the time.
That does not eliminate the loss already suffered. It does, however, allow hard-earned experience to become something more than an isolated misfortune.
The larger significance of predictive storage lies here.
Potato storage has traditionally relied on the knowledge of individuals, specific farms and particular regions. Data-supported systems offer an opportunity to retain that local wisdom while also learning across many stores and many seasons.
The store of the future will still need good insulation, correctly designed airflow, reliable refrigeration, careful sanitation and competent people. No algorithm can compensate for poor physical infrastructure or badly damaged potatoes.
But alongside those fundamentals, the future store may acquire something genuinely new: a memory.
It may learn what trouble looked like before anyone could see it—and speak soon enough for someone to act.