After a season that reduced margins and punished mistakes, an engineering perspective, with its potential and its limitations, on the technologies that are transforming operational information into agriculture’s new asset.
The 2025/26 season left a lesson that goes beyond the cherry sector, even though this was precisely the segment that paid the highest price. According to Frutas de Chile, shipments closed at around 114 million boxes, down nearly 10% compared with the previous season.
Yet, despite the lower volume of fruit shipped by sea, economic returns did not recover as expected. Something has broken in the equation that supported the industry for a decade.
Producing more no longer guarantees higher profits. The industry itself has acknowledged this: analysts at Copefrut summarized the new scenario by explaining that the challenge is no longer to spend less, but to spend better.
According to the seasonal assessment by agricultural economist Jorge González of INIA Quilamapu, toward the end of January one kilogram of 4J caliber cherries was being sold in China for around 10 dollars (8.59 Euros), while L caliber reached barely 2.6 dollars (2.23 Euros).
Almost four times the difference for the same species, in the same week. The market no longer simply pays for fruit: it pays for measurable attributes such as caliber, firmness and quality conditions, penalizing those that do not have them.
Rabobank points in the same direction: the outcome no longer depends on how much is produced, but on how, when and under what conditions the fruit reaches its destination.
All of this does not concern cherries alone. Table grapes must meet extremely precise caliber standards, blueberries compete with origins capable of offering firmer fruit, while avocados, citrus fruit and nuts operate with margins that leave no room for inefficiencies.
The conclusion applies to everyone: when margins narrow, information stops being a luxury and becomes a productive asset. This gives rise to the question at the heart of this article: what can artificial intelligence actually offer to this need to know more and estimate less?
Let us start with the evidence: Chilean agriculture is not a technologically backward sector. Orchards are monitored through weather stations, moisture probes and leaf analyses.
The facilities of major exporting companies use state-of-the-art optical graders, such as Unitec, Tomra or MAF Roda systems, capable of photographing every fruit from all angles and classifying it by caliber, color and defects at speeds impossible for the human eye.
At the heart of modern processing plants, automated measurement has already won. The blind spot lies in everything around it.
Throughout the entire journey of the fruit, a surprising number of decisions continue to be based on estimates: how much fruit is actually present in a plot before harvest; in what condition each lot arrives at receiving; what caused the rejects; what yield each grower achieved.
The industry itself is questioning the problem. In technical industry media, such as Smartcherry, the issue of the representativeness of manual sampling at receiving is raised season after season.
A bin contains around 200 kilograms of fruit from different trees and areas; yet the entire lot is classified on the basis of a small sample.
In day-to-day management, industry suppliers themselves point out how a substantial part of operations continues to depend on paper and isolated Excel spreadsheets.
The best demonstration that the problem is real and widespread is the emergence of an entire Chilean industry of software for agriculture whose commercial proposition consists, literally, of eliminating manual spreadsheets.
The paradox is clear: the machine at the center of the line measures every fruit with precision in the order of tenths of a millimeter, while many of the decisions surrounding it continue to be made through sampling, visual experience and Excel spreadsheets.
This is not negligence: until recently, there was no economically sustainable way to carry out these measurements continuously. This is precisely the space now being contested by new measurement technologies.
Before continuing, a clarification is necessary. The author is not an agronomist: I am a civil mechatronics engineer and I work on the development of industrial measurement systems capable of turning what happens in the field or along a production line into useful data for decision-making.
The readers of this magazine know fruit much better than I do; what I can offer is a different perspective: to explain, without marketing filters, what lies behind the acronym “AI”, what it can measure within an agricultural operation and what it cannot do.
Before continuing, it is useful to clarify some concepts, because many of the frustrations associated with this technology arise precisely from confusion between terms. Artificial intelligence is not a single technology: it is the umbrella term covering systems capable of performing activities that we associate with intelligence.
Under this umbrella, almost everything that works today belongs to the same discipline, machine learning or automatic learning: programs that learn patterns and correlations from data instead of following manually written rules.
To assess its practical applications, it is essential to distinguish the type of data being used: text, numbers or images.
The first application works with language. It is generative artificial intelligence, the kind used by ChatGPT and its equivalents, and today it represents one of the most immediate opportunities to improve the management of any agricultural business with virtually no investment.
Analyzing trends contained in settlement sheets, drafting and reviewing protocols and documentation for audits, or improving communications in English with customers and importers.
Its limitation is clear: it works with information that someone has already recorded. It cannot see the field or the fruit.
The second application works with numerical data and is the most underestimated. This is classical machine learning: models capable of identifying patterns within sets of data.
Many companies have accumulated years of irrigation telemetry, climate data and settlement records that have never truly been used to their full potential.
This historical asset enables very concrete applications: correlating the water and climate performance of each plot with final yield and caliber; identifying anomalies, such as an area that begins behaving differently from others before the symptom becomes visible; or forecasting harvest curves based on the company’s own history rather than using the regional average.
The limitation is not the algorithm: it is the availability of truly meaningful data. An orderly historical dataset that fails to consider the effect of the season, particularly severe pruning or a change in rootstock can generate useless forecasts.
Value emerges when numerical data is combined with the agronomic context known only by field personnel. Companies that have been recording data consistently for years are sitting on an asset they are still not fully exploiting.
The third application is the one capable of seeing: computer vision. And here it is useful to eliminate a common misconception: anyone managing a modern processing facility already lives with computer vision, even if they may not call it that.
Optical graders are exactly this: integrated vision systems, closed and designed for a single purpose, installed inside the machine and extraordinarily efficient at performing one specific task.
The innovation of the last decade does not lie in the technology itself, which has been used for decades for the inspection of industrial components. The real change is that deep neural networks have acquired the ability to recognize objects within messy real-world scenes characterized by shadows, dust and overlapping fruit.
Thanks to this development, computer vision has become independent of the machine: today it is possible to install a camera with a trained model virtually anywhere in the operation, at a fraction of the cost of a sorting line, to perform tasks that no grader can cover.
In practice, it transforms what it sees into data through four operations: detection — is there a fruit here? —, counting — how many have passed? —, measurement — what are its dimensions? — and inspection — does it have a visible defect? —.
They seem like simple tasks, but they are precisely the ones that make it possible to address the blind spot described earlier.
Its operation leads to the most important practical consequence. A vision system is not simply “an intelligent camera”: it is a complex engineering system.
On one side there is the hardware: industrial cameras, equipped with lenses and lighting systems specifically selected for the scene and designed to work under the real conditions of a field or packing facility, environments that are far from being similar to a laboratory.
On the other side there is the model: a program that has learned to recognize objects by analyzing thousands of previously labeled images. The intelligence does not reside in the camera lens, but in the model, and the model learns exclusively from what it is shown.
A system trained on perfectly illuminated apples in a European laboratory may fail when faced with avocados piled on a Chilean conveyor belt, at the actual processing speed and with the backlighting of a warehouse.
The consequence is that, in computer vision, the concept of “plug and play” does not exist. This is why serious projects begin by collecting data directly from the line or the customer’s field.
And the criterion for evaluating any supplier, including the author, should be only one: they must demonstrate that the system works with your fruit and within your operation, not through a catalog video.
In the orchard. The most established application is also the easiest to describe: counting. Carrying out an inventory of plants across large areas using aerial images is now essentially a solved problem and makes it possible to replace inventories carried out on foot, which in companies covering hundreds of hectares are slow, costly or even impractical.
Likewise, vision applied to monitoring makes it possible to identify and quantify visible symptoms, from areas with stressed vegetation to catches in traps, transforming phytosanitary monitoring into quantifiable data rather than a simple impression gathered in the field.
Several levels higher, however, lies the sector’s most publicized and least fulfilled promise: forecasting exportable production by counting flowers or fruit months in advance.
Counting what the camera sees is one thing; predicting how much will actually end up inside a package destined for export after passing through fruit set, weather conditions, agronomic management and harvest is something completely different.
The camera sees the plant structure and the number of flowers, but it does not see the physiology that will determine how many of those flowers the plant will manage to carry through to harvest.
That step cannot be solved by an image: it requires agronomy. Anyone claiming they can do so should be able to show verified results from previous seasons, on the same species and in the relevant geographical area.
At receiving. This is where it is decided which lot can withstand 30 days of ocean transport and which must be processed immediately and where, as we have seen, the sample analyzed is extremely small.
Computer vision does not replace the judgment of the experienced receiving operator: it multiplies their capacity for observation. Visually recording the entire lot, rather than only the handful of fruit that can pass across the inspection table, completely changes the statistical basis on which the decision is made.
There is also another particularly important advantage at this stage: a good system does not merely measure, but also preserves visual evidence of what it has detected.
At a point in the supply chain where decisions are discussed with the grower, being able to show the fruit directly rather than just a number transforms potentially disputed data into a decision supported by evidence.
On the line. The grader already measures what passes through it. The contribution of computer vision concerns everything happening around it: measuring flow before and after the sorter to identify bottlenecks that prevent full use of a machine worth millions.
Counting units, boxes and pallets at points where counting is still carried out manually today; and, above all, transforming waste into timely information.
In many facilities, the causes of rejects are recorded manually and become available when the work shift has already ended. A system capable of classifying rejects in real time instead makes it possible to inform field personnel, during the same harvest, about which practice is causing the damage, rather than discovering it only the following day.
Toward the destination. The 2024/25 season provided a harsh reminder of the cost of being unable to demonstrate the condition of the fruit: the case of the Maersk Saltoro, the vessel that suffered a breakdown in the Pacific and arrived in China 52 days after departure, with around 5 million boxes of cherries on board.
The cargo was valued at around 100 million dollars (85.92 million Euros), according to reports in the specialized press. The fruit was declared unfit and destroyed.
When a situation like this occurs, or when a normal claim concerning quality conditions arises, the difference between recovering the loss and having to absorb it may depend on the availability of evidence: an objective, dated record of the conditions in which the fruit left the facility.
A systematic visual archive of each packed lot does not prevent the incident from occurring, but it radically changes the position of the exporter when dealing with insurance companies and importers.
This is where genuinely useful projects are distinguished from those destined to end up abandoned in some warehouse.
First: computer vision can only see what is visible. An internal defect or physiological condition that has not yet manifested itself on the surface cannot be detected through normal cameras, regardless of how much artificial intelligence is used.
There are complementary technologies, such as spectroscopy or near infrared, that make it possible to address some of these problems, but they belong to another category of solutions and involve different costs.
Second: geometry dictates the rules. There are configurations in which no camera can see what would need to be observed: fruit hidden by foliage or objects covered inside a compact load.
In an industrial project that I had the opportunity to evaluate, the conclusion was that certain positions within the load were invisible with any configuration of cameras, and the correct decision was to communicate this before selling a system that was inevitably destined to fail.
A serious supplier must be willing to explain when a problem cannot be solved through computer vision; it is advisable to be wary of anyone claiming that their product is capable of solving everything.
Third: sometimes the solution is not artificial intelligence. Some variables can be detected more effectively using a simple, robust and inexpensive sensor rather than a deep learning model.
The correct question should never be “how can I apply AI to this problem?”, but rather “what is the simplest and most reliable way to measure it?”.
Sometimes the answer is a camera combined with a trained model; in other cases, a thermometer installed in the right place is enough.
Fourth, and this is the most important limitation: artificial intelligence cannot replace an early detection activity that was never carried out.
The most significant figure from recent seasons comes from INIA Quilamapu: the brown marmorated stink bug accounted for 70% of phytosanitary rejections of cherries destined for the United States in the 2024/25 season and 67.6% in 2025/26.
Every rejection involves fumigation that can deteriorate fruit quality or redirect it toward markets of lower value.
There are tools offering partial support, such as traps equipped with cameras that automate monitoring or the expansion of visual sampling during receiving, but no technology applied at the port can correct what was not detected in the orchard.
The lesson is not that artificial intelligence can autonomously solve the problem of the brown marmorated stink bug. The point is that, when late detection has consequences of this magnitude, it is no longer reasonable to continue relying exclusively on minimal samples and rapidly performed inspections.
1. Which important decision is still made “by eye”?
A technology project should not begin with technology, but with a specific decision that today is made on the basis of insufficient information: how many people to assign to an activity, which lot to ship, or how to determine the settlement for each grower.
If this decision cannot be clearly identified, there is not yet a real project, but only a need that should be explored further before investing.
A simple estimate may be enough to determine whether it is worth proceeding. If the losses caused by the error do not greatly exceed the cost required to improve measurement, the investment will hardly be justified.
But when significant price differences linked to caliber are at stake, or when it is impossible to document a claim concerning the condition of the goods, the question stops being whether it is worth measuring and becomes when it is necessary to start doing so.
Without your own information, there is no model capable of guaranteeing reliable results. It is one of the fundamental rules of any computer vision project.
The advantage is that acquiring images from a processing line or a plot is generally quite simple, involves limited cost and does not necessarily interfere with operations.
One line, one plot, one season. Technology projects that fail in agriculture almost always share the same initial mistake: they started at too large a scale.
A pilot project with precise objectives and a success metric defined in advance makes it possible to validate the solution, correct any errors and gather the information needed before proceeding on a larger scale.
The latest season showed that the market does not necessarily reward those who produce more, but those who better understand what they have.
Knowing how much product is available, the condition it is in, and having the information needed to prove it has, for decades, been a capability concentrated in the automated core of large processing plants.
Artificial intelligence, and computer vision in particular, is extending this capability to the rest of the supply chain, from the orchard to the pallet leaving the facility.
It is not magic and does not claim to be: it is a new measurement tool for an industry in which information and experience have always been decisive.
The next season will once again favor those who know with precision what they have before the destination market is the one to tell them.
The question is no longer whether this technology works. The question is which decision within your business finally deserves to stop being made by eye.
Image source: Stefano Lugli
Kevin Figueroa
Founder of NeuralCrops
01 Oct 2024
One of the key elements of this growth is export innovation: the Cherry Express programme, created about five years ago, has reduced transit times between Chilean ports and China by 10 days, ensuring that the fruit arrives in optimal conditions.
21 May 2024
‘In the south it will be remembered as a very good season. The flowering was very good in 90 per cent of the orchards, reaching production levels not seen in a long time,' says Sebastián Navarro, agronomist and cherry grower in the south of Chile.
11 Sep 2026
Artificial intelligence is transforming fruit growing in Chile: computer vision, machine learning and data help measure quality, yields and waste, reduce errors and improve decisions from the orchard to packing lines and export markets, including cherries and other fruit crops.
11 Sep 2026
In Apulia, HTS identified Pomes virus Greece (PVGR) in sweet cherry for the first time in Italy. The virus was detected in 8.7% of the samples analysed, while one case of Cherry green ring mottle virus also marks the first regional report of CGRMV in southern Italy.
