Raimat, sensors and harvest: IRTA aims to forecast the crop
19 November 2026 · sommelier.vi.cat
Decisions arrive before the harvest does: organising work, anticipating grape deliveries and considering the resources required. The difficulty is making those decisions before the crop is known. Research that seeks to reduce that uncertainty deserves attention. So does resisting the temptation to describe every camera accompanied by artificial intelligence as a crystal ball with a battery.
What has been announced
On 18 November 2026, IRTA reported YIELD4CAST trials at Raimat. Field sensors, drones and instrumented vehicles are being combined to develop and validate crop forecasts. The collaboration sits within the IRTA–Raventós Codorníu joint research unit. The announcement describes ongoing work; it does not provide a final accuracy figure or demonstrate water savings.
The official project record gives a September 2025–August 2029 timetable and identifies Jordi Gené-Mola as its lead. Its approach combines remote sensing, computer vision and crop models, from individual plots to regional scale. This is therefore neither a newly validated commercial tool nor a project that began on the date of the announcement.
Predicting weight does not describe the wine
Our editorial reading starts with a straightforward distinction: estimating production and describing quality answer different questions. A quantity forecast may help plan a season, but it cannot tell us how the wine will taste. Nor does it replace identifying the variety, vintage, location and winemaking decisions when explaining a particular bottle.
Imagine two forecasts with the same central volume, as a planning example rather than a result from these trials. One might have a narrow range and the other a very wide one. They could occupy the same box in a presentation while offering very different reassurance to someone organising a harvest. The uncertainty range and the date of the prediction matter alongside the central number.
Questions worth asking of a model
We suggest three questions for following this project. How far ahead of harvest is the estimate available? What actual crop is it subsequently compared with? What happens when the method is used in a season or plot that was not used to adjust it? These are reading criteria we propose, not shortcomings we attribute to the team.
It will also matter which errors are considered acceptable for each purpose. A forecast informing a regional assessment does not necessarily resolve a particular decision on an individual farm. An aggregate result can be informative without describing every component well. Explaining the scale prevents us from expecting a model to answer a different question from the one it was designed to address.
Water: an objective is not an outcome
If research aims to help manage resources, the next journalistic step is to establish which decisions improve because of it. A reasonable objective is not sufficient evidence. Discussing demonstrated savings would require a comparison, its conditions and its results. We assign no percentages here and do not turn an intention into an already measured benefit.
The same care applies to the word intelligence. A system may provide useful information while still needing agronomic interpretation. The interesting question is not how many technologies fit into the project description, but which specific uncertainty they reduce. That will be a useful way to follow the work as public results become available.
A story to follow
For wine, the interest lies in connecting vineyard observation with earlier decisions. Today we can report where the work is taking place, its timetable and its research objectives. Later we can examine its validation. Keeping those stages separate does not diminish the research: it helps us understand its contribution when the evidence arrives.
Sources and method. IRTA announcement, 18 November 2026, and project record, accessed 19 November 2026; links above. Both come from the research organisation, rather than two independent evaluations. Questions about validation are editorial analysis. We have neither conducted these trials nor tested a forecasting tool. By sommelier.vi.cat.
