Imagine it is midsummer and your potato field looks fine. Is it going to be a good year? For centuries the honest answer was: you won’t know until you dig. Potato yield prediction now offers a much earlier answer. A European research team found that data from the first 60 days after planting explained about 77% of the differences in yield between plots, almost as much as data from the whole season.
The team grew 44 varieties at five sites in four countries over three years. Below, we cover what the researchers measured, how far the forecast can be trusted, and what a grower can do with it.
The short version
- Data from the first 60 days after planting predicted final yield almost as well as data from the whole season. Using only 60 days of data, the model accounted for about 77% of the differences in yield between plots, against 78% using the full season (statisticians call this share R²). It was tested on four varieties it had never seen before.
- A simple five-ingredient equation, built from two drone measurements and three weather records, explained 72% of the variation in yield.
- Forecasts travel well across varieties but not across places. For a completely new location the average miss was 36% of actual yield, so local calibration is essential.
- Only yield could be forecast. Tuber number, overall quality and starch content could not.
- Drone flights after day 60 added nothing meaningful to the yield forecast.
Why early potato yield prediction matters
Potatoes are the world’s third most important food crop and feed over a billion people. Yet a grower usually learns how good a crop is only when the tubers come out of the ground, when every decision about water, fertiliser and crop protection has already been made. A reliable forecast at day 60 would still leave time to act. It could also help breeders drop weak lines months earlier and help buyers plan supply.
What the researchers did
Between 2021 and 2023 the team grew 44 tetraploid potato varieties at five sites in four countries: Zeeland (Netherlands), Fuchsenbigl and Großnondorf (Austria), Valencia (Spain) and Vojvodina (Serbia). Climates ranged from cool and rainy to hot and dry, and roughly half of the trials were irrigated. In total there were 16 field trials and 1,165 plots.
For each plot they collected:
- Drone images at about 40, 60, 90 and 120 days after planting, turned into vegetation cover (how much ground the leaves shade) and greenness scores such as NDVI (an index calculated from the red and near-infrared light the leaves reflect, where a higher value means a denser, healthier canopy).
- Weather and soil data: temperature, rain, humidity, sunshine, day length and soil moisture, summarised for four time windows.
- Maturity ratings for each variety: an expert score from 10 to 90 for how early the plants ripen and begin to die back (high means early maturing, low means late).
- Leaf gene activity, for 16 varieties: leaf samples taken once, when tubers begin to form, to measure how strongly 50 genes were switched on. Most were chosen for their roles in plant development and stress responses, especially tuber formation.
- Harvest results: yield per hectare, tubers per plant, an expert overall-impression score and underwater weight (a stand-in for starch).
To avoid over-representing Zeeland, and because the Austrian sites had only one drone flight a year, the modelling used a balanced set of 606 plots from seven trials.
Yields varied enormously. The median was 45.1 tonnes per hectare (t/ha), from 19.8 t/ha in a non-irrigated Austrian trial in the hot 2022 season to 89.5 t/ha in irrigated Valencia in 2021. Irrigation generally raised yield and tuber number, and at sites with dry conditions it also kept soil cooler and moister. In rainy Zeeland it made little difference.
Finding 1: 60 days of data are almost as good as a full season
To test the forecast fairly, the researchers kept four varieties (Farida, Musica, Valdivia and Charlotte) out of the model’s training entirely, then asked it to predict their yields. The best-performing method was a machine-learning model called kernel ridge regression. In simple terms, it learns from past plots and then forecasts new ones.
Accuracy is shown as a score between 0 and 1, called R². A score of 1 would be a perfect forecast, and 0 would be no better than guessing the average yield every time. So 0.77 means the forecast captured about 77% of the differences in yield between plots.
| Data used up to | R² for yield |
| 40 days | about 0.71 |
| 60 days | 0.77 |
| 90 days | about 0.80 |
| 120 days (full season) | 0.78 |
The 60-day result is only about 1.3% below the full-season one. The first two months cover growth up to and including tuber initiation, the point when tubers begin to form. Drone and weather data together gave the best results, while adding gene activity did not improve them.
Finding 2: A five-measurement equation for potato yield prediction
Machine-learning models are hard to look inside. So the team tested every possible combination of five measurements available by day 60 and fitted a plain linear equation to the best set. It explained 72% of yield variation on the held-out varieties.
| Measurement | Window | Source | Effect on yield | Change in forecast (t/ha per step) |
| Total rainfall | Days 0–40 | Weather record | Lower | −17.3 |
| Days with humidity below 60% | Days 40–60 | Weather record | Higher | +15.2 |
| Highest humidity reached | Days 0–40 | Weather record | Higher | +10.9 |
| Vegetation cover | Day 60 | Drone | Higher | +6.9 |
| Greenness (NDVI) | Day 40 | Drone | Higher | +5.5 |
The forecast starts from a baseline of about 52.7 t/ha, roughly what the equation gives for a field where every measurement is average for the study. Each measurement then moves the forecast up or down, by the amount in the last column, according to how far it is from average. We call a clearly noticeable difference from average a “step” (statisticians call it one standard deviation). For example, if the first 40 days are one step wetter than average, the forecast falls by about 17 t/ha. If there are one step more dry-air days between days 40 and 60, it rises by about 15 t/ha. Add up the movements from all five measurements and you get the forecast. Early rain had the biggest effect, and a wetter-than-usual start went with lower yield. A vigorous, well-closed canopy went with higher yield.
How to use it. Treat the equation as a guide, not a calculator. You can’t type in raw figures such as millimetres of rain, because it was built with each measurement expressed as a difference from the study average. What it does show is which measurements matter most and in which direction. To get a forecast you can rely on for your own farm, the baseline and the numbers would need to be recalculated from your own weather and yield records, which is a job for an adviser or researcher.
Associations, not instructions. The equation shows what tends to go together with higher or lower yield, not what causes it. For example, more dry-air days between days 40 and 60 went with higher yields, but that doesn’t mean drying out a field would raise yield. Those days may simply reflect the kind of site or season the crop grew in. The same caution applies to the rain effect: waterlogged soil starving roots of oxygen is a plausible reason, but the study did not test it. Use the equation to forecast, not to decide what to change in the field.
If you can’t measure something. Several of the inputs measure much the same thing. NDVI, WDVI and the chlorophyll index are three ways of scoring how green and dense the crop canopy is, so if your drone or sensor gives you one of them, another can often be used in its place. Likewise, weather data from a nearby station or an online weather service can replace a sensor in your own field. The authors expect accuracy to stay similar, but this was not tested directly, and the equation would need to be refitted with the substitute.
How far can you trust potato yield prediction?
This is the most important section for anyone planning to use the method.
- New varieties: good. When one variety at a time was held out, the average absolute percentage error (the typical gap between predicted and actual yield, as a share of actual yield) was 16.7%.
- New locations: much weaker. When a whole field trial was held out, that error rose to 36.0%. The authors conclude the dataset did not cover enough locations to generalise to new places. A new region needs local calibration.
- Only yield. Tubers per plant (R² 0.17), overall impression (0.16) and underwater weight (0.11) were poorly predicted because they are mostly set by variety genetics.
- Small test. The headline R² values come from just four held-out varieties (56 plots).
- Research plots, not commercial fields. Plots held 8 to 50 plants. Results from whole fields may differ.
Choosing varieties: yield, stability and starch
The team compared varieties on two things: how much they yielded in a typical trial, and how steady that yield stayed from site to site. Typical yields ranged from 27.3 t/ha (Charlotte) to 68.5 t/ha (Arizona). Varieties whose yield varied by less than about 35% between sites were treated as resilient. Combining all four harvest measures (yield, tubers per plant, overall impression and starch), Orchestra ranked best overall and Flamenco worst, but no variety came top on every measure.
Two takeaways. First, a high yield alone doesn’t make a variety a safe choice, so check how consistent it is across different conditions too. Second, high yield and high starch tended not to go together: varieties with higher yields tended to have lower starch content. Starch is measured here as underwater weight, which was 462-501 g in starch varieties and 303-405 g in table varieties. It stayed steady across sites because it is largely set by a variety’s genetics.
Genes: useful clues, not needed for the forecast
The team compared leaf gene activity in the three best-ranked varieties with gene data (Musica, Sound, Acoustic) and the three lowest-ranked (Agria, Altus, Bartina). Two genes known to hold back tuber formation (StSP5G and StGI1) were less active in the best group, which is the direction you would expect for better-performing plants. Two others were also less active (StLTP3, linked to waterlogging, and StCPK17, a calcium-signalling gene), and a gene linked to drought response (StZNF) was more active. Little is known about these last three, so the authors treat all five as possible targets for breeders rather than proven markers.
This is a small comparison, with gene activity sampled only once per plot. Adding gene data did not improve the yield forecast, and it requires lab analysis of leaf samples, so growers don’t need it.
Putting potato yield prediction to work
| When | What to do |
| From planting | Keep a rain and humidity record from a nearby station or a weather data service. |
| About day 40 | First drone or sensor pass for greenness (NDVI). |
| About day 60 | Second pass for vegetation cover, then run or refresh your forecast. |
| After day 60 | For yield forecasting, later flights can be skipped. |
Growers and advisers. A day-60 forecast leaves time to review irrigation, fertilisation and pest control, the decisions the authors say it can inform. Because the forecast was much less accurate in locations the model hadn’t seen (36% average error), check it against your own past yields before acting on it, and recalibrate it with local records. What you need to measure: rainfall and humidity from a weather station or weather service, plus greenness (NDVI, a score from 0 to 1) at about day 40 and vegetation cover (the percentage of ground shaded by leaves) at about day 60, taken with a drone or a handheld sensor.
Breeders. Early predictions could provide phenotype data well before harvest, which the authors suggest could feed genomic prediction models, improve selection accuracy and shorten breeding cycles.
Anyone building a local tool. Start from your own trial or field data. The authors have published their code at github.com/NIB-SI/Adapt_field_ML.
The bottom line
In this trial, the first 60 days held almost all the information needed to forecast potato yield. Early rainfall, humidity and the strength of the crop canopy were the key signals, and all of them can be recorded with two drone passes and a weather log. The forecast worked well for varieties it had not seen, at sites like those it was trained on. In a completely new region it was much less accurate (36% average error), so it needs local calibration, and it did not predict quality traits such as starch content or tuber number well. They make early forecasting a promising planning tool, and testing across more locations would show how far it can be trusted.
Source
Study: Potato yield can be predicted using drone-captured and environmental measurements early in the growing season
Authors: Angelika Vižintin, Maja Zagorščak, Eva Turk, Maja Križnik, Marko Petek, Katja Stare, Bernhard Wurzinger, Maroof Ahmed Shaikh, Guus Heselmans, Josef Söllinger, Anže Županič, Pieter-Jelte Lindenbergh, Jeroen Bakker, Robert Graveland, Ivana Imerovski, Salome Prat, Christian Bachem, Markus Teige, Bianca Doevendans, Alexandra Ribarits, Jan Zrimec, Kristina Gruden (2026)
Read the full paper: https://www.biorxiv.org/content/10.64898/2026.03.09.709817v2
Leave a Reply