Why AI in agriculture fails in the field
In a demo, crop image analysis looks convincing: the model is trained on large datasets from other people's fields, metrics are high, the map is bright.
On a specific farm, the model's accuracy in the field drops, and some misses surface as yield losses — disease spots and weed patches are found later, when treatment is more expensive and less effective.
The first cause is variability in imaging. Illumination during the day, cloud cover, terrain shadows, different cameras and flight altitude change brightness and spectral characteristics so that the model sees unfamiliar features. Without calibration and on-site data, classification turns into noise.
The second cause is lack of labeling.
Until there are agronomist reference marks of where exactly the problem is in this field and what it looks like in local conditions, there is nothing to fine-tune the model on and nothing to validate it against: third-party datasets do not reflect local crops, crop rotation, and equipment used.
The third is the gap between report and decision: the agronomist gets a layer on the map, but not work priorities and not a clear answer on what to do in a specific grid cell.
Together this produces a familiar scenario — a pilot without results. Reports are generated, crop condition monitoring does not become regular, and the team returns to field rounds.
Below we break down how this is closed: from preparation of imaging data and labeling to the format in which model output is embedded into agronomist decisions.
What changes in the field after implementing AI image analysis
The first measurable shift is the share of areas where the problem is recorded before visual signs.
Imaging in infrared channels shows vegetation stress several days before leaves begin to change color — the agronomist gets a window to treat in time rather than state the fact after the fact.
Next, the shift from a single average indicator for the field to a heterogeneity map works.
Vegetation indices are calculated for each image cell, and instead of "the field is generally normal," the agronomist sees zones with different biomass condition: where the crop is developing on schedule and where it is already lagging.
The heterogeneity map directly reduces the number of repeat treatments.
Instead of a continuous pass over the entire field boundary, machinery enters only problem zones — the treatment area decreases, and with it some repeat visits disappear, which were previously made "just in case" across the whole field.
A separate metric is the time from image to agronomist decision. Imaging, index calculation, prioritization of zones, and a field task card fit into hours: the decision comes with coordinates and rationale, not as "look, something is wrong there."
It makes sense to verify the effect on control zones and in comparison with the previous season on the same field — otherwise there is nothing to confirm the indicators with.
In practice, three values are captured: share of early detections, number of repeat treatments, time to decision.
Imaging modes: satellite, drone, manual field scouting
The three sources do not replace each other but cover different tasks. Satellite field imagery provides regular overview of large areas, a monitoring drone provides detail for a specific zone, manual field scouting provides on-site verification.
Imaging frequency, cloud cover, and interference determine whether a mode will deliver data while a field decision still makes sense.
What we compare with what
| Option | When it fits | Limitations |
|---|---|---|
| Satellite field imagery | Large, even field boundaries, overview of the entire area every few days | Cloud cover and interference: the window of usable images during the season is noticeably narrower than the calendar season; detail is in patches, not per plant |
| Monitoring drone | Complex terrain, small field boundaries, on-demand imaging for a specific date | Limited coverage per flight, requires an operator and flight permits, surface wind |
| Manual field scouting | Diseases and pests indistinguishable from above, anomaly verification | Low frequency, depends on people, point sampling |
| Combined mode | Satellite as background, drone and scouting — pointwise for detected anomalies | Requires coordinated routes and a unified storage for imaging data |
The mode is chosen for the crop, field boundary size, and terrain, not by the example of a neighboring farm. First we fix what exactly needs to be seen and how often, then select the source — otherwise imaging produces beautiful images without agronomic decisions.
Implementation stages: from field audit to model handover
Implementing image analysis is not installing a ready-made model but a sequence of work where each step relies on the previous one. Starting with training on third-party data is easy, but then the model will show good metrics in the report and make mistakes on your fields.
- Field and data audit: we consolidate image history over several seasons, field boundaries, crop rotation, and agronomist records. We look at what is physically distinguishable in the images and for which tasks there is not enough data.
- Solution design: we define the result for the agronomist — a task map, a report with coordinates of hotspots, or variable application zones. For it, we choose imaging channels and update frequency.
- Model training on your own data: we label images of your fields together with agronomists so the model learns local soils, varieties, and machinery. We fix which defects are considered target.
- Tuning: we calibrate sensitivity thresholds, fine-tune on disputed areas, and remove false positives on shadows, field edges, and machinery tracks.
- Control plots: we set aside holdout fields that participated neither in labeling nor in training, and check quality on them under the conditions of a new season.
- Result acceptance: we calculate metrics on control plots, compare with agronomist scouting, and agree on which error is critical. The decision is made based on numbers, not a demo.
- Solution handover: we deploy image processing on your side, provide documentation on data and settings, and train agronomists. Fine-tuning on new seasons is then done by your team.
Labeling, schemes, and regulations: handover package
The value of the project in the field is determined not by the model itself but by your team's ability to support it without us.
Therefore we hand over not a "black box" but a package where each artifact has a purpose and an owner — from the dataset to the regulation for agronomist actions.
- Labeled dataset — images from your fields, class labels (diseases, weeds, growth stages), rules for resolving disputed frames. It allows the model to be fine-tuned on your own.
- Image processing scheme — the path of a frame from imaging to the problem zone map: where the source files are stored, which checks filter out blurry, overexposed, and cloudy images.
- Model update regulation — at what volume of new data retraining is launched, who makes the decision, by which indicators the new version is compared with the working one, and when it is rolled back.
- Instructions for agronomists — how to image a field, at what time and in what weather, how to read the report, and what to do with the model signal on a specific plot.
- Data access rights — roles of agronomist, analyst, and contractor: who sees what, who exports source files, and how access is revoked when an employee leaves.
- Log of settings and sensitivity thresholds — which values were set for your crops and why they changed, so that a season later the review does not start from scratch.
- Knowledge transfer — working sessions with your team: review of typical model errors, independent edits to labeling and regulations.
Case study: early disease detection on a trial field
The trial field — four plots of 40-60 ha with different crops in rotation. Before the project, the problem was looked for by eye: the agronomist walked the field once every 7-10 days, and a spot became noticeable when 15-20% of the area was already affected.
The entire field had to be treated — over the season there were four continuous passes, and part of the product went into the soil and onto healthy areas.
What was changed
Imaging was set up once every three days: a UAV with a multispectral camera plus open Sentinel-2 images as a background for slow changes.
The field was divided into cells, for each a vegetation index was calculated and compared with the baseline of the same plot from previous seasons — this separated disease from uneven irrigation and shading.
Cells with deviation above the threshold went to the agronomist in a checklist, not into an automatic treatment command.
What the outcome was
The share of areas where the problem was found before visual signs grew from 10% to 70% — a time gain of 5-7 days before spots appear.
Continuous treatments decreased from four to one per season, the rest — pointwise by hotspots, and the area under treatment decreased almost threefold.
The false positive rate remains within 15%; we recheck it monthly with ground inspection: thresholds are tuned to the specific field, not taken from a general template.
What to do if fields are already imaged but the data is not used
Images and flights without decisions based on them are a typical situation, and the cause is almost always the same: between the flight and the agronomist's action there is no link that turns a frame into a clear signal.
The data sits in someone else's dashboard or on a disk, with no link to field boundaries and history from previous seasons. We break down step by step what to check in the archive before building something new.
We start the audit of existing data with an inventory: which fields and at which growth stages were imaged, at what frequency, and by the same operator or not.
Then we look at coordinate georeferencing, frame overlap, cloud cover and shadows, and comparability of imaging time for neighboring fields.
If routes changed from flight to flight, the time series are incomparable, and part of the archive will have to be excluded — this is a normal audit result, not a loss.
Imaging quality is more important than archive volume. Processing works reliably when conditions are comparable: stable altitude, same overlap, even lighting, absence of gaps at field edges.
Images from open sources, such as Sentinel-2, should be kept as a separate layer: they have a different frequency, and they cannot be mixed with your own flights without correction.
Problem definition is a key step. The wording "we want AI" does not work: we need to decide which question we are answering — a within-field heterogeneity map, early stress warning, basis for variable fertilization, harvest control.
Then this turns into a measurable result that the agronomist can verify in the field, and only for it is the processing method selected.
We restart the pilot with a narrow scope: two or three fields, one crop, one season, and an agreed metric. Ground checks are mandatory — without them there is nothing to confirm that the map matched crop condition.
On such a scope it is quickly visible where data changes the decision: how many fields fall into the attention zone and how many treatments are assigned based on maps rather than scouting.
Technical concerns: cloud cover, labeling, accuracy
Imaging gaps are not a failure but a normal operating mode: cloud cover removes some dates, so data is reconstructed from the time series, and critical vegetation windows are duplicated with radar images.
We consider the solution working if the field feature is stable also on dates without optical imagery.
At the start, labeling does not require thousands of polygons: we take a stratified sample by zones, growth stages, and crops, and send disputed areas for fine-tuning — labeling volume grows where the model is genuinely uncertain.
How many images are needed to get through a cloudy season?
We look at date density, not their total number: if key development phases fall on two or three dates and there is a radar channel, the model is robust. Sentinel-2 provides regular imaging — for most field tasks this is enough.
How is accuracy validated on new fields and seasons?
Acceptance is performed on holdout fields and in the next season, not on data the model has already seen.
Recall and precision are calculated separately by crop, we monitor the false positive rate, and set a confidence threshold: below it, a plot goes to the agronomist for manual review.
What to do when the variety or sowing scheme changes?
The model is fine-tuned on new data, and the previous version remains available for rollback — if quality on control fields drops, we return to the previous one. This way a season change does not turn into service downtime.
Let's discuss your fields: where to start and what will be needed
To discuss the task for your farm concretely rather than in general terms, the project assessment should rely on your source data. Nothing needs to be collected specially — what has already been accumulated in the fields is enough. The minimum set that greatly speeds up the review:
- Field boundaries with coordinate georeferencing — plot borders in any form: an export from a precision farming system, a map, or marks directly on an image.
- What is already being imaged: sensor type (regular camera or multispectral), flight frequency, who performs the imaging — your specialists or a contractor.
- Ground reference surveys: where the agronomist recorded weeds, lodging, stress hotspots — even notes and phone photos are suitable.
- Field history: crop rotation, emergence and harvest dates, treatments — this is needed to distinguish a real signal from crop-specific features.
- The task in decisions: what exactly you want to receive — a map for variable treatment, early stress notification, or loss assessment.
- Constraints: field connectivity, depth of the image archive, where data is allowed to be stored.
Some data may be missing — this does not block the conversation; we will collect part of it during the work.
In response, we come back with a project assessment and work plan: which fields we start with, over what period we accumulate observations, how we validate accuracy against your reference surveys, and what we hand over to the farm.
We fix the project boundaries at the start: which plots and decisions are included in the first stage, and what we deliberately leave out of scope so as not to blur the result.
Describe the task in the scope it has now — we will come back with an assessment and plan.
