Pest detection in the field: where manual scouting breaks down
A complete walkthrough of a plot of several hundred hectares is infeasible: the agronomist follows a route and inspects a sample of points — diagonally across the field, counting pests on plants within a fixed frame.
Such manual field scouting gives a snapshot for a specific day, but the sampling density is orders of magnitude lower than the area, and insects in the field are distributed unevenly.
The infestation hotspot is always a local patch: a lowland, field edge, area near a shelterbelt or last year's plowed land.
The probability that the route will pass exactly through it is low, and pheromone traps catch adult flight and almost do not show where the larva is already feeding.
Between two scouting passes, one to two weeks pass.
During this time, the pest manages to produce a generation, and the patch grows severalfold, while some species move deeper — to the underside of the leaf, inside the stem, to the root collar, where visual inspection by leaf no longer gives anything.
Hence late pest detection has its cost: the decision is made not by hotspot but by the entire field, treatment is applied across the area, and yield losses are calculated after damage has become irreversible and the economic threshold has been exceeded by a wide margin.
We examine how this is closed — how regular monitoring of the entire area is built, and why a decision on a hotspot can be made before losses become visible.
Measurable effect: scouting speed, field coverage, early warning
Previously, scouting relied on the agronomist's route: in a day, sample points across several hectares are covered, and the decision "treat or wait" is made based on them.
Pest detection changes three things: the share of surveyed area, the time from hotspot appearance to signal, and the set of failures that stop occurring.
Field coverage, scouting speed, and regularity
A grid flight with frame overlap covers the entire perimeter in one pass: the surveyed area grows from sample units of percent to full coverage, and not because the agronomist walks faster.
Then the images are processed in batches — the model annotates hotspots, coordinates are compiled into a hotspot map by grid cells, and it shows not only "where" but also how the patch grows from visit to visit.
Regularity is maintained by schedule: once a week during the active flight phase, less often outside it, without manual route replanning.
Early warning and response time
The signal arrives the same day: the time from hotspot appearance to notification is reduced from weeks — while the pest is seen by eye — to hours, while the patch occupies a few grid cells.
The model's confidence threshold is tuned to the crop: in some cases it is more profitable to catch a false positive and recheck, in others it is more important not to miss.
A whole class of failures is removed — inspector fatigue and subjectivity, missed edge zones, different readings of the same symptom by two specialists.
We keep recognition accuracy on validated observations from iNaturalist, supplementing the sample with our own field images — so the model does not "drift" on a new variety or background.
Pest detection methods: comparison of approaches
The monitoring method is chosen not by model accuracy, but by what signal is needed and at what frequency.
Pest detection from images works only when the sign is distinguishable in the frame — therefore the key question is not the algorithm, but the imaging platform and its frequency.
What we capture and how often
| Option | When it fits | Limitations |
|---|---|---|
| Camera traps | Continuous monitoring at fixed points: greenhouse, rows of one crop, orchard | Cover individual plants, require power and maintenance |
| Drone survey of the plot | Large blocks, review once every few days, general hotspot map | Depends on weather, the picture becomes outdated between flights |
| Selective inspection as a method | Alarm clarification: confirm pest species and hotspot boundaries | Slow, small area per pass, result depends on the observer |
| Manual pheromone traps | Assessment of pest flight dynamics at a control point | No image, hence no automatic classification |
Where each method falls short
No single option covers the entire block, and they fall short in different ways.
Continuous monitoring catches the moment an outbreak begins, but gives a point picture of several plants; a drone survey builds a distribution map, but misses rapid changes between flights; selective inspection is accurate by species, but does not scale to large areas.
Therefore, sources are combined: traps raise the alarm, drone survey clarifies boundaries, inspection confirms the species before treatment.
The detection loop remains the same — only the frame format and arrival frequency change, and this determines which methods can be integrated without rebuilding the system.
How we implement the system: stages of engineering work
The work proceeds in a fixed order: each stage delivers an artifact that is checked before moving to the next. Skipped field surveys result in a one-size-fits-all monitoring scheme — and the model confidently makes mistakes on a specific crop.
- Field survey. Visit to plots: crop and rotation map, treatment history, zones of past outbreaks. The output is a list of pests by development stage, not a wish "we want detection."
- Monitoring scheme. We choose image sources — equipment-mounted cameras, drone survey, stationary traps. We fix capture frequency, angle, and frame-to-coordinate georeferencing so the signal arrives accurate to the row.
- Image collection and annotation. We need frames under different light and at different growth stages. Disputed frames are annotated by two people, and discrepancies are reviewed: otherwise noise goes straight into training.
- Training and setting trigger thresholds. We select the threshold by the cost of error, not by a nice number: missing a hotspot is more expensive than an extra visit, so we tune the precision-recall balance to the agronomist's regulations.
- Validation on control plots. We test on fields not involved in training — different crop, different weather, different imaging equipment.
- Handover and training of agronomists. Scouting regulations, response to signals, review of false positives. Without this, even an accurate model remains unused.
How we check quality
Acceptance is based on control plots: share of missed hotspots, number of false positives per hectare, repeatability between surveys.
We separately look at degradation — how the system behaves under changing lighting or a new crop, so that failure is manageable rather than sudden.
Delivery package: models, schemes, regulations, and access
We consider the system delivered not when the model has shown the required accuracy on a test set, but when the farm operates it without us.
Therefore, the delivery package includes not only the trained model, but everything the agronomist, engineer, and plant protection service need for daily work.
Technical part
- Configured pest detection model, fine-tuned on images from specific fields: trigger thresholds and pest classes that actually occur on the farm are fixed.
- Layout plan for cameras and traps: field map georeferencing, installation height and angle, overlap zones — reinstallation on a new perimeter takes hours, not weeks.
- Panel access rights: agronomist, engineer, and manager roles, segmentation by plot, action log, and exchange keys with the farm's accounting system.
- Module for exporting observations and accumulating images for fine-tuning, with a backup procedure.
Documentation and regulations
- Signal processing regulation: to whom and within what time a notification goes, how a trigger is confirmed, and in what order escalation proceeds.
- Agronomist instructions: how to read the field summary, confirm or reject a trigger, and mark false positives.
- Model passport: what data it was trained on, where the applicability boundaries are, and what signs require reconfiguration.
- Handover act with acceptance log: what was checked and on which fields, who is responsible for operation.
The package is handed over in working form: access is issued to responsible employees, documentation is written with examples from your fields. Describe the task — we will return with an estimate and handover plan.
Case study: how yield losses were reduced on trial plots
Before the project, the farm scouted trial plots manually: the agronomist recorded hotspots in a log, and a full round took several days. On average, about 60% of hotspots were found, while the rest were identified only at the stage when the pest moved to neighboring rows. From first signal to treatment, five to seven days passed.
We assembled a pipeline: plot imaging, agronomist image annotation, model fine-tuning on every cycle, and a hotspot map with coordinates that goes to the crew.
The model outputs a probability, and the trigger threshold is tuned to the crop and growth stage — this removes false signals on withered leaves and shadows. Training was done on the open stack PyTorch; the weights and code remain with the customer.
On the same plots, the share of detected hotspots rose above 90%, response time dropped to one day, and the area of re-infestation dropped to 8% versus a quarter of the field previously.
Some fields remained a control plot with the old scouting scheme: the numbers there did not change, which made it possible to separate the model's effect from the weather season.
Applicability boundaries of the case study
The case study does not transfer to any farm without tuning. The result was obtained on one crop and one season: on other pests and different lighting, accuracy will have to be confirmed anew, and hotspots in shading and along the field edge are detected worse.
The model does not replace the agronomist — it shows where to look, and the treatment decision remains with the specialist.
What to do if part of the fields is already covered by cameras and sensors?
Whatever is already in the field can almost always be used: supports, power, communication channel, and the cameras themselves are reused without replacement.
The problem is that surveillance cameras were installed for other tasks — perimeter control, equipment tracking — and in terms of field of view, mounting height, and illumination, they are only partially suitable for pest detection.
An inventory gives two answers: what share of the area the camera covers with quality sufficient for recognition, and where blind spots remain.
Cameras and sensors cover different halves of the task. The camera records the fact and localizes the hotspot, while humidity and temperature sensors and pheromone traps provide context by which the model distinguishes seasonal background from an outbreak.
The combination works by matching time and coordinates: a single trigger without sensor confirmation goes into a verification queue rather than to an agronomist notification.
Partial field coverage is not a reason to postpone launch. The system works on plots where model confidence is above a set threshold; the rest is covered by scheduled route inspections.
The control boundary shifts gradually: as equipment is added, the area under automatic monitoring grows without rebuilding the logic or re-annotating.
We do a pilot launch on one plot: we take the already accumulated archive, annotate it, and fine-tune the model for a specific crop and pest.
Within a few weeks, accuracy and false positive rate are visible under real field conditions — after that, expansion to other plots proceeds with a clear risk rather than blindly.
Usually only a little needs to be added: point cameras in blind spots, a weather station, and correct georeferencing of images.
Technical concerns: accuracy, false positives, weather
Model accuracy is not a single number but a set of modes: at acceptance, we count separately correct triggers for each pest, misses, and false alarms that cost an agronomist visit. The reference point is phytosanitary survey standards EPPO.
Can small pests really be distinguished?
Thrips and whitefly in a frame from working distance are fractions of a percent of the area; detection by whole individual does not work. We shoot in macro mode with a fixed distance and decide by damaged area, while disputed frames go for additional annotation.
What happens when shooting in fog and at night?
In fog, the visible channel loses contrast — we work in tandem with infrared illumination and spectral indices. At night, some pests are less visible in IR; separate classes are set up for them. In dense fog, the system responds "uncertain" and suggests re-capture.
How do you deal with false positives?
The detector finds candidates, a second classifier confirms them on an enlarged fragment, and only then does a notification go out. We select the threshold according to the cost of error on the farm: in some cases a miss is more expensive, in others an extra visit. Thresholds are calibrated using agronomist confirmations.
What to do without stable connectivity?
Recognition happens directly on the device: the frame is processed on site, and metadata and disputed frames go to the server. Results accumulate in a local buffer and synchronize when the network appears — scouting does not depend on coverage.
Let's discuss the task: we will select a scheme for your crops and fields
To return with an estimate and work plan rather than general considerations, we need a minimum of source data on the farm.
A pilot launch on one plot usually takes two to four weeks — the timeline depends directly on how complete the field materials are and whether there are already captured and annotated frames.
What to send for assessment:
- Field map: plot boundaries, areas, crop rotation scheme, and breakdown by crop.
- List of crops and phenophases where it is important to catch the pest — from emergence to grain filling.
- What is already on the plot: stationary cameras, drones, weather station, equipment telemetry — and whether there is access to archives.
- Images and frames from the past season, including those where infestation is confirmed, and clean control plots.
- Treatment log: dates, plots, what was used — to link system signals with actual results.
- Who on site makes the decision and in what form it is convenient for them to receive the signal: map, plot list, notification.
- Communication constraints, flight regulations, and requirements for field data storage.
This is enough to select an imaging scheme, estimate the annotation volume, and give timelines for a pilot launch on your plot. If some data is missing — just say so: we will discuss what to replace it with; often the agronomist's ground observations over a season are sufficient.
Describe the task — crops, area, what is already working in the fields. We will return with an estimate, work plan, and boundaries of the pilot launch; after that, the decision is yours.
