Hive monitoring technology: what sensors and cameras can tell you

Measurement, inference and diagnosis

Every hive technology does the same three things, and most confusion comes from running them together. A sensor MEASURES a physical quantity: a load cell measures mass, a probe measures temperature, a camera records pixels. Someone then INFERS a colony state from the measurement: a nectar flow from a gain in mass, brood from a steady temperature, a mite from a cluster of pixels. A DIAGNOSIS — this colony has this disease, this colony has lost its queen — is a further step again, and in beekeeping it is made by looking at the bees.

The field is substantial. A 2025 systematic review found 135 peer-reviewed studies from 1990 to April 2025, most using environmental sensing, sound or vibration, imaging or weight. It also found that 55% used a single sensor type, that true combination of sensors was largely absent, that many systems were still prototypes, and that long-term, multi-season field testing was the main gap.

What each technology measures, and what that can support

The table keeps the measured quantity and the inferred state in separate columns, and grades each interpretation separately. 'Cannot show by itself' is the column to read first: it is the honest limit of each instrument.

How mature is each method?

HoneyHQ grades monitoring methods on a five-rung ladder: prototype, controlled research, field research, validated decision support and established practice. A method moves up only on evidence from real colonies, beyond one dataset, for a stated decision. It does not move because one paper reports high accuracy, because it uses machine learning, or because it is on sale; a commercial claim without independent testing sits beside the ladder, not on it. Nor is the ladder a pipeline: many methods will stay where they are.

Hive scales

A hive scale is load cells under the hive, read every few minutes. It is the most mature instrument in beekeeping because the measurement is unambiguous: it records the mass of everything on the platform. A scale on one reference colony tells a beekeeper when a flow started and stopped, when supers are needed, and how fast winter stores are going — cheaply and without opening the hive.

The daily curve carries more. Foragers leaving in the morning make the hive lighter; nectar, pollen and water coming back make it heavier. Researchers who delayed forager departure by one to three hours with screen gates confirmed that the size and timing of the morning drop reflect flight, and showed that within-day change can reveal colony-level effects that periodic inspections miss — where the effect is large compared with the scale's precision.

• Calibrate with a known load, and again after the hive is moved.
• Load cells drift and respond to temperature; a slow trend over weeks can be the scale rather than the colony.
• Tare after every box added or taken, and write down when.
• Read the curve against the forage calendar and the weather: a gain with nothing in flower is a reason to look, not a flow.

Forecasting weight is a separate task from explaining it. A 2026 study forecast weight on a hive held out from training with an average error of 0.94 kg — on two hives from one German dataset — and its authors call for testing across sites before relying on it.

Temperature, humidity and carbon dioxide

Worker bees hold the brood nest at about 32–36 °C, and a probe beside brood shows that steadiness while the air outside swings. That makes temperature useful for broad states — brood rearing under way, the winter cluster moving — but it reads the temperature where the probe is, not the colony. A probe a frame away from the brood reads lower whether or not there is brood. A steady 35 °C is consistent with brood nearby; it is not proof of brood, and a cooler reading is not proof of none.

Temperature can also catch a swarm leaving. In ten Latvian colonies, a probe above the nest showed a warm-up of about 1.5–3.4 °C in the 8–20 minutes before take-off, and a simple rule detected nine swarms at that stage. That is detection at the event — useful for finding the swarm — not a forecast days ahead.

Relative humidity depends on temperature as well as on water, so it can change with no change in the air's moisture. Brood-nest humidity of roughly 50–60% has been recorded in African honey bee colonies, with a gradient from the centre outwards, but whether bees regulate it actively is still discussed. No source read supports a universal humidity threshold for hives, and HoneyHQ gives none.

Carbon dioxide builds up as the colony breathes; workers respond by fanning, which limits it. A CO2 sensor follows that balance. Air composition appeared in only 10 of the 135 studies in the 2025 review, which links the low uptake to cost and limited interpretability. Electronic noses — arrays of gas sensors whose pattern is classified by a model — have separated Varroa-infested from uninfested colonies in field trials with a misclassification error of about 17% in the best case, from one research group. That is research, not a disease detector.

Sound and vibration

A microphone records sound in air; an accelerometer embedded in the comb records vibration through the wax. They are different signals. Both carry real information — reviews of hive sound regard it as one of the most promising approaches — and both carry everything else too: rain, wind, traffic, the beekeeper's hands, the box itself.

Acoustic classifiers recognise a queenless colony in the hives they were trained on, and the researchers report that they generalise poorly to hives they have not heard. A 2026 study of 53 colonies matched recordings to field Varroa checks and found the 122–396 Hz band most informative, without yet reporting how accurately infestation could be classified. Estimating colony strength from sound is at the preprint stage.

The strongest swarm-warning evidence comes from comb vibration. Accelerometers in 25 hives in France and England gave warnings averaging 22 days (instantaneous spectra) or about 10 days (ten-day change) before swarming. The same study records 3 of 18 swarms missed by one method and 5 of 18 by the other, with 0 and 2 false alarms in 14 colonies that did not swarm; the first method worked only within the swarming season, daytime foraging made the spectra unusable, and there was no held-out test set. Promising field research from one programme — not a swarm forecast for your apiary.

At the entrance: counters, cameras and tags

Optical counters send bees through narrow tunnels with light beams; the order in which beams are broken gives the direction. Cameras do the same job with a model tracking bees across frames, and can also pick out pollen loads and, in some rigs, mites. Both measure departures and arrivals. Neither has a standard method for validating its counts, which makes counters hard to compare.

RFID tags glued to individual bees record each tagged bee's departures, returns and trip durations. They have shown, for example, shorter-term changes in foraging after a single dose of a neonicotinoid. They measure the tagged bees only, tags fall off or go unread, and the tag itself is a load: RFID is a research method for individual behaviour, not whole-colony monitoring. Radio tags have also been used to follow Asian hornet workers back to nests up to 1.33 km away — a research and surveillance technique for trained trackers.

Varroa: what automated detection can and cannot show

Varroa is where the technology is moving fastest, and where the gap between a headline number and a beekeeping decision is widest. Four different things are called 'Varroa detection', and they are not interchangeable:

1. A mite visible on an adult bee — what an entrance camera can see.
2. Mites fallen onto a board under the colony — natural mite fall, which shows a trend rather than a level.
3. Mites in brood cells — where most of the colony's mites are, and where a camera at the entrance cannot see.
4. The colony's infestation — mites per hundred adult bees, which is what treatment thresholds refer to, measured by an alcohol or detergent wash or a sugar roll.

Recent systems are genuinely good at the first. A 2026 study fitted hives in Taiwan with an entrance imaging chamber and reported 97.4% mean average precision for finding mites in its test images. A 2025 method classifying short video clips of foragers, trained on a public set of 607 videos from one hive, reported 0.986 accuracy on its clips. An earlier rig that films bees shaken from a frame estimated 5.80% infestation of the filmed bees against 5.52% counted by hand in the same video.

The established methods are not replaced by any of this. The table below is HoneyHQ's record of the manual methods, with what each can and cannot show; an automated count is best read as a reason to do one of them sooner.

Cameras on the comb, and thermal imaging

Photographing frames with the bees removed and classifying every cell is well developed: one open system finds cells and sorts them into seven content classes — eggs, larvae, capped brood, nectar, honey, pollen and empty — with an F1 score of 94.3% for its best network. That gives a measured inventory of brood and stores. It still means lifting the frame, it is sensitive to lighting, angle, wax colour and bees left on the comb, and classifying a cell as 'capped brood' is not recognising foulbrood or a virus. A camera fixed in a closed hive does not perform an inspection.

A thermal camera images infrared radiation from the surface facing it — on a closed hive, the outside of the box. It does not see through wood. Calibrated images of hive walls have been related to approximate population, and thermography of exposed winter clusters shows where heat is made. A warm patch on a wall shows heat escaping there; it is not a picture of the comb and not a diagnosis.

DNA in the hive

Swabs of comb, hive debris, honey, bee bread or bees can be tested in a laboratory for the DNA of a pest or pathogen. For an exotic mite this can be the most sensitive tool available: in a 2025 field trial in Papua New Guinea, comb swabs detected Tropilaelaps in every infested colony, more often than any other method compared. Honey DNA has been used to survey nine pathogens and parasites across 102 samples from 18 countries, and sequencing honey lists the plants, microbes and pests whose DNA reached it.

That is why DNA tools belong in surveillance — is it here, and where should inspectors look? — rather than in deciding what a single colony needs. The biosecurity guides for Tropilaelaps and the yellow-legged hornet set out the official routes.

Combining sensors

Independent signals that agree make one explanation more likely and others less so. A sudden 2 kg loss at midday is ambiguous; the same loss with a warm-up in the minutes before, a surge of departures and warm still weather points to a swarm rather than a harvest. Combining sensors is how a monitoring system could reduce false alarms. It is also, on the 2025 review's count, largely absent from the published work so far.

Fusion is not certainty. A combined model still depends on how its training events were confirmed, on the colonies, hives and climate it learned from, on each sensor's calibration, and on whether it transfers to another apiary. Its honest output is a confidence that an event happened, which leads to a check.

Comparing the technologies

There is no best hive sensor: which is useful depends on the question. The comparison below sets out what each records, whether it runs continuously, where it sits, what it costs in power and data, and the confounders that matter most. Nothing is scored.

Edge computing, connectivity and power

A hive monitor has to sample, store and send its data, usually far from mains power. Processing at the hive — on a single-board computer, a microcontroller or a small machine-learning accelerator — means sending a result ('42 departures this hour', 'no mites seen') instead of hours of video or audio. That saves bandwidth and power, keeps working when the connection drops, and keeps raw video of a site private. The cost is a model that is harder to update and check.

Power follows from what is sampled. A temperature probe read every ten minutes runs for a long time on a small battery; a camera or a continuous microphone needs far more, and the 2025 review notes that solar-powered hives often fail to harvest enough energy for continuous monitoring. Sampling less often, sleeping between readings and processing at the hive are the usual trade-offs.

Data quality and the management log

Most wrong conclusions from hive data come from the data, not the model. The commonest problems:

• Calibration and drift: a sensor that is slowly wrong produces a trend that is not there.
• Placement: the same probe reads differently a frame away; the 2025 review notes that temperature varies with probe position.
• Missing data and lost packets: a gap is missing, not zero. A counter that stopped sending did not count no bees.
• Clock errors: a sensor whose clock has drifted puts events at the wrong time, which breaks any comparison between sensors.
• Outliers: a knocked hive, a bird on the roof, a probe pulled during an inspection.
• Maintenance: propolised sensors, blocked tunnels, flat batteries.

The single most useful addition to any monitored hive is a written log of what the beekeeper did. Opening the hive, adding or removing a box, feeding, harvesting and moving the colony all change the signals dramatically — and a system without the log will read them as colony events.

Weather, landscape data and digital twins

Weather explains many changes that would otherwise look like colony events: bees do not fly in cold rain, a wet roof is heavier, and a sunny afternoon after rain brings a rush at the entrance. A weather record at or near the apiary should sit beside every other signal. HoneyHQ's guide to climate, weather and honey covers what weather does to forage and flows.

Satellite vegetation data describe the land: greenness, land cover, drought. They are a proxy for forage, not a record of it. Across four European countries, the satellite start of the green season did not match the start of the nectar flow recorded by hive scales, though it was related to the flow's midpoint. A landscape proxy cannot say what the bees actually collected, and no satellite can identify a honey's botanical origin; pollen and DNA analysis of the honey can.

A digital twin is a dynamic digital representation of a real system, kept linked to it by data, that can be used to explore what happens next. A dashboard of readings is not one, nor is a static model or a stand-alone simulation. A review of agricultural digital twins from 2017 to 2020 found 28 use cases; whether any operational digital twin of a bee colony exists, outside research concepts, is not established in the sources read.

Automation, robotics and the treatment boundary

Some automation is routine: scales that weigh continuously, loggers that record unattended, alerts sent to a phone. Robotics is experimental. In one 2024 study, cooperating camera robots in an observation hive at a university followed the queen across two combs for weeks, imaging workers, brood and stores while she rested and producing about 1.4 TB of images a day — a research instrument in a research facility, not a field beehive.

Privacy, data ownership and disturbance

A connected hive is a connected device. Change default passwords, keep firmware updated, and know what happens to the data if the service closes. Exact apiary locations are sensitive — to theft, vandalism and disease-control planning alike — and there is rarely a reason to publish precise hive coordinates. HoneyHQ's own tools ask for none.

• Can the raw readings be exported, in an open format, with timestamps and units?
• Is calibration metadata — when, against what — kept with the data?
• Is there an interface for other software, and is it documented?
• If the account or the company goes, does the historical record go with it?

Non-invasive does not mean zero disturbance. Monitoring that replaces some inspections spares the colony being opened; but an entrance tunnel or camera chamber narrows the entrance, lights and enclosures change it, probes sit in the brood nest, tags load individual bees, and every false alarm that sends the beekeeper into the hive is a disturbance the monitoring caused.