HoneyHQ

How to read a claim about hive sensors and AI

A monitor that 'predicts swarms with 91% accuracy' or 'detects Varroa with 97% precision' may be describing good research — and still not tell you how it will do on your hives. Six questions, three numbers and one diagram that separate the two.

BeesReviewed 2026-11-01

The words, without the marketing

'AI-powered' says nothing about what a system does or how well. The useful words describe the task.

Terms used about hive-monitoring software, and what each means
TermWhat it meansA hive example
Artificial intelligenceThe broad field of making computers do tasks that seem to need judgement. Too broad to tell you anything about a product.—
Machine learningMethods that learn a pattern from labelled examples rather than following rules written by hand.A classifier trained on recordings labelled 'queenright' and 'queenless'
Deep learningMachine learning with many-layered neural networks, usually needing large labelled datasets.A network trained on thousands of images of bees with and without mites
Computer visionGetting information from images or video.Counting bees at the entrance from video
ClassificationPutting a whole input into a category.This clip shows a bee with a mite: yes or no
Object detectionFinding and boxing each object in an image.A box round every mite visible in a frame
SegmentationLabelling every pixel or region of an image.Outlining each comb cell and its contents
TrackingFollowing the same object across video frames.One bee followed from landing board to entrance
Time-series forecastingPredicting the next values of a measurement from its past.Tomorrow's hive weight
Anomaly detectionFlagging readings unlike a hive's usual pattern, without saying why.An unusual overnight temperature

Six questions to ask of any claim

  1. Predict or detect WHAT, exactly? A mite visible on a bee, or the colony's infestation? A swarm leaving, or preparation weeks before?
  2. How far ahead? Recognising a state in recorded data, catching an event as it happens and forecasting it days ahead are three different achievements.
  3. Checked against WHAT? A model is only as good as its ground truth: an alcohol wash, queen cells seen on inspection, a swarm observed — or labels inferred from the same sensor?
  4. How many false alarms, and how many missed events? Accuracy alone hides both.
  5. How many hives, apiaries, regions and seasons? One apiary for one summer is a start, not a validation.
  6. Was it tested on hives it had never seen? If not, the score describes the hives it learned from.

What a reported accuracy means

'97% accurate' is meaningless until three things are attached: 97% at what task, on which dataset, by which measure. The common measures answer different questions.

Performance measures and the question each answers
MeasureThe question it answersWatch for
AccuracyWhat share of all answers were right?Looks excellent when the event is rare, even for a useless model
Recall (sensitivity)Of the real events, how many were found?High recall with many false alarms
PrecisionOf the alarms raised, how many were real?High precision that comes from rarely raising an alarm at all
SpecificityOf the quiet periods, how many were correctly left quiet?Can be high while precision is poor, if quiet periods vastly outnumber events
F1 scoreA single balance of precision and recallHides which of the two is weak
Mean average precision (mAP)How well boxes drawn by a detector match the real objects in imagesSays nothing about what the detections mean for the colony
Mean absolute errorHow far a forecast is from what happened, on average, in the measurement's unitsAn average can hide occasional large misses

Rare events make accuracy misleading. If a swarm, a queen loss or a disease occurs in 1% of monitoring windows, a model that always says 'nothing happening' is 99% accurate and finds nothing. For rare events, ask for recall and precision — or the counts of events found, missed and falsely raised — and prefer precision-recall curves to ROC curves, which can look reassuring on imbalanced data.

1,000 windows, each square 10 windows

Square marked E: the 10 windows with the event.

A model that always says “nothing”
Accuracy 990 / 1,000 = 99%. Recall 0 / 10 = 0%. Useless.
A model that finds 8 events and raises 40 false alarms
Accuracy (8 + 950) / 1,000 = 95.8% — lower — yet recall 8 / 10 = 80%, and precision 8 / 48 = 17%: most alarms are false, and each false alarm costs an inspection.
Arithmetic, not data. In 1,000 monitoring windows of which 10 contain the event — a swarm, a queen loss — a model that always answers “nothing happening” is 99% accurate and finds none of them. For rare events, ask how many real events were found (recall) and how many alarms were real (precision), not the accuracy alone. Sources: The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. Drawing: HoneyHQ (HEKNO Ltd), original.

Data leakage: when the test is too easy

A model is trained on some data and scored on data it has not seen. But if images or windows from the same hive, day and lighting fall on both sides of that split, the model can score well by recognising the hive rather than the thing it is meant to detect. This is data leakage, and a survey across 17 scientific fields found it in hundreds of machine-learning studies, inflating their results.

  1. Weakest

    Random images or windows

    Frames from the same hives, days and lighting fall on both sides. The model can recognise the hive rather than the mite.

    What the score tells you: How well it fits these hives. Usually the most flattering figure.

  2. Stronger

    Whole hives held out

    Some hives are never seen in training and are used only to test.

    What the score tells you: How it does on new hives in the same apiary and set-up.

  3. Strongest

    Another apiary, region or season

    External validation: different bees, boxes, cameras, weather.

    What the score tells you: The nearest thing to how it would do on your hives.

Three ways to test a model, in increasing order of how much the result says about a hive the model has never seen. When images or windows from one hive or one session sit in both training and test sets, information leaks between them and performance looks better than it is; acoustic queen-detection work reports exactly this difficulty in generalising to new hives. Sources: Leakage and the reproducibility crisis in machine-learning-based science; Audio-based identification of beehive states; An intelligent monitoring system for forecasting and anomaly detection in precision beekeeping. Drawing: HoneyHQ (HEKNO Ltd), original.

So look for how the test set was chosen. A random split of frames is the weakest; whole hives held out is stronger; a different apiary, region or season — external validation — is the nearest thing to how a model would do on your hives. 'No held-out test set' and 'not reported' are findings in themselves.

Will it work on my hives?

A model trained in Taiwan, on one hive design, under one lighting rig, may not perform the same in Scotland, on a National hive, in winter or with a different camera. Bees differ by subspecies and strain; boxes differ in material and shape; microphones and cameras differ; seasons change what a colony sounds and looks like. Acoustic queen-detection researchers report this difficulty directly: their models generalised poorly to new hives.

None of this diminishes good research. A strong result on one apiary's data is a real result about that apiary. It is simply a different claim from 'it works', and the scope travels with the number.

Predictions: swarming, queens, failure, flows and yields

Prediction is where claims run furthest ahead of evidence. Three different things are called prediction: recognising a state in data already recorded, detecting an event as it happens, and forecasting it ahead of time. Only the last lets a beekeeper prevent anything, and it is the least established.

Can technology detect this? Where the evidence has reached
QuestionMethods studiedFurthest so farWhat is establishedStill experimentalWhat actually confirms itWhat to check
Can technology measure a colony's Varroa level?Entrance camera, Microphone, Gas sensor array (electronic nose), DNA testing of hive materialField researchCameras can find mites that are visible on bees, with high reported performance on the authors' own datasets.Turning mites seen on passing bees into the colony's infestation; sound and odour signatures of infestation; performance on other hives, cameras and seasons.An alcohol or detergent wash, or a sugar roll, of about 300 bees from the brood area.Do a wash or roll; read any automated count as a reason to, not a replacement.
Can technology predict swarming?Accelerometer in the comb, Temperature sensors, Hive scale, MicrophoneField researchA swarm leaving shows as a temperature warm-up minutes before and a sudden weight drop as it goes.Forward warning days ahead: shown in one programme's colonies from comb vibration, with missed swarms and false alarms, and only within the swarming season.Queen cells containing larvae, seen on inspection; or the swarm itself.Inspect for queen cells.
Can technology tell whether a colony has lost its queen?Microphone, Temperature sensorsField researchNothing at the level of practice.Acoustic classifiers recognise queen absence in the hives they were trained on and generalise poorly to new hives.Eggs and young larvae, or the queen, found on inspection.Look for eggs.
Can a hive scale show a nectar flow?Hive scale, Weather station, Satellite vegetation dataValidated decision supportA sustained net gain beyond the daily cycle, on a colony with no feeding or added boxes, is the standard instrument reading of a flow.Telling nectar from rain, feed and robbing without a log; satellite proxies for the start of a flow.Fresh nectar in the combs, and the forage in flower.Check the log and the weather; look at the combs.
Can technology count bees at the entrance?Entrance counter (light beams), Entrance camera, RFID tagsField researchCounters and cameras produce departure and arrival counts that track activity.Absolute accuracy, which no standard method validates; reading the difference between departures and arrivals as losses.A hand or video count at the same entrance.Check the counter against video before relying on its totals.
Can a camera inspect the brood?Comb photography and cell classification, Thermal camera, Temperature sensorsField researchSoftware can classify cell contents in photographs of frames lifted out of the hive.Imaging without opening the hive; recognising brood disease from images.An inspector looking at the comb.Inspect the brood pattern and any discoloured or sunken cappings.
Can technology estimate how strong a colony is?Comb photography and cell classification, Microphone, Thermal camera, Hive scaleField researchCounting brood and stores from frame photographs gives an inventory of the comb.Population from sound, from the heat of the hive walls, or from weight; how well any transfers between hives.Frames of bees and brood counted on inspection.Count frames of bees and brood; population is not the same as health.
Can technology detect disease or pests?DNA testing of hive material, Gas sensor array (electronic nose), Entrance camera, Radio tracking and harmonic radarField researchLaboratory DNA tests find a target organism's DNA in swabs, debris or honey; for Tropilaelaps a comb swab was the most sensitive method in a 2025 trial.Odour and image signatures of disease; any general detector. No sensor in the sources read recognises foulbrood or viruses in a closed hive.Examination of the colony and, for notifiable diseases, a laboratory test arranged through the official bee inspectorate.Inspect; report suspected notifiable diseases and pests to the authority.
Can technology tell me something is wrong?Hive scale, Temperature sensors, Humidity sensors, MicrophonePrototypeA reading far outside a colony's own pattern can be flagged.Flags that are reliable enough to act on: without labelled events, how many flags are real cannot be measured.Inspection, with the management log and the weather.Check the log and the weather, then look.
Can technology forecast the honey crop?Hive scale, Weather station, Satellite vegetation dataPrototypeA scale records what has come in so far.Forecasting what will come in from weather, phenology and past seasons; no forecast in the sources read has been validated across regions.The harvest weighed.Treat any yield forecast as an estimate with wide uncertainty.

These are inferences about the colony, so none is graded as established practice. 'What actually confirms it' is the reference any model has to be judged against — and the check a beekeeper makes before acting.

Checked 2026-11-01 against: Buzzing with Intelligence: A Systematic Review of Smart Beehive Technologies (Sensors, 2025); Recent developments on precision beekeeping: A systematic literature review (Journal of Agriculture and Food Research, 2023); Tools for Varroa Management: A Guide to Effective Varroa Sampling & Control, Ninth Edition (Honey Bee Health Coalition (Keystone Policy Center), 2026); The prediction of swarming in honeybee colonies using vibrational spectra (Scientific Reports 10:9798 (Ramsey, Bencsik and colleagues), 2020); Remote detection of the swarming of honey bee colonies by single-point temperature monitoring (Biosystems Engineering 148:76–80 (Zacepins, Kviesis, Stalidzans, Liepniece and Meitalovs), 2016); Audio-based identification of beehive states (IEEE ICASSP 2019 (Nolasco, Benetos and colleagues), 2019); Using within-day hive weight changes to measure environmental effects on honey bee colonies (PLoS ONE 13(5):e0197589 (Meikle, Holst, Colin, Weiss, Carroll, McFrederick and Barron), 2018); Approaches, challenges and recent advances in automated bee counting devices: a review (Annals of Applied Biology 180:73–89 (Odemer), 2022); Automatic detection and classification of honey bee comb cells using deep learning (Computers and Electronics in Agriculture 170:105244 (Alves and colleagues), 2020); Long-wave infrared imaging for non-invasive beehive population assessment (Optics Express 19(1):399 (Shaw, Nugent, Johnson, Bromenshenk, Henderson and Debnam), 2011); Improved monitoring of honey bee colony strength via audio IoT sensors, modulation tensorgrams and recurrent neural networks (arXiv preprint 2607.20386 (Abdollahi, Zhu, Guimarães, Coallier, Maucourt, Giovenazzo and Falk), 2026); Evaluation of methods for early detection of Tropilaelaps mites in European honey bee (Apis mellifera) colonies (Scientific Reports 16:22309, 2026); Honey as a source of environmental DNA for the detection and monitoring of honey bee pathogens and parasites (Veterinary Sciences 7(3):113 (Ribani, Utzeri, Taurisano and Fontanesi), 2020); An intelligent monitoring system for forecasting and anomaly detection in precision beekeeping (Scientific Reports 16:7080 (Huet, Bougueroua and Metidji), 2026); Mapping the phenology of European honey bee nectar flows (MSc thesis, University of Edinburgh (Blomstedt)).

For swarming, the strongest forward evidence is from comb vibration in one research programme, with warnings of days to weeks, missed swarms and false alarms recorded, and no held-out test. Temperature and weight catch a swarm as it goes. No monitoring system in the sources read prevents swarming; it can only prompt the inspection that finds queen cells. For colony failure, no validated forecast exists. Honey-yield forecasts from weather, flowering and past seasons are research, and any single figure from one should be read with wide uncertainty.

Swarming, queens, comb and anomalies: what each study tested
StudyTaskDataChecked againstTest designReported result, and what it measuredOutput describesMaturityLimits
Swarm warning from comb vibrationClassification: Classify night-time vibration spectra as from a colony preparing to swarm or not, and raise a daily alarm. Predicts ahead of the event (Averaged warning of 22 days (instantaneous spectra) or about 10 days (ten-day spectral change)).18 swarming and 14 non-swarming datasets; 25 colonies; 3 apiaries; France (Avignon, Jarnioux) and England (Nottingham); 2014 and 2015 swarming seasonsSwarm exit times identified on the daily vibration spectrograms; colonies that did not swarm labelled as not preparingSeveral apiaries; split: no held-out test set; external validation: noClassification error 6.3% (instantaneous); 1.92% (spectral change) — share of spectra misclassified across the whole collection used to tune the method; Missed events 3 of 18; 5 of 18 — swarming datasets with no daily alarm, by method; False alarms 0 of 14; 2 of 14 — non-swarming colonies given a daily alarm, by method. Errors: Raw hourly alarms fired in every colony, swarming or not, which is why the daily night-time alarm was used.The whole colonyField researchNo held-out test set, and several frames from one hive counted as separate datasets. The instantaneous method worked only within the swarming season; autumn swarms were not tested; daytime foraging made spectra unusable.
Swarm take-off from one temperature probeEvent detection: Detect a swarm leaving from the temperature above the brood nest, read once a minute. Detects an event as it happens (Up to about 20 minutes before take-off).Nine swarms; 10 colonies; apiaries: not reported; Latvia; May to August 2015Swarms observed leavingOne apiary; split: not reported; external validation: noNo performance figure in the source read. Errors: Not given in the abstract.The whole colonyField researchDetects a swarm as it leaves — too late to prevent it, though it can tell a beekeeper to look for the cluster. One season, ten colonies.
Queen absence from hive audioClassification: Classify short audio segments as from a hive with or without a queen. Recognises a state in recorded data.Recordings from the hives of a public audio project; colonies: not reported; apiaries: not reported; Not stated in the abstract; Not stated in the abstractQueen status recorded by the beekeepers when the audio was madeRe-analysis of a public dataset; split: whole hives held out; external validation: noNo performance figure in the source read. Errors: Performance fell on hives not seen in training.Each image, clip or time windowField researchThe authors report difficulty generalising to new hives — the central problem for any acoustic classifier.
Comb cell contents from frame photographs (DeepBee)Segmentation: Locate every cell in a comb photograph and classify its contents into seven classes.Labelled comb images released with the paper; colonies: not reported; apiaries: not reported; Portugal (research group); Not stated in the abstractCells labelled by handOne apiary; split: not reported; external validation: not reportedF1 score 94.3% — balance of precision and recall for classifying cell contents, best network (MobileNet); Recall (sensitivity) 98.7% — cell detection rate — cells found in the image. Errors: Per-class errors are not given in the abstract.Each detected objectField researchNeeds frames lifted out and bees removed. Classifying contents is not recognising disease.
Forecasting hive weight, temperature and humidity, and flagging anomaliesAnomaly detection: Forecast weight, temperature and humidity, and flag readings that depart from the forecast or the colony's profile. Predicts ahead of the event (Set by the user; forecasts made one day at a time).Two hives from the open HOBOS dataset; 2 colonies; apiaries: not reported; Germany; 2016–2019For forecasts, the later readings themselves; for anomalies, none — no labelled anomalies existedRe-analysis of a public dataset; split: whole hives held out; external validation: noMean absolute error 0.94 kg (weight); 0.61 °C (temperature); 3.8 points (humidity) — average forecast error on the held-out hive; Share of points flagged 1.8–35% of weight points, by method — share of points flagged as anomalous — not accuracy, since nothing was labelled. Errors: Precision and recall of anomaly detection cannot be measured without labels; one method flagged far more in testing than in training.The whole colonyPrototypeTwo hives in one place; the dashboard ran on simulated data. An anomaly flag says a reading is unusual for that hive, not why.

'No held-out test set' and 'not reported' are findings in themselves: a result tuned and scored on the same data says how well the method fits those colonies, not how it will do on others.

Checked 2026-11-01 against: The prediction of swarming in honeybee colonies using vibrational spectra (Scientific Reports 10:9798 (Ramsey, Bencsik and colleagues), 2020); Remote detection of the swarming of honey bee colonies by single-point temperature monitoring (Biosystems Engineering 148:76–80 (Zacepins, Kviesis, Stalidzans, Liepniece and Meitalovs), 2016); Audio-based identification of beehive states (IEEE ICASSP 2019 (Nolasco, Benetos and colleagues), 2019); Automatic detection and classification of honey bee comb cells using deep learning (Computers and Electronics in Agriculture 170:105244 (Alves and colleagues), 2020); An intelligent monitoring system for forecasting and anomaly detection in precision beekeeping (Scientific Reports 16:7080 (Huet, Bougueroua and Metidji), 2026).

Anomaly detection is not diagnosis

Anomaly detection says 'this hive differs from its expected pattern'. Diagnosis says 'this hive has disease X'. The first can be genuinely useful without knowing the cause — an unusual night is worth a look — but it is a different statement.

A 2026 study shows the difference well. It forecast weight, temperature and humidity on two hives from a public German dataset and compared four anomaly detectors. There were no labelled anomalies, so the authors could report only what share of points each method flagged — between about 2% and 35% of weight readings, depending on the method — not how many flags were right. That is honest reporting, and it means the flags are prompts, not findings.

From an alert to a decision

A monitoring system that is useful to a beekeeper follows a different chain from the one marketing implies. Not sensor → diagnosis → action, but:

  1. ObservationWhat was measured and what changed: a 2 kg drop at 12:40, an unusual night-time sound.
  2. ConfidenceHow sure the system is that something happened, and on what evidence.
  3. Possible explanationsSwarm, harvest, robbing, a box removed — with the management log and the weather already checked.
  4. Next checkThe established check that settles it: look for queen cells, do an alcohol wash, look for eggs.
  5. Human decisionThe beekeeper decides what, if anything, to do. Treatment decisions stay with the beekeeper.

The cost of false alarms and missed events

Accuracy is not usefulness. Every false alarm costs time, an unnecessary inspection and a disturbance to the colony; enough of them and the alerts get ignored. Every missed event costs whatever the event costs — a swarm lost, a Varroa load left to climb. A system with fewer false alarms that misses more events can be better or worse depending on what the event is and how cheap the check is. The useful questions are how many alarms per hive per month, how many were real, and what checking each one takes.

The claims on this page

Every substantive claim is placed on a tier, and the top two tiers must also state what they are not claiming.

Emerging research

Real studies exist and point somewhere, but they are small, in vitro, in animals, or not yet replicated. Direction is not the same as a result.

Emerging research

AI can reliably predict when a colony will swarm.

Comb-vibration spectra gave warnings days to weeks ahead in 25 hives in one research programme, with missed swarms and false alarms recorded, within the swarming season only, and without a held-out test set. Temperature and weight detect a swarm as it leaves. No forecast validated across apiaries was found.

What this does not claim: That swarm warning is impossible: the vibration evidence is genuine field research.

  • The prediction of swarming in honeybee colonies using vibrational spectra — Scientific Reports 10:9798 (Ramsey, Bencsik and colleagues) (2020) · Peer-reviewed article Link
  • Remote detection of the swarming of honey bee colonies by single-point temperature monitoring — Biosystems Engineering 148:76–80 (Zacepins, Kviesis, Stalidzans, Liepniece and Meitalovs) (2016) · Peer-reviewed article

Commonly claimed, but not supported

Widely repeated and not supported by current evidence. Listed so the correction is here rather than only the claim.

Commonly claimed, but not supported

A high accuracy figure shows a model is good at detecting a rare event.

When an event is rare, a model that never detects it can still score high accuracy. Precision and recall, or counts of events found, missed and falsely raised, are needed, and precision-recall analysis is more informative than ROC analysis on imbalanced data.

What this does not claim: That accuracy is never useful — on balanced data it is a fair summary.

  • The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets — PLoS ONE 10(3):e0118432 (Saito and Rehmsmeier) (2015) · Peer-reviewed article Link

Sources

  • Leakage and the reproducibility crisis in machine-learning-based science — Patterns 4:100804 (Kapoor and Narayanan) (2023) · Peer-reviewed article Link
  • The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets — PLoS ONE 10(3):e0118432 (Saito and Rehmsmeier) (2015) · Peer-reviewed article Link
  • Audio-based identification of beehive states — IEEE ICASSP 2019 (Nolasco, Benetos and colleagues) (2019) · Peer-reviewed article Link
  • The prediction of swarming in honeybee colonies using vibrational spectra — Scientific Reports 10:9798 (Ramsey, Bencsik and colleagues) (2020) · Peer-reviewed article Link
  • Remote detection of the swarming of honey bee colonies by single-point temperature monitoring — Biosystems Engineering 148:76–80 (Zacepins, Kviesis, Stalidzans, Liepniece and Meitalovs) (2016) · Peer-reviewed article
  • An intelligent monitoring system for forecasting and anomaly detection in precision beekeeping — Scientific Reports 16:7080 (Huet, Bougueroua and Metidji) (2026) · Peer-reviewed article Link
  • Beehive-entrance imaging and deep learning for real-time monitoring of Varroa destructor in apiculture — Journal of Invertebrate Pathology 214:108465 (Su, Wu, Lin and Liu) (2026) · Peer-reviewed article Link
  • Vit4V: a video classification method for the detection of Varroa destructor from honeybees — Proceedings of the IEEE/CVF CVPR Workshops 2025, pp. 5343–5351 (Giovannesi, Russo and Beraldi) (2025) · Peer-reviewed article Link
  • Buzzing with Intelligence: A Systematic Review of Smart Beehive Technologies — Sensors (2025) · Systematic review Link

How sources are selected and weighted is set out in the sources and evidence policy.

Read next

Tools that use this

HoneyHQ saves your cookie choice on this device. Optional analytics run only if you accept — set a preference now, or change it later on the cookies page.