AI Varroa Mite Counting: How Accurate Is It Really?

The Short Answer

Camera-based mite counting works best on sticky boards, and the published evidence there is genuinely good: in controlled studies, AI counts track careful human counts closely, and at high mite loads they beat tired human eyes. The evidence is much thinner for the two other things people mean by "AI mite counting": photographing live bees on a frame or at the entrance, and photographing the mites from an alcohol wash. For that last one we could not find a single peer-reviewed validation.

Meanwhile the reference method hasn't changed. The Honey Bee Health Coalition's Tools for Varroa Management, ninth edition (June 2026), still builds every treatment threshold on mites per 100 adult bees from a roughly 300-bee wash. At the new thresholds the decision point is 3 or 6 mites in that sample. At those numbers counting takes seconds. What makes a count wrong is usually the sample, not the eyes counting it.

This article looks at what the studies measured, where the software gets it wrong, and how to judge a vendor's accuracy number. Full disclosure: HiveSense doesn't count mites from photos. The last section explains why.

"AI Mite Counting" Means Three Different Things

Accuracy claims only mean something once you know what the camera was looking at. The three approaches are very different problems.

What the Published Studies Found

Sticky boards: the strongest evidence

• Divasón et al., Sensors, 2024. They trained an object detector on smartphone photos of sticky boards: 64 images at 48 megapixels, containing 807 mites. It reached a mean average precision of 0.907 on the validation set. The authors said lower-resolution images "fail to capture enough Varroa mite details". Detection ran on a GPU, not on the phone.
• Yániz et al., Agriculture, 2025 (VarroDetector). This study covered 204 sheets photographed under controlled lighting with several smartphone models at a minimum of 48 megapixels. The software's counts were "highly correlated with the control (R² = 0.98 to 0.99)", including on a phone it had never been trained on. Above 50 mites per sheet it was more reliable than trained human counters. The tool is free and open source, runs on a desktop rather than a phone, and its documentation still tells you to check the results by eye.
• Scutaru et al., Insects, 2025 (BeeVS scanner). This study covered 85 sheets from 5 hives over 17 weeks. The scanner's counts correlated with the reference at 0.998, against 0.967 for human visual counts. Human error ran 15–18% on sheets with more than 10 mites, while the scanner stayed under 1%. On sheets with 0–2 mites, though, the scanner counted 64 mites where people found 21. Its most common mistake was reading red or purple pollen as mites.
• Bugnon et al., Agrarforschung Schweiz, 2021 (ApiZoom, with Agroscope). They reported that under optimal conditions the model "approaches human visual acuity". Results depended on the camera, the light at the apiary, and how much debris was on the tray.

So on sticky boards, image counting really can match or beat people, provided you control the image. That's a real result, and it's worth saying plainly.

Live bees on frames and at the entrance

• Bilik et al., Sensors, 2021. Across 600 images, the best detector reached an F1 score of up to 0.874 for spotting infested bees but only 0.727 for detecting the mites themselves.
• Bjerge et al., Computers and Electronics in Agriculture, 2019. This camera unit filmed bees shaken off a frame in front of the hive. In one test video of 1,775 bees it measured 5.80% infestation against a ground truth of 5.52%. That's promising, but it is a single test.
• BeeScanning is a Swedish app that photographs bees on brood frames. By its own published figures it finds roughly 8 in 10 of the mites that are optically visible. It applies a correction factor calibrated against alcohol washes to account for mites hidden under the bees. We found no peer-reviewed validation, but BeeScanning is at least transparent about the gap, which is more than most vendors offer.

The underlying limit is biological. On visually checking adult bees, the HBHC guide says mites are difficult to see "unless mites are on thorax or top of abdomen". A camera faces the same problem as your eye: it can only count the mites that are in view. That is why BeeScanning has to apply a correction factor at all.

Photos of an alcohol wash

This is the use most phone apps advertise: do your wash, pour the mites onto a white surface, photograph them, and let the app count. We searched PubMed and the wider literature and found no peer-reviewed study validating it. That doesn't make it useless. Washed mites on a white background are an easier image problem than sticky-board debris. It does mean that any accuracy number you see for it is unverified.

About That "95.3%" Figure

The most-cited number in this space comes from a page on beefamily.net, published under the HiveLog AI name. It compares "65% Manual Counting", "82% Phone + Manual Count" and "95.3% AI-Powered Detection". When we checked it on 2026-09-22, the page gave no study, footnote or method for any of those three numbers.

To be fair, a missing citation doesn't mean the tool performs badly. It means nobody outside the company can check it. Before you trust an accuracy figure from any vendor, ours included, ask it five questions:

• Accurate compared with what? Another person's count, an expert's count, or a full-colony mite estimate?
• At what mite counts? The BeeVS study shows software can be excellent at 100 mites and badly wrong at 1.
• Under what imaging conditions? The strongest results used 48-megapixel images and controlled lighting.
• How many samples, from how many apiaries?
• Is it published anywhere you can read?

Why the Alcohol Wash Is Still the Reference

The HBHC ninth edition is direct about this. The alcohol or soap wash is "a more reliable method than the powdered sugar shake". Sticky boards are listed under less reliable methods, because a board is "a whole-colony measure and not a per-bee measure". The UK National Bee Unit calls natural mite drop "not a particularly accurate method for determining mite numbers".

Wash versus sugar roll is less settled than forum arguments suggest. A 2025 study in the Journal of Apicultural Research found the sugar shake "more prone to underestimating mite levels". A 2022 review in Bee World concluded the comparative evidence was "sparse and contradictory". Either way, the wash is what the thresholds are built on.

Here is the arithmetic that matters. HBHC's new thresholds are 1% during the dormant and population-increase phases, and 2% at peak population and population decrease:

Counting six mites at the bottom of a jar isn't where a count goes wrong. It goes wrong in the sample. HBHC says "using fewer than 300 bees will not give you an accurate sample", and the bees should come from brood-nest combs, not the outside frames. The guide also says sampling results "are an estimate of the infestation". It recommends repeated sampling so you can see whether your management is working, instead of treating one wash as the final answer.

Use the varroa calculator to turn a count into a percentage, and the varroa mite treatment guide for what to do once you're over the line.

When a Camera Tool Genuinely Helps

Being honest cuts both ways. Image counting has real advantages in some situations:

• High-count sticky boards. When a board holds 80 mites among the debris, people get tired and miss them. The studies show software doing better than trained counters in exactly this situation.
• Many hives on a board-based monitoring routine. Photographing a stack of boards is faster than counting each one under a magnifier.
• Non-lethal monitoring. Frame-photo tools like BeeScanning don't kill a 300-bee sample. Some beekeepers reasonably value that, as long as they accept the correction factor that comes with it.

If you use one of these tools, calibrate it against your own alcohol washes a few times each season. That way you'll know how it performs in your apiary, with your phone and your light.

Frequently Asked Questions

Can a phone app accurately count varroa mites?

On sticky boards, yes, under the right conditions. Published studies of smartphone photos of sticky boards report counts that track expert counts closely, but they used 48-megapixel images under controlled lighting. We found no peer-reviewed study validating a consumer app that counts mites from a photo of an alcohol-wash sample, so accuracy figures for that use are vendor claims until someone publishes the method.

Is AI mite counting more accurate than counting by eye?

At high mite loads it can be. In a 2025 study of an AI sticky-sheet scanner, human counting errors were around 15 to 18 percent on sheets with more than 10 mites, while the scanner was under 1 percent. At very low counts the result flipped: on sheets with 0 to 2 mites the scanner counted 64 mites where humans found 21, mostly by mistaking red or purple pollen for mites.

Why is the alcohol wash still the standard method?

Because it measures mites per bee on the adult bees that matter, from a known sample size. The Honey Bee Health Coalition calls the alcohol or soap wash more reliable than the powdered sugar shake and lists sticky boards among the less reliable methods, since a board is a whole-colony measure rather than a per-bee one. Every treatment threshold in its guide is expressed as mites per 100 adult bees.

How many mites in an alcohol wash means I should treat?

The Honey Bee Health Coalition’s ninth edition guide, published June 2026, sets the prompt-control threshold at 1 percent during the dormant and population-increase phases and 2 percent at peak population and during population decrease. In a standard 300-bee sample that is 3 mites and 6 mites respectively. Check your local extension service or bee inspector as well, since thresholds vary by region.

Does HiveSense count mites from photos?

No. HiveSense records the mite counts you make yourself, by alcohol wash, sugar roll, CO2 or sticky board, and keeps them per hive alongside your treatments so you can see the trend. It does not analyse photos for mites, and we do not claim it does.

The HiveSense Angle

HiveSense doesn't count mites from photos, and it doesn't pretend to. That's a deliberate choice. The evidence above says counting is rarely the weak link in a wash. The weak links are taking a proper sample and keeping the result. That second part is software's job, and it's the part we built.

A mite check in HiveSense records the method (alcohol wash, sugar roll, CO2 or sticky board) and the count. For a standard ~300-bee sample it shows you the mites per 100 bees right away. Counts sit per hive next to your treatment log, so a count from before treatment and one taken 10–14 days after end up side by side. That comparison is what tells you whether a product actually worked, which matters more every year as amitraz resistance spreads.

The question people ask us most is how the voice feature works, and this is a good example. With a jar in one hand and gloves on, say "counted six mites" during a voice inspection and the varroa count field fills in for you to check before you save. The transcription runs on the phone, so it works in a yard with no signal. More on the tracking side is on the varroa tracker page.