Help us train our bee detector

What the bee detector actually is

The bee detector is an image classifier called bee-swarm-id. It looks at a single photo of an insect cluster and sorts it into one of three classes: honey bees, bumble bees, or wasps. It is a whole-image classifier rather than an object detector, so it never draws a box around an individual insect. It judges the picture as a whole, which is the same thing a beekeeper does from across a garden.

The weights are open. The model is a CAFormer backbone fine tuned on bee and wasp imagery, published at Swarmed/bee-swarm-id on Hugging Face under an MIT licence. Anyone can download it, audit it, or run it against their own photographs.

Why a machine still needs people

Because the mistakes it makes are the expensive kind. Wasps are the hard class: a yellowjacket nest under a soffit and a honey bee swarm on a fence post can occupy the same few hundred pixels of a phone photograph. Get that wrong in the cautious direction and a beekeeper drives out to a wasp nest. Get it wrong in the other direction and a real colony is written off as wasps and never gets a beekeeper at all.

Human labels are what moved those numbers. The version of the model running today was retrained on 4,664 consensus labels collected on this page, and the held out comparison against the version before it is stark. Confident wasp calls went from correct about 15 percent of the time to about 72 percent. Recall on honey bees, meaning the share of real swarms the model actually recognises as swarms, went from 64 percent to 98 percent. In plain terms, roughly a third of genuine swarms used to be quietly filed as something else.

Wasps are still the binding constraint. Recall on the wasp class sits near 54 percent, and the reason is simply scarcity: there were 416 wasp images in the training set against several thousand of bees. Every wasp you tag here is worth more than several bees.

How one tap becomes training data

No single vote decides anything. A photograph needs three independent people to look at it, and a label is only recorded once at least two thirds of them agree. Disagreement is informative too: an image that never reaches consensus is usually genuinely ambiguous, and those are worth knowing about.

The queue you are served is deliberately mixed rather than random. Most of it is drawn from the band where the model was unsure of itself. A fifth is drawn from photographs it was extremely confident about, which sounds wasteful and is not: a model that is confident and wrong is invisible until a human checks, and we found dozens of cases where beekeepers physically visited, found wasps, and the model had called them honey bees at the very top of its confidence range. The rest comes from a backlog of photographs belonging to swarms a beekeeper later confirmed were wasps, which are the single most valuable images we have.

So far this has produced 8,630 agreed labels from about 60,000 individual votes, including 430 confirmed wasps and 44 bumble bees.

One thing we tried that failed

The obvious shortcut is to skip the crowd and train on public natural history photographs instead. We tested exactly that, adding thousands of iNaturalist images of honey bee swarms, bumblebees and vespid wasps to the training set, evaluated against an identical held out split.

It didn’t help, and on the wasp class it hurt slightly. The reason is domain shift. A naturalist photographs an insect deliberately, close up, in focus, often on a flower. A member of the public photographs a swarm from three metres away, into the light, one handed, because there are thousands of bees on their car. Those are different distributions, and the second one is the one that matters. Which is why this page exists: labelled in-situ phone photographs are the lever, and there is no dataset you can download to replace them.

Where the labels end up

The classifier is not an end in itself. Its verdict decides whether a report counts as a honey bee swarm in everything downstream, so a wasp nest misfiled as a swarm quietly distorts the maps, the seasonal curves, and the trap placement guidance that beekeepers actually act on. Cleaner labels here mean the published work rests on honey bee reports rather than on a mixture of bees and wasps.

You can see what that supports on the Swarmed data hub, which collects our open analysis across 39 countries. The 2026 North American swarm season report tracks how early spring arrived and what it did to peak swarming. The forage outlook maps what is in bloom and where. Across all of it we try to be honest about the fact that a map of reported swarms is partly a map of where people live.