Lab · Technology

AI pest identification

Quick answerVerified September 23, 2026

What is ai pest identification?

Software that names a pest from a photograph, usually a convolutional neural network trained on labelled insect images. Accuracy claims above 95% are common and are usually true — of the conditions they were measured in. Field conditions are not those conditions, and the gap is where the buying decision lives.

Reviewed by LTK editorial team

Where it fits

Triage and pre-sorting, where the job is reducing a large pile of images down to the ones a human should actually look at. Volume reduction is a real saving even at imperfect accuracy, because the cost of a mistake is a second look rather than a wrong treatment.
Fixed imaging rigs with controlled lighting and a small, known species list. The published 96%-plus results for stored-product species were achieved on top-down images of five known beetles — that is a realistic description of a mill or warehouse monitoring station, and a poor description of a crawlspace.
Stretching scarce taxonomic expertise. Manual morphological identification needs wing venation and body segmentation read under a microscope by someone trained to do it, and there are not enough of those people. DNA barcoding is accurate but too costly and logistically awkward for routine field volume.
Surveillance programmes where the output is a trend line rather than a verdict on one specimen — mosquito sorting and similar high-count, low-stakes-per-image work.

Where it does not

This is the part the brochure leaves out, so it is the part worth reading.

Anywhere the answer is legally or commercially binding. A WDO report, a quarantine call, or an invasive-species determination is not a place for a probabilistic guess with no stated reasoning. CNNs are black boxes — the published work reaches for Grad-CAM visualisation specifically because the model cannot say why it decided what it decided.
Species that look like each other. Inter-species similarity is the named hard problem: published work uses Prodenia litura and the meadow moth as the standing example of two species that are strikingly similar externally. If your two candidate species are the ones that matter commercially, similarity is exactly the case the model handles worst.
Life stages. Intra-species variability across adult, pupa, larva and egg is as large as the difference between some species. A model trained on adults is not a model that identifies what you actually find in an inspection.
Small or dark specimens, and cluttered real-world images. Where average accuracy is quoted at 96%, per-class accuracy on a large field benchmark tells a different story: Miridae 42%, aphids 48% — driven by small size, varied appearance and high intra-class variance. THE AVERAGE HIDES THE FAILURES, and the classes it fails on are not random.
Any system that always returns an answer. The best published practice is a model that REFUSES to predict on low-resolution, blurred or confusing images, precisely because misclassifying an unseen invasive species as something benign is the catastrophic outcome. Ask a vendor what their tool does when it does not know. If the answer is that it always gives its best guess, that is a defect being sold as a feature.

Sources

See what members use