WeedNet AI Learns 1,593 Weed Species From Citizen-Science Photos, With University of Arizona Help

Every gardener knows the moment. Something green pops up between the carrots or along the drip line, and the question is whether it is a volunteer worth keeping or a weed about to drop a thousand seeds. A team of researchers at Iowa State University and the University of Arizona has built an AI model meant to answer that question from a photo, and it learned largely from pictures taken by ordinary people.

The model, called WeedNet, was described in a study published September 30 in the journal Nature Communications. According to the paper, WeedNet identified weeds with 91.02 percent accuracy across 1,593 species, and 41 percent of those species were identified with 100 percent accuracy in testing. When the team fine-tuned the global model for one region, a version built for 84 Iowa weeds reached 97.38 percent accuracy.

Two of the paper’s authors, Zi K. Deng and Nirav Merchant, are with the University of Arizona’s Data Science Institute in Tucson. The study’s corresponding authors are Soumik Sarkar and Arti Singh of Iowa State.

Built on backyard photos

The hardest part of training a plant-identification AI is getting enough reliable pictures. Expert-labeled weed images are scarce and expensive to produce, and many look-alike species differ only in a leaf hair or a flower detail.

The WeedNet team turned to iNaturalist, the free community science platform where users post photos of plants and animals and other users help confirm the identifications. An earlier preprint version of the study describes how the researchers filtered iNaturalist down to weed-relevant species using invasive-species lists and other references, dropped any species with fewer than 100 research-grade images, and ended up with roughly 14 million images. The published paper credits iNaturalist as its data source.

The model was first taught to recognize general visual patterns from unlabeled images before it was trained on labeled weeds. That approach, called self-supervised learning, lets the AI absorb leaf shapes, stem structure, and textures before it ever has to name a species.

Where it gets confused

The most useful findings for gardeners may be about the model’s weak spots, because they mirror the weak spots in human identification.

According to the published abstract, the diversity of the images used in training, covering all growth stages and the plant traits experts use to tell species apart, was crucial to performance. The preprint’s tests show why. On a set of web images of 84 Iowa species, accuracy was 65.8 percent for seedlings, 80.2 percent for plants in the vegetative stage, and 85.7 percent for flowering or fruiting plants. A tiny seedling with a couple of leaves simply does not carry much information.

Look-alike species were harder still. In the preprint, accuracy fell to 60.6 percent on a set of 24 look-alike species, and grasses were the toughest group. When the researchers fed the model close-up photos of the grass collar, the spot where the leaf blade meets the stem, it correctly identified 11 of 15 grass species.

Some of the mix-ups involve plants Arizona land managers know well. University of Arizona Extension lists saltcedar among woody perennial weeds, and the preprint reports that one saltcedar species, Tamarix ramosissima, was identified with 100 percent accuracy in testing. Two related tamarisks, with a few hundred training photos each compared with more than 5,500 for that cousin, scored zero, and the preprint’s examples show them being labeled as the better-photographed species. The team attributes many errors like this to imbalance in the training data: species with far more photos tend to pull uncertain identifications their way.

To keep the model honest, the researchers built in two safeguards. One flags images that do not look like anything the model was trained on. The other, called conformal prediction, returns a short list of possible species instead of a single confident answer when the model is unsure. In one example from the preprint, the model offered Palmer amaranth at 56 percent and spiny amaranth at 44 percent rather than picking one.

From research model to a tool you can try

WeedNet already sits inside a working tool. Iowa State combined its weed model with an insect model in a web-based tool called Pest-ID. In an October 2025 feature, the university described how a user chooses insect or weed, uploads a photo, and within seconds gets the scientific name, notes on the plant’s impact, and an “Ask our AI” box for management questions. The researchers said the chatbot’s answers draw on expert-developed and curated bulletins and papers.

As of an April 2026 report in the Iowa Capital Dispatch, Pest-ID was a webpage being tested by farmers, with a phone app planned. Singh told the outlet the weed model was trained on 15 million images and could identify 1,500 weed species, while the insect model covered 4,000 species. A two-year, $400,000 National Science Foundation grant announced in March is supporting work to add crop disease identification.

Sarkar has argued that specialized models have an edge over general-purpose AI. “Most can detect a charismatic species like a monarch butterfly,” he said in the 2025 feature. “But with specific invasive species, our models do much better than the general-purpose tools.” The preprint reports that across roughly 60 invasive plants listed by the National Invasive Species Information Center, WeedNet averaged 96.67 percent accuracy.

The team also tested the model on images from a drone and a ground robot in Iowa fields, which points toward future machines that can spot and treat individual weeds. For now, though, the practical version for most people is a phone and a photo.

What it means for a desert garden this fall

The timing is good for Arizona gardeners. In a 2019 University of Arizona Extension newsletter, turf agent Kai Umeda described a winter weed season that began when October rains germinated seeds, bringing on cheeseweed, common groundsel, filaree, red brome, London rocket, shepherd’s purse, and sowthistle. His advice was to be ready after each rain to deal with emerging weeds while they are small, noting that most herbicide labels call for treating weeds at a small growth stage, usually under 4 to 6 inches. In landscaped beds, he wrote, hoeing and hand-pulling are among the few options.

A University of Arizona Extension bulletin on landscape weeds puts identification first on its list of steps. Knowing a weed’s life cycle tells you when it is vulnerable. Winter annuals, for example, sprout from late summer through early spring and die off in late spring or early summer once temperatures climb past about 85 degrees. For annual weeds, the bulletin’s first mechanical advice is to cut, pull, or hoe them before they go to seed.

The WeedNet research suggests a few ways to get better answers from any photo-based ID tool. Wait a little if you can, since young seedlings are the hardest to name. Take more than one picture, including leaves, any flowers or seed heads, and for grasses, a close-up of the collar. Treat a short list of possibilities as a prompt to look closer rather than a final verdict. When the stakes are high, the same Extension bulletin recommends bagging a sample, ideally with flowers and roots, and taking it to a Cooperative Extension office, nursery, or garden center.

There is also a way to give back. Every research-grade weed photo posted to iNaturalist from a Valley yard or wash adds to the kind of image collection that made WeedNet possible, and the researchers’ own results show that varied photos, covering every growth stage, are what make the model better. The weeds coming up after the next fall rain are worth a quick picture before they go in the bin.