We’ve used artificial intelligence (AI) to identify a pest stink bug. Now we’re working with our partners on WeedScan, a new app for identifying and reporting weeds.
A photo of a hand holding a phone that is using the WeedScan app. Behind the phone is a photo of weeds.
Blackberries are delicious but one of Australia’s worst weeds.

Weeds cost Australia’s agriculture industry at least $5 billion a year. They make up about 12 per cent of our flora, a higher proportion than any other continent. Identifying and reporting harmful weeds is important for managing both farmed and natural landscapes. Enter the idea for a national scale weed identification and reporting app. 

Identifying weeds

CSIRO botanist Alexander Schmidt Lebuhn specialises in daisies. There are more than 25,000 species in this family. Many daisies, such as dandelions, are successful weeds because their seeds can travel long distances on the wind. 

In recent years, Alexander has expanded his work as a taxonomist classifying daisies. He now uses artificial intelligence (AI) to build apps to identify invasive insects and weeds. The latest app is WeedScan, a collaboration with NSW Department of Primary Industries (DPI) and the Centre for Invasive Species Solutions, funded by National Landcare. 

Alexander is coordinating field work from Tasmania to the Top End to take photos of weeds growing in bushland and on farms.

“We are working on 300 priority weed species that are causing problems in different places all over Australia,” Alexander said. 

“Firstly, we identify each weed plant carefully. Then we take photos of the weed from multiple different angles. We try to find weeds growing at different life stages, such as before it flowers and after it sets seed.

“The next step is using AI to train a weed image classification model so the AI can accurately recognise each weed species. Then the model will be embedded in an app for people to use on their phones,” he said. 

A photo of a man standing outside holding onto a large weed.
Weed experts Richie Southerton (pictured) and Andrew Mitchell are photographing weeds in the wild. The images help train an AI model to recognise priority weeds. 

Test driving WeedScan

The current weed image classification model for WeedScan contains 57 weed species. NSW DPI has built a website prototype featuring a subset of these. 

A group of farmers, land managers and biosecurity staff tested the prototype in Bathurst today. WeedScan is planned to be released for free public use in 2023. 

Once released, WeedScan app users will be able to identify priority weeds by simply taking a photo on their phone. App users will also be able to submit records to biosecurity staff, supporting local action to manage weeds. 

A close up photograph of a small weed in the ground.
Xanthium spinosum is a weed in the daisy family.

33 comments

  1. To promote responsible and effective chemical use for control of identified weed csiro should include a tab with that information. Eg what chemical at what growth stage and how to minimise misuse or ineffective use of chemicals.

  2. Yes being rural on land important to know whats what and control the spread.

  3. Is this based on iNaturalist? Can it be used in conjunction with iNat? I am assuming as it’s CSIRO that the data is FAIR and will be lodged in the ALA?

  4. It looks great but some weeds get a bad rap. Invasive weeds such as dandelions are popular with bees, and swanbush/milkweed plants are used by butterflies for food and as a nursery for their eggs and babies. Same with some species of daisies. Hopefully there will be relevant info on the ‘weeds’ that provide shelter and food for insects too? I would love to share this with my local council in South Australia whenever that is allowed. Good work!

  5. A great idea but I hope the app does not jut operate on-line (you have to have internet service. The app should have an option for submitting a photo(s) for analysis. Weeds are not just where there is an internet service

What do you think?

We love hearing from you, but we have a few guidelines.