Skip to content
menu icon

Online gallery puts everyone in the weeds control picture

University of Sydney Weed Research director, Associate Professor Michael Walsh
Photo: Supplied: Michael Walsh

Key points

  • An opensource platform of weed images has been developed
  • Called Weed-AI, this industry first is a key step in the ongoing development of weed recognition and weed-specific control
  • It has been developed by researchers at the University of Sydney

An open-source platform of annotated weed images has been developed. Called Weed-AI and developed by researchers at the University of Sydney, it aims to alleviate a major chokepoint in the development of weed recognition technology – the lack of searchable, weeds images.

The platform is a grains industry first. As a searchable, weeds image data platform, it will help facilitate machine learning algorithm research and development, which is necessary for weed recognition in cropping systems.

University of Sydney Weed Research director, Associate Professor Michael Walsh, says although publicly available weed datasets exist, they are often fragmented and collected, stored and managed by multiple institutions.

Weed-AI is different. It enables open-source community contributions. The platform has been designed to support the uploading, storage and downloading of standardised weed imagery.

The accompanying metadata allows the weed images to be sorted according to their agricultural context, such as weed species growth stages and crop situation.

This structure is important in allowing the development of relevant weed recognition algorithms.

Associate Professor Walsh says weed recognition technologies have improved markedly over the past decade and in-crop, site-specific control is fast becoming a reality. This would create the chance to narrow the control of in-crop weeds from whole paddock treatments to weed-specific applications and achieve substantial reductions in weed control inputs.

“Good weed control programs have allowed growers to benefit from reducing weed densities to very low levels in their cropping fields, but what next? The weeds are still there and even at low densities still pose a significant threat,” says Associate Professor Walsh.

“The dilemma for growers is that they can’t afford to relax their control programs, but the field-wide application of weed control treatments targeting low weed densities, at less than one plant per square metre, is extremely wasteful.”

Using annual ryegrass and turnip weed as ‘guinea pig’ weeds, researchers established the Weed-AI database with annotated images of these weeds in wheat and chickpea crops. Then using a machine learning process, these images were used to build weed recognition algorithms.

Associate Professor Walsh says weed recognition accuracy of 60 to 70 per cent was achieved for turnip weed with the initial algorithm. “This was good, but we are obviously aiming for near perfect accuracy.”

The results were not as promising for annual ryegrass with less than 10 per cent accuracy, highlighting this weed’s recognition challenge.

Improving accuracy for both species requires the collection and annotation of many more images. “Essentially the more images that are collected for the training dataset, the more accurate the algorithm will be.

“By making weed recognition technology more readily available, we hope that the weed control industry can focus on the development of much needed alternate weed control technologies suited to in-crop, site-specific weed control.”

The team’s next steps are aimed at streamlining the use of Weed-AI to facilitate user experience and platform capability. “We also need to broaden its potential use across a wider range of weed species and cropping systems. We aim to establish a gold standard weed image database that significantly advances the development of site-specific weed control technologies for the Australian grains industry.”

More information: Michael Walsh, 0448 847272, m.j.walsh@sydney.edu.au

back to top