Scientists are using artificial intelligence to develop fast and accurate ways of identifying pest plants and insects.
Researchers from the Universities of Waikato and Canterbury and from Manaaki Whenua Landcare Research are working together on a project that involves ‘big data’, training computers to recognise species of plants, insects and fungi. It’s a MBIE Smart Ideas project.
The research, based on the application of an artificial intelligence technique called “deep learning”, will be on show at the University of Waikato stand at next week’s Fieldays. Deep learning refers to the way computers can learn to recognise complex patterns by implementing sophisticated algorithms that manipulate large networks of artificial neurons, arranged in multiple layers and feeding them with big quantities of annotated data.
Dr Michael Cree and Associate Professor Eibe Frank from the University of Waikato, Dr Varvara Vetrova from the University of Canterbury, Dr Jerry Cooper from Manaaki Whenua and Dr Brent Martin from Enviro-mark, a subsidiary of Manaaki Whenua, are all working on the project and say their work will be particularly useful for border control, and for farmers.
Dr Cree says once completed, the technology developed in this project will enable farmers, for example, who find a rogue or unrecognisable insect or plant on their property, to take a photo, upload it, and their computer will identify it in a matter of seconds. “We started this work before brown marmorated stink bugs became an issue, but our research is especially relevant with the recent find of the insects in containers of used cars coming into the country,” he says.
Dr Cooper is acting portfolio leader for biosystematics at Manaaki Whenua. He says while this type of research isn’t new, they’re designing general methods to identify organisms that are especially difficult to distinguish and require considerable expertise to identify. “Organisms that we have little knowledge of or perhaps haven’t seen before.”
The correct identification of potential pests, pathogens and weeds is critical to national biosecurity, he says. “We need to monitor the borders to make sure nasty organisms don’t enter the country. New incursions need to be spotted early, when we can do something about them.”
In managing existing problems, scientists need to make sure they are targeting the right species, and not, for example, digging out precious native tussock species thinking they are Nasella (ponytail grass).
Correct identification of problem species currently requires skilled people and time, Dr Cooper says. “We need to cost-effectively speed-up and simplify the process. Our research is investigating the use of cutting-edge image pattern recognition to identify species from simple photographs taken with a smartphone and identified by an app.”
Currently the researchers are testing prototypes on three groups of organisms; plants, fungi and insects. For plants they are looking at Coprosma with many species that are difficult to distinguish from each other. For fungi, the focus is on distinguishing microscopic spores (carried by the wind) of several plant pathogens. For insects, they are looking at a group of moths difficult to identify.
While the stink bug earned its name from its tendency to release an odour when disturbed or when crushed, many other insects share these same characteristics, including some species of ants, beetles and other bugs, Dr Cree says. “It’s a perfect example of how our prototype based on deep neural networks [DNN] might be used. So the Ministry of Primary Industries has expressed interest in our work.”
Dr Vetrova, formerly at Waikato and Manaaki Whenua-Landcare Research, is the lead on the project, currently in its research phase. She says once the prototype is developed, scientists will be able to seamlessly extend it to a group of new species of interest. “Normally ‘deep learning’ requires large amounts of data to develop and train models. However, this is not the case in practice when obtaining labelled data is very costly,” she says.
“We are designing methods that will enable us to build accurate and fast-targeted models based on very small datasets provided by people who discover something they’re not familiar with.”
Fieldays runs 13-16 June at Mystery Creek near Hamilton.