Environment Canterbury has taken its first steps into utilising Artificial Intelligence (AI) to identify pests from aerial imagery.
A successful pilot programme in South Canterbury has shown that AI may provide a more efficient way to inspect large swathes of land for biosecurity threats such as Nassella tussock.
The tussock is an invasive grass that is spreading in Canterbury. It displaces productive pastures and is unpalatable to stock. Only 70 years ago, Nassella forced some farmers to leave the land after it ran rampant, taking over entire farms, leaving no room for crops or stock.
Each year, Environment Canterbury’s biosecurity team spend months inspecting properties and ensuring landowners identify and control this pest plant and control its spread – a requirement of the Canterbury Regional Pest Management Plan.
Environment Canterbury recently partnered with Lynker Analytics to develop and train an AI program to identify the pest plant from aerial imagery. High resolution imagery, collected by a small plane, was captured with the support of the landowner.
Lynker Analytics managing director Matt Lythe said the test image relies on high quality images. The photo used in the trial was 13GB in size and contained 3.6 billion pixels.
This high number of pixels creates a processing challenge, as the machine learning process needs to assess every pixel. When training the AI model, we provided examples of Nassella and ‘not-Nassella’ in the image
The machine learning model was trained by splitting up the large image into non-overlapping 512x512 pieces. These are called train, validation, and holdout sets, Matt explained. This allowed the programme to learn over time to identify a pixel or group of pixels as either ‘Nassella’ or ‘not-Nassella’.
Nassella Tussock coverage predicted by AI
The benefits for the regional council in having AI identify pests through imagery mean greater efficiencies in identifying unknown infestations, monitoring, and potentially for initial inspections, reducing the need for biosecurity staff to physically inspect the land as often.
And while it’s still a while away from being implemented across the region, landowners may in future be able to use the AI to ascertain infestation levels on their own properties and plan control programmes accordingly.
Environment Canterbury biosecurity team leader Matt Smith said the initial results showed the AI to have around a 90 per cent success rate at identifying large mature Nassella tussock. More refining of the model is required to increase that identification rate.
“An AgResearch study found that hand control operations remove about 30-35 per cent of Nassella tussock, mostly very small plants, which means about 65-70 per cent is left behind,” Matt said.
Mature Nassella tussock can produce up to 100,000 seeds, so it’s important to remove all infestations prior to viable seed being produced each year.
Environment Canterbury’s Nassella tussock programme is mostly carried out on foot, but Matt said there could be many benefits to wider biosecurity programmes in future, including monitoring pest trends over time.
Nassella Tussock coverage predicted by AI
“In biosecurity, things have been done the same way for a long time and now we have this amazing technology that is becoming available, and we are looking to the future. It’s a big part of how we can protect the region going into the future.
“We’re embracing new technologies for biosecurity, especially when they can increase productivity both for our work programmes and for affected landowners,” Matt Smith said.