https://ift.tt/UiKH6IC
Agricultural services provider PGG Wrightson has partnered with Kiwi artificial intelligence specialist Inde to destress the livestock counting process by combining Microsoft Azure, Power Apps and drones.
Livestock counting has been a traditional bugbear for farmers, requiring time, cost and effort to confirm animal numbers for audit purposes, especially for large properties and in varied or challenging terrain.
Rounding up and bringing animals in from the paddocks can take up to a week for large properties with multiple mobs of cattle or sheep. For dairy cows, stock counts can range from 200 to 2,000.
Accuracy was essential, as discrepancies could lead to significant financial losses.
“In today’s environment, farmers are counting every penny, so every animal that’s missing is a big hole in your books,” said Roger Lamb, national operations manager of livestock at PGG Wrightson, which commissions stock agents to audit the process.
“Some animals might be worth up to $20,000, so having any of them missing means your accounting is out by a lot.”
Farmers were also keen to promote the wellbeing and productivity of their animals, and lower the risk of potential injuries and stress, by reducing the need to herd them.
“Traditional farm audits can take five or six staff off the road for a week,” said Roger Lamb, project owner at PGG Wrightson. “I thought: there has to be a better way.”
While the team looked into drones and AI-powered stock counts, no one was doing exactly what they were looking for. They decided to build the solution, now dubbed SkyCount, with the help of AI partner Inde.
Inde chief technology officer Rik Roberts threw himself into the mission, first by working out which drones and cameras were right for the task of capturing the imagery needed to train the AI platform.
“It was about finding a balance,” Roberts said. “Larger drones can carry bigger cameras, but these posed issues with flight planning.
“If you’re flying them near controlled airspace, you need to get permission from the Civil Aviation Authority. However, smaller drones flying too low can frighten the animals.”
Testing brought up some interesting challenges: Infrared cameras couldn’t “see” sheep through their thick fleece while still cameras required lots of shots to be stitched together, making the final product blurry, resulting in lost or duplicated animals.
Also, while beef cattle were unbothered by the sound of drones, dairy cattle didn’t like it at all, causing stress.
After months of testing, it was determined mid-sized, quieter drones could operate closer to animals without distress. DJI’s Mavic Enterprise Series drones, which shoot regular video through HD cameras rather than infra-red, were selected for the job.
The next hurdle was training the AI models.
“Our team spent countless hours collecting high-quality aerial images we could feed our model, including the characteristics of different breeds, to photographing them in various topographies and at different times of the day, to ensure shadows, trees and rocks were not mistaken for animals and vice versa,” Roberts said.
Tens of thousands of images were captured and processed in Microsoft Azure to train the AI model.
The team also needed to establish automated mission planning to execute flight plans consistently and efficiently in remote areas with limited connectivity.
“If the property does not already have a digital map, we’ll do an initial mapping mission,” Roberts said. “We can define all the drones’ flight altitudes and their directions, and angle the cameras so that the second time, the third time, fourth time we go out for a stock count, they’re flying the same missions.”
SkyCount debuted at the Fieldays agricultural show in June,