Neural networks and drones may improve seabird counts

June 6, 2021
Drone image captured over a seabird colony on Steeple Jason Island. (Wade Sedgwick and Vivon Crawford, WCS)

Drone image captured over a seabird colony on Steeple Jason Island. (Wade Sedgwick and Vivon Crawford, WCS)

Researchers from Duke University's Marine Lab have developed an approach combining neural networks and drones that can locate and count hundreds of thousands of seabirds, helping conservationists obtain a more comprehensive view of how climate change and other dynamics have affected wildlife populations.

The Duke researchers tracked birds far from North Carolina — specifically, they examined species of albatrosses and penguins that lay eggs on the Falkland Islands, a remote archipelago off the coast of southern Argentina, where birds far outnumber the more than 3,000 human inhabitants. Their contribution, detailed in a May 22 Ornithological Applications paper, could enable scientists to track seabird populations more accurately and efficiently than they can through manual counting.

"Traditional monitoring is expensive and very time-intensive, and it can introduce disturbance to the birds," Madeline C. Hayes, a research technician at the Duke Marine Lab and the first author on the paper, told The Academic Times. "With our automated workflow, we were able to count these birds with a really high accuracy by reducing human error and disturbance. We believe that this is the most accurate mapping of this area ever done."

Drones had previously been used to count seabird, whale and porpoise populations from an aerial perspective, allowing researchers to spot wildlife in otherwise remote or inaccessible areas. But although drones provided a helpful vantage point, the birds captured in the images still needed to be counted by hand — a labor-intensive task, especially since half a million mating pairs of black-browed albatrosses build nests along the Falkland Islands' craggy coastlines each year. 

Argentina's Wildlife Conservation Society has for several years conducted population monitoring with drones. But it often tracks only a subset of the island-wide bird population and estimates the total number of birds based on that relatively small sample. Hoping to increase its capacity, the Society recruited the team at Duke to incorporate deep-learning technology that would help automate the process. 

Researchers employ an ordinary, off-the-shelf drone, albeit one with a sensitive camera capable of taking extremely high-definition pictures — one pixel corresponds to just half a centimeter, or less than 0.2 inches, on the ground. The drone captures images by carving parallel lines along an automated flight path. Hayes then stitches each of the images together to create a comprehensive map of the area.

"It gives you a single snapshot of that colony at that point in time," Hayes said. "It lets you explore the patterns and processes of the larger ecosystem that would be missed if you're just looking at one single image."

The neural network learns in a similar way as the human brain — taking in new information and adjusting its detection methods accordingly, Hayes explained. She had to label thousands of birds by hand to teach the network how to locate both species' unique shape, size and color.

Black-browed albatrosses, named for the distinctive black dash above their eyes, can live for 30 years or more and have a wingspan between 210 and 250 centimeters (approximately 6.9 to 8.2 feet), according to the Australian Antarctic Program. That's a short length compared with other species: Wandering albatrosses, for instance, can reach a wingspan approaching 12 feet. The colonizing birds are also noted for their exceptional flying ability, with Smithsonian Magazine reporting that some species soar for the first five or six years of their life without ever touching the ground. And earlier this year, a Laysan albatross hatched a chick at the age of 70

Meanwhile, southern rockhopper penguins, named for their tendency to jump over environmental obstacles, have pointy yellow feathers above their eyes and are among the smallest penguins in the world. 

Both the albatross and the penguin species mostly feed on fish, krill and other ocean inhabitants and reside around the waters and islands north of Antarctica. Their numbers can reflect the health of entire ecosystems: The birds' roles as marine predators mean that population declines may be indicative of systemic issues such as climate change and unsustainable fishing practices. 

Conservationists' counts can also reveal the state of a species itself. Black-browed albatrosses had previously been considered endangered by the International Union for Conservation of Nature but have recently seen a notable population increase. Southern rockhopper penguins, on the other hand, are designated by the IUCN as a vulnerable species, according to the World Wildlife Fund.

The two species frequently nest together — albatrosses build mud mounds to store their eggs, and the much smaller rockhopper penguins nest amid those albatross colonies. "So, if you're on the ground, just looking straight onto a colony, it's going to be hard to even find those penguins, let alone accurately count them," Hayes said.

Despite those challenges, the neural network could identify black-browed albatrosses with nearly 98% accuracy and southern rockhopper penguins with approximately 87% accuracy, using drone images collected in 2018 and 2019. This high level of accuracy gives researchers confidence that their deep-learning technique could be deployed in upcoming analyses to provide a faster count of birds across the entire island chain. The neural network could also be trained to detect the population dynamics of other elusive seabirds.

"The model would be able to learn really quickly what that other type of penguin or albatross looks like," Hayes said. "Seabirds, in general, are good indicators of ecological change. So if we can study what's happening at these local levels, then they can act as a canary in the coal mine for the rest of the ecosystem."

The study "Drones and deep learning produce accurate and efficient monitoring of large-scale seabird colonies" published May 22 in Ornithological Applications, was authored by Madeline C. Hayes, Patrick C. Gray and David W. Johnston, Duke University; Guillermo Harris, Wade C. Sedgwick and Vivon D. Crawford, Wildlife Conservation Society; Sarah Crofts, Falklands Conservation; and Natalie Chazal, North Carolina State University.

Saving
We use cookies to improve your experience on our site and to show you relevant advertising.