- Unmanned aerial vehicles (UAVs) have great potential for surveying wildlife, especially species that assemble in large numbers and that are easily disturbed by human presence.
- Scientists creatively combined high-tech UAVs and computer-vision algorithms with rubber ducks to assess the potential of aerial imagery to count seabirds relative to traditional survey methods.
- They found that both human and semi-automated computer counts of colony-nesting birds from UAV-derived images were more accurate and less variable than counts made by observers on the ground.
- Combining UAV-derived imagery with artificial intelligence can help scientists more accurately estimate population sizes with less variability.
Jarrod Hodgson is one of very few scientists who have used rubber ducks as part of their Ph.D. research.
Hodgson and colleagues at the University of Adelaide compared the accuracy of counts of birds on an Australian beach from images taken from an unmanned aerial vehicle (UAV) to counts by ground observers. They brought in the ducks to serve as faux seabird colonies, each with a known number of individuals.
Their findings, published this week in the journal Methods of Ecology and Evolution, suggest that aerial imagery can offer scientists more accurate counts of at least some species than even experienced observers on the ground.
Image vs. in-person observations
UAV-derived imagery is increasingly being used to survey and monitor wildlife, including detecting and monitoring individual koalas and surveying orangutan and chimpanzee nests, but few researchers have tested the accuracy of UAV-based data collection relative to other, traditional methods.
Counting birds and other colonial species from the ground is liable to miss some animals and double-count others. It also requires experts to invest time visiting a site, sometimes repeatedly, to collect the data, and their presence may scare or alter the behavior of the animals they are trying to study. Thus the drones and the ducks.
The researchers simulated 10 breeding colonies of greater crested terns (Thalasseus bergii), a local seabird, by placing commercial, life-size plastic duck decoys above the beach’s high-water mark, in typical tern nesting habitat.
The birder-UAV competition consisted of surveying and estimating the size (i.e. number of individuals) of each fake tern colony. Each colony contained between 463 and 1,017 decoys.
Experienced seabird counters using spotting scopes and binoculars stood 38 m (123 feet) from the fake birds, the distance shown to be as close as people can approach terns before they fly off.
They competed against a team of citizen scientists brought in to count birds in images taken from a standard compact digital camera mounted facing downward on a small, off-the-shelf quadcopter UAV. The UAV took photos at 30 meters (98 feet), 60, 90 and 120 meters above the beach. The researchers limited each flight to seven minutes.
The volunteers counted the numbers of birds in the UAV images in a standardized way, using a grid overlaid onto each photograph. These counters could zoom in and out of each grid cell to ensure they had counted all individuals. At least seven people, none of whom had knowledge of the experimental setup, counted the birds of each colony; most had never counted colony-living birds before.
Even passionate, experienced researchers get tired counting hundreds of individual animals in photos, which can lead to inaccurate tallies. So the researchers joined forces with university computer-vision scientists to develop software with an object-detection algorithm to semi-automatically detect and count the target animals in a given drone-captured photograph.
“We were interested in understanding if we could semi-automate the detection process without compromising data quality,” Hodgson said in an email to Mongabay-Wildtech.
The software “learns” from training data to distinguish birds in an image from the background. The team manually delineated a proportion of the birds in the photos and several areas of background without birds to “train” the algorithms. They used increasing amounts of training data until they reached an amount that would minimize the data-collection effort needed for the software to run on its own and make accurate identifications.
“This allowed us to investigate the relationship between the amount of training data required, and therefore user time, and accuracy,” Hodgson said. “We identified 10% training data as a threshold above which little improvement in count accuracy was achieved for our application.”
The researchers have released the code for their algorithm via open access so that others can use and adapt it for their applications. “As the algorithm is trained to detect the target animal from a set of delineations made by the researcher,” Hodgson said, “it can be easily applied to many different animals and habitats.”
Improving the efficiency of wildlife observations
The team compared the numbers of birds calculated by the ground birdwatchers, citizen scientist counters, and the computer software counter to the true number of fake birds put into each replica colony.
The bird counts from the UAV-derived images were between 43 and 96 percent more accurate than ground observer counts, depending on the sample height.
The counters in the field consistently counted fewer birds in each colony than there actually were, whereas the volunteers counting from the images did not over- or underestimate the true number of individual birds in each “colony.” The counts made by the on-site observers also varied more than either the manual or software-based counts of birds in the images taken by the UAV.
The automated counts that had used at least 10 percent of the data for training matched the counts made by the volunteers, with greater speed and efficiency and fewer person-hours needed.
“The ability to collect data with higher accuracy, higher precision, and less bias than the existing approach confirms that [UAVs] are a scientifically rigorous data collection tool for wildlife population monitoring,” the authors say in their paper. “This approach also produces a permanent record, providing the unique opportunity to error-check, and even recount with new detection methods, unlike ground count data.”
The quality of the image mattered. For the six colonies with high-quality aerial imagery—obtained when wind speed was slow and equipment behaved as expected, limiting vibration—the UAV-derived counts were between 92 and 98 percent more accurate than field-based ground counting.
“Consequently,” the authors wrote, “ecologists should determine the minimum required GSD”—ground sample distance, representing image resolution, or the number of pixels that represent the animal of interest—“for their context and optimise their sensor accordingly (e.g. resolution, focal length) relative to sample height.”
UAV-aided surveying and monitoring are best suited for species that can be seen from above, Hodgson said. “Our results have important implications for a range of species. We consider they are especially relevant to aggregating birds, including seabirds like albatrosses, surface nesting penguins and frigatebirds, as well as colonial nesting waterbirds like pelicans. Other types of animals that are easily seen from above, including hauled-out seals and dugongs, are highly suited to drone monitoring.”
Hodgson stressed the importance of making wildlife well-being the top priority.
“We are still learning about how wildlife react to the presence of drones, and more research is required to quantify these responses in a range of species and environments,” he said. “The results of such research will help to refine and improve drone monitoring protocols so that drones have minimal to non-existent impact on wildlife. This is particularly important for species that are prone to disturbance and where traditional methods involving close proximity to species are not possible or desirable.”