New Machine Learning Tool Detects Endangered Black Petrels in Fishing Lines 30 Times Faster than Humans
Dragonfly Data Science, a Wellington-based data science company, has developed a groundbreaking machine learning tool that can identify black petrels caught in fishing lines at a rate 30 times faster than humans. This innovative technology marks the first time machine learning has been employed to identify protected species trapped by fishing vessels.
Black petrels, also known as tākoketai, are native to New Zealand and are classified as nationally vulnerable, putting them at high risk of extinction. They are particularly susceptible to commercial fishing, often diving for bait as it is dropped into the ocean. If they become hooked, they are pulled under the water and drown.
Over the past two decades, the risk to black petrels from fisheries in New Zealand has decreased due to various mitigation measures such as bird scaring lines, night setting, and sinking lines to a depth that birds cannot detect. However, this year alone, eight black petrel captures have been reported by commercial fishing operators, with four birds found dead and the remaining four released unharmed. In some instances, the birds can survive if they go for leftover bait while the line is reeled in.
To address this ongoing issue, Dragonfly Data Science embarked on a research and development project in 2015 to determine if video footage could be utilized to detect deceased black petrels captured on long-line fishing hooks. By reviewing footage from 2019 and 2020, the company created a machine-learning algorithm that can quickly scan video clips and identify captured black petrels.
The tool is comparable in effectiveness to human reviewers, correctly identifying approximately 85% of seabird captures. However, its speed is a distinct advantage, capable of scanning one minute of footage every second, which is 30 times faster than human reviewers can process. The tool employs box-drawing technology to outline the detected petrels, assisting in visual recognition.
While the tool is still in the research and development phase, Dragonfly data scientist Henry Zwart believes it works best in conjunction with human reviewers to augment the results. However, he anticipates that in the future, it could potentially replace human reviewers altogether. Zwart acknowledges the challenging conditions the tool must contend with, including sunstroke, sea spray, salt, and fog that can obscure the lens, along with the fleeting nature of petrel captures. Despite these obstacles, the tool has achieved encouraging results.
The findings from a 2017 monitoring program have been instrumental in informing recent government initiatives aimed at recording commercial fishing activities. Approximately 300 vessels are currently being outfitted with cameras to provide independent and accurate data on fishing practices. While the fishing industry has largely embraced this endeavor, concerns exist regarding privacy, workplace filming, and access to the footage.
Biz Bell, the managing director of Wildlife Management International, has been monitoring black petrels on Aotea/Great Barrier Island for nearly three decades on behalf of the Department of Conservation and iwi Ngāti Rehua. He emphasizes the importance of gaining knowledge to protect and enhance the black petrel population, highlighting that the presence of these seabirds on the island surprises many visitors and describing his visits to the colony as the highlight of his year.
Dragonfly Data Science’s machine learning tool represents a significant breakthrough in addressing the threat to black petrels from commercial fishing. With its ability to swiftly identify captured birds, this technology has the potential to mitigate harm and contribute to the protection and preservation of this vulnerable species. As research and development continue, the hope is that advancements like this tool will pave the way for even greater conservation efforts in the future.
