Can machine learning help save the whales? How PNW researchers use tech tools to monitor orcas
Being an orca isn’t easy. Despite a lack of natural predators, these amazing mammals face many serious threats – most of them brought about by their human neighbors. Understanding the pressures we put on killer whale populations is critical to the environmental policy decisions that will hopefully contribute to their ongoing survival.
Fortunately, marine mammal researchers like Holly Fearnbach of Sealife Response + Rehab + Research (SR3) and John Durban of Oregon State University are working hard to regularly monitor the condition of the Salish Sea’s southern resident killer whale population (SKRW). Identified as J pod, K pod and L pod, these orca communities have migrated through the Salish Sea for millennia. Unfortunately, in recent years their numbers have dwindled to only 75 whales, with one new calf born in 2021. This is the lowest population figure for the SRKW in 30 years.
For more than a decade, Fearnbach and Durban have flown photographic surveys to capture aerial images of the orcas. Starting in 2008, image surveys were performed using manned helicopter flights. Then beginning in 2014, the team transitioned to unmanned drones.
As the remote-controlled drone flies 100 feet or more above the whales, images are captured of each of the pod members, either individually or in groups. Since the drone is also equipped with a laser altimeter, the exact distance is known making calculations of the whale’s dimensions very accurate. The images are then analyzed in what’s called a “photogrammetric health assessment.” This assessment helps determine each whale’s physical condition, including any evidence of pregnancy or significant weight loss due to malnourishment.
“As a research tool, the drone is very cost effective and it allows us to do our research very noninvasively,” Fearnbach said. “When we do detect health declines in individuals, we’re able to provide management agencies with these quantitative health metrics.”
But while the image collection stage is relatively inexpensive, processing the data has been costly and time-consuming. Each flight can capture 2,000 images with tens of thousands of images captured for each survey. Following the drone work, it typically takes about six months to manually complete the analysis on each season’s batch of images.
Obviously, half a year is a very long time if you’re starving or pregnant, which is one reason why SR3’s new partnership with Vulcan is so important. Working together, the organizations developed a new approach to process the data more rapidly. The Aquatic Mammal Photogrammetry Tool (AMPT) uses machine learning and an end-user tool to accelerate the laborious process, dramatically shortening the time needed to analyze, identify and categorize all of the images.
Applying machine learning techniques to the problem has already yielded huge results, reducing a six-month process to just six weeks with room for further improvements. Machine learning is a branch of computing that can improve its performance through experience and use of data. The faster turnaround time will make it possible “to more quickly identify whales of concern and provide health metrics to management groups to allow for adaptive decision making,” according to Vulcan.
