Norm Van Vactor stands on the bank of the Wood River facing a mysterious looking grey box with blinking lights to start the morning.
“So that's the habitat and the home for our ‘bird’, and it's a completely self-contained unit,” he said, pointing to the box.
At exactly 8:30am, the box opens and a drone emerges. It quickly takes off on an automated flight path, flying to an observation position above the right bank of the Wood River. It hovers there for a few minutes, with its camera fixed on the river bed.
The goal is to feed the images into a computer model to generate more accurate salmon escapement counts. For over 70 years, the Alaska Department of Fish and Game has used counting towers in Bristol Bay to track salmon escapement. That means techs hand count each fish that swims by. But new technology is on its way to the Bay, Van Vactor is a project lead.
“I feel strongly that there are some tools out there, like this, that could significantly enhance the quality of our data collection that helps manage this fishery,” he said. Van Vactor noted that this project wouldn’t have been possible just a few years ago.
“We currently have this amazing convergence of technology,” he said. “This drone in the box technology combined with machine learning and AI combined with things like Starlink Mini combined with lithium batteries and solar panels–it's all happening just in the last year or two, and it's all coming together to make something like this possible.”
Van Vactor’s remote control displays a livestream from the drone’s camera that shows hundreds of salmon. Each fish is making its way upstream to spawn, slowly and steadily swimming through the sunlit, shallow water.
“I've been engaged and involved in the commercial fishery here for 40 years,” he said. “These last two weeks to me have just been phenomenal, because it's given me a perspective that I've never been part of–Mother Nature doing its thing.”
The project’s goal is to explore technology that could improve salmon escapement counts, which help inform the management of the Bristol Bay fishery. Most escapement counts across the Bay currently come from nine Fish & Game counting tower sites.
Van Vactor’s drones are just downstream of the Wood River towers – scaffolded towers on each side of the river tethered by ratchet straps. There, Fish & Game tower crews rotate through continuous shifts over 24-hours, counting each salmon that swims by with a hand-clicker in 10 minute intervals. This is how the department has been tracking salmon escapement since 1955, to ensure millions of sockeye return to spawn future generations.
Just upstream, the University of Washington’s Fisheries Research Institute has a lab on the shore of Lake Aleknagik. It’s one of the partners supporting this research, alongside Bristol Bay Regional Seafood Development Association and Bristol Bay Economic Development Corporation. At this stage, Fish & Game isn’t directly involved in the project. Joel Tibbs is one of the project’s technicians in the lab.
“Tower counts have been successful for so long, and they've been an implemented technology that's really worked for the industry for decades now,” said Tibbs. “But understanding how we might be able to step into easier and more accurate methods for measuring fish is a really important part of management.”
Tibbs pulls up the footage from that morning’s drone flight on his computer. There’s a bold yellow line across the river on the screen and his task this morning is to press a key on his computer each time a fish swims past that yellow line. But this hand counting is only temporary. By manually identifying the fish, Tibbs and his fellow technicians are training an AI-powered model.
“That is data hopefully informing the model, like, oh, that was a fish, because something crossed the line, and I pressed F at that time,” he explained. “The model will then hopefully be able to latch on to that and count fish more accurately.
Another member of the team is Ian Chiu, who’s working remotely to develop the machine learning side of the project.
“The end product is that we'll be able to feed in a video, and then we'll have a computer vision model that will be trained on picking out salmon,” he said.
Chiu says that they’ll likely need several thousand drone images for the model to learn how to reliably identify a fish, and multiple rounds of training and corrections.
“It might need a lot more training for very specific use cases,” he said. “For example, if there's a shadow along the river, or if the conditions are poor.”
The rapid expansion of AI is controversial. But Chiu believes this project is an important application of the technology.
“AI, especially for environmental sciences, is kind of counterintuitive sometimes because AI takes up so many resources,” he said. “But I think there are certain applications where AI does really help, and I think this is one of them.”
Back on the bank of the Wood River, Norm Van Vactor agreed with that sentiment.
“I really think it's our future. I really do,” he said. “When you think about it, we could be sitting at the fish and game cabin a couple 100 feet up the river from here, we wouldn't be out in the bugs and mosquitoes, and we would be monitoring a machine like this.”
The crew will test their models against Fish & Game tower counting data in the project’s next phase.
Get in touch with the author at jessie@kdlg.org
