Edge models provide researchers with new insight into marine ecosystems

Complete Post-Analysis Complemented by Real-Time Updates on the Most Important Data

For several years, Jonas Hentati Sundberg has used a Norwegian-built sailing drone to conduct his research. The drone autonomously navigates the waters between Öland and Gotland, collecting sonar data, among other things, which is later analyzed.

But when the drone was deployed this spring, it was equipped with a new combination of hardware and software: a Raspberry Pi and self-developed models for processing and analysis.

– With edge models, we filter the collected data in real-time. The most interesting data is sent directly to shore every ten minutes, via satellite link. The rest is analyzed later when the drone comes ashore and is emptied of data, says Joakim Eriksson, AI Developer at AI Sweden.

This isn’t the first time that Joakim Eriksson and Jonas Hentati Sundberg have worked together. Two and a half years ago, they developed models that automatically count nesting guillemots on Stora Karlsö. Last winter, a chance conversation revealed the potential for a new collaboration – on drones. In the drone project, solutions similar to those for counting guillemots will be used to eliminate time-consuming manual work.

– But we aren’t just limiting manual work, we’re also reducing the risk for human error, which makes our research results more robust, says Jonas Hentati Sundberg.

The new technology will allow researchers to study the ecosystem in a new way: in real-time for the most critical events, and historically for a more complete understanding.

– What we want to know in real-time could be data indicating an ecosystem collapse or unexpected new patterns that we haven't seen before. This would allow us to react immediately, such as by activating additional sensors to gather more information, says Jonas Hentati Sundberg.

The comprehensive, complete analysis funded by the Arctic Research Foundation and Stockholm Resilience Centre will be supported by artificial intelligence.

The AI models currently used by Joakim Eriksson and his colleague Sofia Nguyen can, for instance, differentiate between the seabed, waves, and fish. "This allows us to already calculate the fish mass under the drone.‘’

"The next step is to develop algorithms that can identify individual fish and perhaps even determine their species," says Sofia Nguyen, AI developer at AI Sweden.

The Goal: Robust, Generalizable, and Possibly Commercialized

Joakim Eriksson and co-worker Sofia Nguyen handle the AI components in the drone project. They note that the current drone has its limitations:
"For instance, we need to conserve electricity. In the long run, we would like to develop our own drone to optimize the hardware conditions necessary to run more advanced AI models onboard," he says.

Another way forward is to install AI models on SLU's research vessel Svea, where the sonar systems are more advanced, and the connectivity and electricity supply are better.
"Perhaps we could even install the solution on commercial ships and use them as a sensor network," says Jonas Hentati Sundberg.

The drone project has attracted international interest. A trip to Coats Island in Hudson Bay, Canada, is booked in the weeks following Midsummer, where the technology will be tested. A similar collaboration is also in the works with a research group in Australia. The goal is to take the next step in development together with them.

"We want to make our solution more robust and generalize it so it works on more hardware platforms and for more ecosystems. Eventually, we might even commercialize it," says Jonas Hentati Sundberg.