Using AI to map the seafloor
Watersurface in the Marshall Islands. Photo: Reinhard Dirscherl/ullstein bild via Getty Images
We have a more accurate map of the surface of Mars than we do for Earth’s ocean floor. Right now, researchers have a blurry, indirect picture of the seabed from satellite imaging, some sonar data and samples collected from ships.
Yes, but: There's an avalanche of data about the chemical, physical and biological properties of the ocean, and scientists are beginning to use machine learning to tie that data to limited scientific records in hopes of piecing together a better picture.
Why it matters: The ocean is the planet's largest carbon sink thanks to processes that critically depend on the seafloor, and sediment layers hold a record of the ocean's chemical and climatic history. Understanding the ocean floor is key for climate forecasts and predicting geohazards like tsunamis as well as for efforts to mine the seas for methane and minerals.
"The seafloor is one of the more remote places on the planet. It is hard to get a lot of data," says Warren Wood, a geophysicist at the U.S. Naval Research Laboratory in John C. Stennis Space Center in Mississippi. Instead, researchers want to find proxies for what's happening in the deep ocean. "If you can have some idea of what is out there without going out and measuring it, it is a phenomenal achievement."
What's down there
A census: A few years ago, Adriana Dutkiewicz and Dietmar Müller at the University of Sydney used machine learning algorithms to take a census of the sediments on the seafloor. Previous maps were drawn by hand in the 1970s and depicted "a barren seascape," Müller says. "Our new map illustrates an enormous diversity of seafloor geology."
- For example, the corpses of diatoms — plankton that produce about a quarter of the oxygen we breathe — fall to the deep ocean in a so-called "marine snow" that traps carbon on the seafloor.
- With their map, the researchers found diatom accumulations on the seafloor don’t directly correspond to blooms on the surface that can be seen in satellite images. "This disconnect demonstrates that we understand the carbon source, but not the sink," Müller says
Tying it to the top: They then analyzed 14,400 deep sea sediment samples taken around the world and found a link between the types of sediments that collect on the seafloor and properties on the surface of the water, like salinity and temperature, as well as the seabed's terrain. Knowing that the ocean floor is a reflection of what is happening at the surface, they want to use the surface to learn about the ocean basin.
"The sea surface is more accessible," Wood says. "Satellites are taking pictures of it all the time. If we can use that to understand things about the sea floor, that is impressive."
The nitty gritty: Wood and his collaborators are using machine learning to try to predict properties like porosity (how much solid and liquid there is in a sediment, which heavily influences its strength), total organic carbon content (which indicates where methane gas is being generated) and the flow of heat in sediments (which is an important determinant of the speed of sound underwater and therefore to sonar technologies).
Looking back: Most recently, the team in Sydney looked at how the depth of sediments on the ocean floor changed over time, especially in the last 25 million years as continents collided, and mountains eroded and were shed into oceans by rivers as they formed.
- They found the depth has changed by more than 200 meters over the past 150 million years.
- That contributed to a rise in sea level that was offset by the oceans aging and deepening during the same time period, so in the long run, sea level actually fell, Müller says. "Understanding the evolution of the ocean basins through geological time is important for long-term predictions of the fate of our planet."
Yes, there's more data and more computing power. But there is also increased interest from the private sector to apply machine learning to geology, especially in mining, says Lorenzo Perozzi, who points to a collaboration between IBM Watson and the Canadian mining company Goldcorp. Perozzi, a data scientist at Geolearn and postdoctoral researcher at the Institut National de la Recherche Scientifique in Québec City, is working on applying machine learning algorithms to predict rockburst in underground mines.
"Machine learning can be used to predict properties where we don't have a measurement," Wood says. "We're not quite to the place where we can pin an important geological discovery on machine learning. I’m confident we’ll get there but right now we're using it to better understand the processes and properties of the seafloor."