AI is helping read the ocean’s oxygen history from fossil shells

A sediment core preserves more than layers of mud. It also contains the shells of seafloor-dwelling foraminifera: organisms whose different environmental tolerances help researchers infer how much oxygen their surroundings held. Reading that record requires a census of the kinds of shells left behind.[3]

Researchers automated the census off Peru with cropped shell photographs and a convolutional neural network. The best model reached 92% validation accuracy from a 5,860-image collection split 80:20 between training and validation; separate application images tested the resulting counts.[1]

Processing after training and calibration takes about 35 minutes per sample, mostly photography. The authors report manual picking and identification sometimes exceeding a day, without a paired timing test. Washing and sieving remain manual.[1]

A census becomes a chemical clue

The oxygen estimate comes from a second relationship, established by ocean science. Researchers compare modern shell assemblages with oxygen conditions at the places where they were collected, then use the resulting calibration to interpret fossil communities. A 2021 study calibrated three approaches using 45 surface-sediment samples spanning 0.03–2.88 millilitres of dissolved oxygen per litre. Assemblage composition and shell shape gave consistent reconstructions; a method based on one species’ porosity produced more ambiguous estimates.[3]

That distinction explains why recognising every photograph correctly is only part of the job. What matters for an oxygen reconstruction is how classification errors change the mixture of environmental indicators. Confusing two similarly tolerant species can matter less than moving specimens between groups associated with very different oxygen conditions.[3]

The oxygen reconstructions closely tracked one another (R² = 0.992). Their fitted slope was 0.935, near one for identical estimates. Both used the BFAex index; neither directly measured ancient seawater.[1]

Where the counts diverged

The public dataset contains paired counts for 31 samples from two cores covering approximately 18,000 years. Adding its 15 reported taxon groups gives 16,183 manually classified specimens and 18,321 automated classifications, a 13.2% larger total. Our calculation also exposes differences concealed by a single overall accuracy figure.[2]

Bolivina humilis remains the most numerous group in both tallies. At the other extreme, the automated count of Buliminella elegantissima includes 13 samples where the manual count was zero. An oxygen-history study and a search for the first appearance of a rare species would therefore place different demands on this same output.[2]

The authors attribute some excess counts to fragments or other particles. They recommend expert review and regional training; detailed changes through time await separate work.[1]

There is a substantive history to recover. In separate research published in 2020, Zeynep Erdem and colleagues used three deeper Peruvian-margin cores to estimate an approximately 30-micromole-per-kilogram oxygen decline between the Last Glacial Maximum and the Holocene at the oxygen minimum zone’s lower boundary. That earlier result concerns different cores and depths. Automating the census could make it practical to ask how such changes unfolded between widely spaced samples.[4]

Report an error or suggest a correction