A small team of marine biologists is working on a project that would have seemed slightly ridiculous ten years ago, somewhere along the coast in La Jolla, in a building that smells slightly of warm electronics and salt. From a desk, they are observing coral reefs. Instead of using a livestream or a porthole, millions of underwater photos were used to create dense, multi-layered, three-dimensional reconstructions. This idea has been pursued for years by Scripps’ Sandin Lab, and it’s possible that technology has finally caught up to the ambition.
It’s difficult to ignore the numbers. Each dive returns about 350 gigabytes of photos. Three to four hundred dives are recorded annually by the team. A typical university computing cluster would be bent sideways by that type of data load, and the lab was bending for a considerable amount of time. Everyone in the building was aware of the environmental irony of using powerful servers to study a warming ocean, the slow processing, and the uncomfortable energy bills.
The calculations appear to have been altered by a recent in-kind donation from Dell Technologies. AI is used in their pilot project, Concept Astro, to plan demanding computing jobs for when the grid is cheapest and cleanest. On paper, it sounds boring. In actuality, the Sandin Lab is reporting a roughly twofold increase in image throughput, a twenty percent decrease in energy costs, and a thirty-two percent decrease in emissions. Speaking with professionals in this field gives me the impression that minor infrastructure changes like this are more important than the larger announcements.
The lab’s director, Stuart Sandin, has discussed the work with a level of measured optimism that is uncommon in coral science at the moment. Reefs sustain damage. Marine heatwaves, pollution, overfishing, etc. However, his team continues to uncover a more subdued discovery that doesn’t always garner media attention. Certain reefs are stable. Some are even making a comeback; these are typically those that are close to communities that have taken significant responsibility for their own waters. Without reliable data, it is more difficult to tell that story, which is one reason the new infrastructure is important.
Scripps contributed to the development of CoralNet, an open-source machine learning platform that contains the other half of the puzzle. In order to train the algorithm to identify coral, algae, and the peculiar grey textures of dying reef floor, the Living Oceans Foundation has been providing it with expertly annotated images from locations like Fiji and the Great Barrier Reef. According to preliminary testing, CoralNet achieves accuracy of over 90%, which is about what a skilled human annotator can do. However, the machine completes the task approximately a thousand times more quickly. The speed at which that alters the economics of reef monitoring is difficult to ignore.
Additionally, the digital reefs can now be seen at scale in the Eliza and Stuart Stedman space, a new visualization lab. Members of the lab, such as McClaran Shirley and Nicole Pederson, have been seen pointing at coral colonies in front of these wall-sized replicas, much like an architect might stroll through a model city. The paint is hardly dry, but it already feels lived in.
It’s still unclear if all of this represents a true turning point in coral science. AI cannot reduce ocean temperatures, and servers cannot save reefs. However, compared to a year ago, those performing this work can now see more, more quickly, and with less energy waste. That is not insignificant. Sometimes the most significant improvements are the quiet ones.
