GreenTalk#5 - The Coral Vision

Summary

Professor Philip Heller presents Coral Vision, a set of projects that use bioinformatics and machine learning to study marine ecosystems and how they may respond to climate change. He emphasizes the ocean’s importance to life, including its role in the global oxygen cycle and its ecological and cultural value.

Sponges and reef ecosystems

Sponges filter seawater and can process dissolved organic nutrients that many other organisms cannot use directly. They convert some of this material into particles that become available to the wider community. Sponges also pump and strain large volumes of water. In a video demonstration, fluorescent dye injected near sponges was rapidly drawn through them and expelled.

Sponge skeletons contain spicules. Some are made of calcium carbonate, while others are made of silica, a glass-like material.

Coral reefs support high biodiversity, shelter juvenile fish, and provide food and other benefits to people. They also host important biogeochemical processes, including nitrogen fixation. However, ocean warming can cause corals to lose their photosynthetic symbionts, leading to bleaching. Ocean acidification, caused when carbon dioxide reacts with seawater, threatens calcium-based coral and shell structures.

Because silica spicules may withstand acidification better than calcium-based structures, Heller proposes investigating whether some sponge reefs could persist or become more prominent as conditions change. The best roles for sponge propagation in reef management remain open research questions. In response to a question, Heller noted that sponges may be less brittle than branching corals and may resist some stresses differently. He suggested that reefs with both corals and sponges could provide the best combination of habitat and ecological functions.

Studying ancient sponge communities

Sediment cores from the Bering Sea preserve evidence of past marine communities. Heller describes a period about three million years ago when the region had high productivity of glass sponges.

The conventional approach—visually identifying a small amount of material under a microscope—can miss much of a core sample. It is also difficult to identify sponge species from individual spicule shapes, and some spicules may have broken during burial.

Heller’s team is developing Spiceter, a workflow that uses a FlowCam to photograph particles in a sample, then applies deep learning to classify spicule shapes and estimate their proportions. Comparing sponge assemblages from different core layers with past ocean temperatures may help identify which kinds of sponges did well during warmer periods. This work relies on the assumption that sponge spicule shapes have evolved relatively slowly.

Coral Vision and reef monitoring

NOAA’s Autonomous Reef Monitoring Structures (ARMS) are plate-like structures placed on the seafloor and retrieved after about 18 months. The organisms that colonize them provide a snapshot of reef communities.

NOAA’s Reef Assessment and Monitoring Program deployed ARMS across Pacific sites, including Hawaii, the Line Islands, Samoa, and the Marianas, to help establish baselines for relatively undisturbed reefs. The project produced thousands of images that would be laborious for specialists to identify manually, so Coral Vision uses computer vision to analyze them.

The image-analysis process involves:

• Segmentation: locating and outlining organisms or other features in an image.
• Identification: classifying the segmented objects.
• Quantification: calculating measurements such as organism abundance or plate coverage.

An off-the-shelf Mask R-CNN model struggled with underwater images, which often contain irregular, overlapping, or visually ambiguous organisms. The team is using transfer learning to adapt a pretrained model to recognize specific reef organisms and features, using manually labeled images.

Current target applications include juvenile corals, sponges, crustose coralline algae, overall plate coverage, and plate boundaries or bolt holes. The algae are of interest because they secrete material that helps bind reefs together.

A separate coral-restoration project grows juvenile corals on plugs made from coral-skeleton-like material. Researchers gradually increase temperature and acidity to select corals that tolerate projected future conditions, then transplant the survivors onto reefs.

Heller demonstrated an interface that displays model-detected objects according to a confidence threshold. Undergraduate and graduate students are helping build and improve the models.

Researchers and sources featured

• Philip Heller — Speaker and Coral Vision project lead
• Rusty Brainard — NOAA coral-reef researcher associated with the ARMS work
• Molly Timmers — Scientist who worked with Brainard on ARMS
• NOAA — National Oceanic and Atmospheric Administration, including its Reef Assessment and Monitoring Program
• Hawaii Institute of Marine Biology — Group identified as conducting the juvenile-coral restoration work
• FlowCam — Used to photograph particles for the sponge-spicule project
• Mask R-CNN — Pretrained image-segmentation model discussed in the machine-learning section
• Barbara — Event host
• Kevin Moore — Lecturer who asked questions during the discussion