In partnership with

Coral Reefs Health Monitoring

Mapping hard and soft coral in underwater imagery to track reef health over time.

25% of marine life

<1% of the seafloor

open-source models

Source code Try the demo Computer VisionMachine Learning Live

Coral reefs are among the richest ecosystems on Earth — and among the most threatened. Marine biologists track their health by photographing the seabed on research dives, but turning thousands of those images into numbers is slow, painstaking work, and the reporting lag blunts conservation’s ability to respond in time.

Together with ReefSupport, we built a computer-vision pipeline that maps coral in benthic imagery — outlining and classifying each colony automatically — so researchers can measure how coral cover is growing or declining across a protected area, far faster than by hand.

From a dive photo to coral cover — capture, segment each colony, classify hard vs soft, and measure coral cover over time A diver photographs the reef; the model outlines and classifies every coral colony, and the result becomes a coral-cover figure that can be tracked over time.

Our tools tap on AI and computer vision for increasing the capabilities of coral reef and marine monitoring in examining benthic/seabed features.

– ReefSupport

Why coral reefs matter

Reefs cover a sliver of the ocean yet underpin a huge share of marine life — and the livelihoods of millions of people.

Biodiversity hotspots

Less than 1% of the seafloor, yet home to around 25% of all marine species — habitat, nursery and shelter for countless fish and invertebrates.

Coastal & economic backbone

Reefs sustain the fisheries and tourism that millions rely on, and act as natural breakwaters that shield coastlines from storms and erosion.

Climate, medicine & culture

They help cycle carbon, yield compounds with real medical promise, and hold deep cultural meaning for coastal and Indigenous communities.

Reefs under pressure

Coral reefs are squeezed by many forces at once — most of them driven, directly or indirectly, by people. Tap each to learn more.

Rising sea temperatures make corals expel their algae and bleach — turning white, weakening, and falling prey to disease. Acidification weakens their skeletons too.

Runoff, sewage and debris smother corals and feed the algae that compete with them for light and space.

Removing grazing fish lets algae overgrow corals, and destructive methods like blast and cyanide fishing damage reefs directly.

Dredging, construction and land clearing bury reefs in sediment and nutrient-laden runoff.

Anchoring, trampling and over-visiting fragile sites — and harvesting for souvenirs — wear reefs down.

Non-native species outcompete or prey on reef life and reshape the habitat.

Limited awareness and patchy enforcement let reef degradation continue unchecked.

Mapping coral, colony by colony

The system is built around instance segmentation: rather than just labelling a photo, it finds each individual coral colony, traces its exact outline, and classifies it. We started with the key distinction — hard versus soft coral — with the framework designed to grow into finer functional groups and to adapt to reefs anywhere in the world.

How the model maps a reef — a benthic photo runs through the segmentation model, which outlines and labels every coral colony Instance segmentation traces the exact outline of each colony — hard coral in teal, soft coral in orange — rather than just drawing boxes.

The models are open source and come in a range of sizes that trade speed against accuracy, so a survey team can pick the right balance for their hardware and their reef.

The open-source instance-segmentation model outlining hard and soft corals on real underwater imagery The open-source model segmenting hard and soft corals on real benthic imagery.

Why underwater vision is hard

Reading a reef from a photo is far harder than it looks — water itself works against the camera.

Light & colour loss

Water absorbs and bends light, so underwater images lose contrast and colour and grow hazier the further away the subject is.

Visual noise

Caustics, backscatter and drifting particles add artifacts, and busy reef backgrounds blur the line between a coral and its surroundings.

Variable & scarce data

Conditions shift constantly and gear isn't standardised — and well-labelled underwater datasets are rare, which makes models hard to train and to generalise.

Models are trained on reef imagery from around the world to cope with this variety:

Coral reefs photographed in different regions across the world.

Conclusion

Automated benthic analysis turns a reporting bottleneck into fast, repeatable measurement — quantifying the long-term growth or decline of coral cover, and giving reef managers the timely picture they need to act.

The benthic imagery analysis system in action — courtesy of ReefSupport.