With coral reefs in fast decline throughout the globe, researchers from the College of Hawaii at Mānoa have pioneered an AI-based surveying device that displays reef well being from the sky.
Utilizing deep studying fashions and high-resolution satellite tv for pc imagery powered by NVIDIA GPUs, the researchers have developed a brand new methodology for recognizing and monitoring coral reef halos — distinctive rings of barren sand encircling reefs.
The research, just lately printed within the Distant Sensing of Setting journal, might unlock real-time coral reef monitoring and switch the tide on international conservation.
“Coral reef halos are a possible proxy for ecosystem well being,” mentioned Amelia Meier, a postdoctoral fellow on the College of Hawaii and co-author of the research. “Seen from area, these halo patterns give scientists and conservationists a novel alternative to watch huge and distant areas. With AI, we are able to recurrently assess halo presence and dimension in close to actual time to find out ecosystem well-being.”
Sea-ing Clearly: Illuminating Reef Well being
Beforehand attributed solely to fish grazing, reef halos may also point out a wholesome predator-prey ecosystem, in accordance with researchers’ current discoveries. Whereas some herbivorous fish graze algae or seagrass close to the protecting reef perimeter, hunters dig across the seafloor for burrowed invertebrates, laying naked the encircling sand.
These dynamics point out the world hosts a wholesome meals buffet for sustaining a various inhabitants of ocean dwellers. When the halo modifications form, it alerts an imbalance within the marine meals internet and will point out an unhealthy reef atmosphere.
In Sizzling Water
Whereas making up lower than 1% of the ocean, coral reefs supply habitat, meals and nursery grounds for over 1 million aquatic species. There’s additionally enormous industrial worth — about $375 billion yearly in industrial and subsistence fishing, tourism and coastal storm safety, and offering antiviral compounds for drug discovery analysis.
Nonetheless, reef well being is threatened by overfishing, nutrient contamination and ocean acidification. Intensifying local weather change — together with the ensuing thermal stress from a warming ocean — additionally will increase coral bleaching and infectious illness.
Over half of the world’s coral reefs are already misplaced or badly broken, and scientists predict that by 2050 all reefs will face threats, with many in vital hazard.
Charting New Horizons With AI
Recognizing modifications in reef halos is vital to international conservation efforts. Nonetheless, monitoring these modifications is labor- and time-intensive, limiting the variety of surveys that researchers can carry out yearly. Entry to reefs in distant places additionally poses challenges.
The researchers created an AI device that identifies and measures reef halos from international satellites, giving conservationists a chance to proactively tackle reef degradation.
Utilizing Planet SkySat pictures, they developed a dual-model framework using two kinds of convolutional neural networks (CNNs). Counting on pc imaginative and prescient strategies for picture segmentation, they educated a Masks R-CNN mannequin that detects the perimeters of the reef and halo, pixel by pixel. A U-Internet mannequin educated to distinguish between the coral reef and halo then classifies and predicts the areas of each.
The crew used TensorFlow, Keras and PyTorch libraries for coaching and testing hundreds of annotations on the coral reef fashions.
To deal with the duty’s massive compute necessities, the CNNs function on an NVIDIA RTX A6000 GPU, boosted by a cuDNN-accelerated PyTorch framework. The researchers acquired the A6000 GPU as individuals within the NVIDIA Tutorial {Hardware} Grant Program.
The AI device rapidly identifies and measures round 300 halos throughout 100 sq. kilometers in about two minutes. The identical activity takes a human annotator roughly 10 hours. The mannequin additionally reaches about 90% accuracy relying on location and might navigate varied and complex halo patterns.
“Our research marks the primary occasion of coaching AI on reef halo patterns, versus extra widespread AI datasets of pictures, equivalent to these of cats and canine,” Meier mentioned. “Processing hundreds of pictures can take lots of time, however utilizing the NVIDIA GPU sped up the method considerably.”
One problem is that picture decision generally is a limiting issue within the mannequin’s accuracy. Course-scale imagery with low resolutions makes it troublesome to identify reef and halo boundaries and creates much less correct predictions.
Shoring Up Environmental Monitoring
“Our long-term aim is to rework our findings into a strong monitoring device for assessing modifications in halo dimension and to attract correlations to the inhabitants dynamics of predators and herbivores within the space,” Meier mentioned.
With this new method, the researchers are exploring the connection between species composition, reef well being, and halo presence and dimension. At present, they’re wanting into the affiliation between sharks and halos. If their hypothesized predator-prey-halo interplay proves true, the crew anticipates estimating shark abundance from area.
