Keeping coral reefs healthy with AI

With nearly a quarter of the world’s fish depending on coral reefs for shelter and food, the health of these so-called ocean rainforests is critical to the overall wellbeing of marine life.

The reefs also protect coastlines from tropical storms, providing food and income for one billion people, and generating $9.6bn a year in tourism and recreation. But reefs are being endangered by overfishing, bottom trawling, global warming and unsustainable coastal development.

For a while now, scientists have been monitoring the number and variety of fish around a reef to ascertain overall reef health. This involves sending human divers into the sea to capture video footage and photos of coral reefs, a dangerous and time-consuming undertaking that could also frighten fish into hiding.

In 2019, the Sulubaaï Environmental Foundation, a Filipino non-profit organisation dedicated to protecting marine life, started Project: Corail together with Accenture and Intel to use artificial intelligence (AI) to monitor, characterise and analyse coral reef resilience using an artificial reef.

Called Sulu-Reef Prosthesis (SRP), the artificial reef was placed in the reef surrounding the island of Pangatalan in the Philippines. Fragments of living coral were planted on it and will grow and expand, providing a hybrid habitat for fish and marine life.

Underwater video cameras, equipped with Intel chips and Accenture’s AI-powered video analytics capabilities, were then placed in strategic locations near the SRP to count and classify marine life as they pass by using a deep learning classification model.  

But before that, Accenture had to build a dataset of fish to train the model because few such datasets were available publicly, according to Lee Joon Seong, the firm’s managing director and applied intelligence lead for Southeast Asia.

To do so, it developed a computer vision algorithm capable of detecting objects moving in front of the camera. Each time a big enough object is detected, the image is saved, sent to a cloud repository and subsequently annotated together with scientists at the Sulubaaï foundation. Since May 2019, Accenture has collected more than 67,000 images. 

To train its deep learning classification model, Accenture used TensorFlow to perform transfer learning on a ResNet-101 convolutional neural network with a public dataset enriched with the dataset it had built.

Lee said the model can detect fish with an accuracy of about 93% during lab tests. But as Pangatalan island is situated in a remote location, his team has faced issues with power and internet access, which hindered proper accuracy testing.

“The accuracy will also depend on the quality of the water on that day, and we will need to collect further images and re-train our deep learning classification model to enhance the accuracy,” Lee told Computer Weekly.

Engineers from Accenture and Intel are already working on the next prototype, which will feature an optimised convolutional neural network and a backup power supply. Infrared cameras could also be used to capture videos at night to provide a complete picture of the coral ecosystem.

Beyond monitoring indicators of coral reef health, Accenture said the technology could be used to track the migration of tropical fish to cooler climates, or to monitor intrusions in protected or restricted underwater areas.

Asked if it can be used to identify new marine species and those that are thought to have been extinct or close to extinction, Lee said the chances are slim.

“It is unlikely as they are very rare, and we will be very lucky to have one swimming in front of the camera,” he said. “But it is not impossible – we have detected rare marine life, such as the hawksbill sea turtle.

“The deep learning algorithm is unlikely to be able to classify it, but the motion detection capabilities would act as a ‘camera trap’ and allow scientists to review the images once an unidentified creature appears.”