Conservationists have launched a new AI tool that can sift through petabytes of underwater imaging from anywhere in the world to identify signs of abandoned or lost fishing nets—so-called ghost nets.
Each year, around 2% of the world’s fishing gear—including roughly 80,000 square kilometers of fishing nets—is lost in the oceans. Those nets threaten marine wildlife like seals, turtles, and dolphins, which get tangled in the abandoned nets and drown. Lost fishing gear also decomposes in water and becomes a major contributor to plastic pollution.
Detecting abandoned fishing nets resting on the ocean or sea floor is extremely difficult. Nets are often as thin as a finger and are all but invisible to humans looking at underwater images.
To identify ghost nets’ exact locations so they can be removed, WWF Germany and partners Accenture and Microsoft’s AI for Good Lab created GhostNetZero.ai. The online platform crowdsources high-resolution underwater data—called side scan sonar images—from research institutes, governments, offshore wind-power companies, and other groups that routinely gather such scans.
A Convolutional Neural Network (CNN), made up of DeepLabV3 with a ResNet50 backbone, is embedded into GhostNetZero.ai to scan the sonar data for signs of ghost nets. It’s accurate in identifying ghost nets in solar scans 94% of the time.
“Every side scan sonar image includes geolocation and other kinds of metadata, so if the AI recognizes a ghost net in a sonar image, it can tell us the ghost net’s exact location,” said Gabriele Dederer, WWF Germany’s ghost nets project manager.
Catching ghost nets
Once AI identifies a ghost net, Dederer and her team work with local divers and fishermen to figure out how to remove those nets. The team asks local partners to verify that the nets are where AI believes them to be, and to provide more granular details such as the nets’ approximate sizes.
From there, it’s up to local maritime and conservation groups to set in motion the logistical process to remove the ghost nets.
“Divers help us determine whether these are very big nets or smaller nets, and they can help judge how big a retrieval ship is needed,” Dederer said. “Our work has to be viewed region-by-region, and if there’s already a local protocol, we’ll work within that because all this retrieving is a huge logistical and financial task.”
WWF Germany currently is working closely with local teams in France, Estonia, and Sweden, and it plans to expand its partnerships in future.
Dederer came up with the idea for analyzing crowdsourced sonar data several years ago. But it was only earlier this year, after Accenture and Microsoft learned about her efforts and offered to help, that AI was integrated into the project.
The CNN is hosted on Microsoft Azure cloud, and was developed using PyTorch libraries. It uses NVIDIA A100 TensorCore GPUs for training and inference.
Christian Bucher, who worked as Microsoft’s liaison to the ghost net project, noted that another upside of working with local teams is that their on-the-ground verifications can be fed back into the model to improve its accuracy and efficiency.
“AI is very much about pattern recognition, and we used PyTorch to help detect the different types of ghost nets because sonar images [of ghost nets] are almost always different,” Bucher said. “Utlimately, this is a relatively simple machine learning task and once we figured out how to do the segmentation, we reached an accuracy of 94%.”
Read more about the ghost net effort in additional news coverage.
