Graduate and undergraduate student members of the NRC–University of Victoria research team presented their project at the first-year information fair.
The ocean is one of the planet's most powerful natural systems for carbon sequestration. Among the most effective agents of this process are blue carbon ecosystems—coastal habitats such as mangrove forests, seagrass meadows and salt marshes. These rich environments capture and store vast amounts of carbon through plant growth and sediment accumulation. However, unlike many terrestrial ecosystems that cycle carbon back into the atmosphere, a significant portion of the carbon our oceans sequester remains locked away in underwater soils for centuries, or more. Preserving and restoring these ecosystems is essential to strengthening the ocean's role in climate regulation and Canada's goal of achieving net-zero carbon emissions by 2050.
Supported by the National Research Council of Canada's (NRC) Ocean program, a research team from the University of Victoria has partnered with the NRC to monitor ocean health and its capacity for carbon sequestration using a novel suite of sensors, machine learning and a remotely operated vehicle (ROV). The project also includes a cybersecurity component to assess vulnerabilities in the ROV's sensor network to protect data privacy.
Dr. Mohammad Mamun, a research officer at the NRC, explains the project "leverages interdisciplinary expertise in artificial intelligence for developing models that combine data confidentiality, cybersecurity and underwater systems to create an integrated framework for monitoring blue carbon ecosystems and advancing ocean health intelligence." At its core, it's a transformer-based machine learning model pre-trained on historical publicly available data from Ocean Networks Canada to predict chlorophyll concentrations—a key indicator of phytoplankton activity and of the potential for carbon sequestration. High levels of chlorophyll means a high potential for carbon absorption. This model helps researchers assess the health of marine ecosystems and their capacity for carbon uptake.
The beauty of this project is that these kinds of models are created using federated learning, a technique for training AI models without transferring all the data to a central location. This is an important distinction because not all organizations are willing to share their data, though many are open to using it for AI training if it remains protected. With this technique, instead of transmitting data to a centralized AI system, the AI model is sent to the devices where the data reside—in this case, a suite of sensors. The model learns in situ from the local data. As a result, only the learned insights are brought back, keeping the raw data private while allowing the AI to continuously improve.
This approach becomes especially valuable in complex environments like underwater systems. Many of us are familiar with the Internet of Things (IoT), where sensors connect to the cloud, but the Internet of Underwater Things (IoUT) introduces an added layer of complexity. In this underwater context, the sensor suite itself houses the predictive AI model, which is composed of both hardware and software. These sensors are connected to a ROV, which in turn communicates with the cloud. Each added connection point—sensor to ROV, ROV to cloud—increases the risk of cyber attack, making the privacy-preserving nature of federated learning not just a benefit, but a necessity.
"We wanted to understand ocean health, which is closely linked to carbon sequestration—a process that a healthy ocean performs effectively," adds Dr. Navneet Kaur Popli, an associate teaching professor of electrical and computer engineering at the University of Victoria in British Columbia. She was responsible for developing the sensor suite, the ROV that housed the sensors, and the AI algorithms used for federated learning.
Ocean-going drones for predicting changes in ocean health
The researchers used the publicly available data from Ocean Networks Canada collected at 4r key regions in the Salish Sea, a marginal sea of the Pacific Ocean located in British Columbia. Using this data, researchers applied machine learning techniques to predict chlorophyll levels. Then, by analyzing patterns in environmental factors such as rising water temperatures, ocean acidification and increased salinity, the team developed models that can accurately forecast changes in chlorophyll concentration—and thus, ecosystem productivity—30 days, 1 year and 2 years in the future. These predictions are essential for understanding how mangrove-like ecosystems, seagrass meadows and salt marshes in coastal British Columbia may respond to climate change, and how their ability to sequester carbon might shift over time.
Access to this information can help create a competitive advantage for some industries, scanning optimal locations for seaweed aquaculture or eco-tourism. By using federated learning, these sensors help keep data private, which is important for companies that want to keep their information from their competitors.
Cybersecure ocean monitoring
The team's work also involves applying machine learning to cybersecurity to help develop better techniques for detecting potential cyber attacks on ocean monitoring systems.
Attackers can target websites, the cloud, drones and wired or wireless communication networks. As Dr. Popli explains, if an adversary manipulates or blocks values from the sensors and proper cybersecurity measures aren't in place, the data might still appear normal—even when they should be raising warning signals. "If an application is not secure, it's vulnerable to attacks and its predictions are meaningless. Without security, I can't trust whether the data reflect reality."
To ensure the integrity and security of underwater data collection, a Federated Learning-based Intrusion Detection System—developed originally developed for Internet of Underwater Things networks—is used to enable distributed, privacy-preserving anomaly detection. This safeguards sensor data, which is critical for accurate carbon accounting. Furthermore, an AI-enabled simulation framework adapted from autonomous ships (known as Maritime Autonomous Surface Ships, or MASS) is used to model cyber attack scenarios targeting automatic identification system protocols. This framework makes it possible to monitor the resilience of the sensing and communication infrastructure essential for blue carbon monitoring.
By integrating all of these technologies into the project, we are advancing the scientific understanding of blue carbon processes, supporting the blue economy and contributing to the delivery of secure, scalable and intelligent ocean monitoring systems. These systems facilitate climate change mitigation, improve sustainable management of marine resources and support evidence-based policy development.
Fostering research and skill development
Although this project has commercial applications and contributes to furthering the NRC's research objectives, it also provides a great learning opportunity for students interested in pursuing a career in ocean-related sectors where data must remain private and protected. Furthermore, this research can help generations after us. "By developing highly qualify personnel and a future workforce armed with the knowledge and confidence to address the underwater cyberworld, we're also looking ahead to ways we can help ensure governments and industry have access to quality data sets they can trust, and need, particularly as we move toward a carbon economy," says Dr. Mamun.
To date, the testing on the sensor suite and ROV has taken place in the University of Victoria's lab. For the next stage, which will take place by the end of 2025, the team plans to deploy the ROV in Burrard Inlet in the Salish Sea.