Cleaning the world’s oceans has long been a race against time and tide. For organizations tasked with removing plastic pollution, the primary challenge is not just the act of collection, but the logistical nightmare of interception. By the time a satellite identifies a patch of plastic, the currents have often shifted, leaving cleanup vessels chasing a ghost that has already moved miles from its last recorded position.
To solve this, researchers at the EPFL (École polytechnique fédérale de Lausanne) have developed a system that uses AI to track ocean waste from space and, more importantly, predict where it will be tomorrow. Known as the ADOPT project, this initiative leverages machine learning to bridge the gap between static satellite imagery and the dynamic reality of ocean currents.
The project, led by the Laboratory of Computational Science for Environment and Earth Observation (ECEO) in collaboration with the Swiss Data Science Center, transforms satellite data into an actionable operational tool. By combining high-resolution imagery with predictive drift modeling, the team is providing a reliable spatial and temporal window for NGOs and cleanup operations to deploy their resources with precision.
As a former software engineer, I find the technical leap here particularly compelling. The team isn’t just using AI for image recognition; they are using it to correct the inherent flaws in traditional physical simulations of the ocean, creating a hybrid model that is far more accurate than either approach alone.
From Sparse Data to Daily Surveillance
The foundation of the ADOPT project began with the Sentinel-2 satellites operated by the European Space Agency. While these satellites provide a wealth of data, they possess two significant limitations for real-time cleanup: a resolution of 10 meters per pixel and a revisit time of six days. In the fast-moving environment of the open sea, a six-day gap is an eternity.
To overcome this, the EPFL team engineered an AI capable of “knowledge transfer.” They trained the system on the high-quality data from Sentinel-2 and then adapted it to work with the nanosatellites of PlanetScope. These smaller, more numerous satellites offer a revolutionary shift in capability, providing daily surveillance with a precision of three to five meters per pixel.
At this resolution, the AI is not searching for individual plastic bottles—which would be nearly impossible from orbit—but rather “windrows.” These are long, concentrated bands of debris formed by converging currents. By identifying these windrows, the system can pinpoint massive accumulations of plastic that are prime targets for collection.
Comparing Satellite Capabilities
| Feature | Sentinel-2 (Baseline) | PlanetScope (ADOPT) |
|---|---|---|
| Pixel Resolution | 10 meters | 3 to 5 meters |
| Revisit Frequency | Every 6 days | Daily |
| Primary Target | General debris patches | Concentrated “windrows” |
Predicting the Drift: The Hybrid AI Approach
Identifying a plastic patch is only half the battle. Because ocean currents and winds are volatile, a target spotted on Monday may be dozens of kilometers away by Tuesday. This “drift” is the primary reason many cleanup missions fail to intercept the debris they are targeting.
The ADOPT project addresses this by merging classical physics with machine learning. Traditional models use wind and current data to simulate movement, but these simulations often contain imperfections. To correct these, the researchers integrated historical data from thousands of GPS-equipped drifting buoys.
The AI analyzes the discrepancy between where the physical models predicted the buoys would go and where they actually ended up. By learning from these errors, the AI recalibrates the trajectories in real-time. This allows the system to provide organizations like The Ocean Cleanup with a precise forecast of where a plastic mass will be 24 hours after the satellite image was taken.
Open Source and Future Constraints
In a move that emphasizes ecological collaboration over proprietary gain, the EPFL and its partners have released the detection and drift codes as open source. This ensures that the global community can integrate these tools into their own environmental monitoring systems without the barrier of licensing fees.
Despite the breakthrough, the project faces a persistent physical obstacle: clouds. Because the current system relies on optical sensors, heavy cloud cover can blind the satellites, leaving gaps in the surveillance. The research team has identified the integration of radar technology—which can “see” through clouds—as the next critical frontier for the technology.
The transition from a proof-of-concept to a global intervention force depends on this ability to maintain a constant eye on the ocean, regardless of weather conditions. By moving the logic from “where is the plastic now” to “where will the plastic be,” the ADOPT project shifts the paradigm of ocean cleanup from reactive to proactive.
The project concludes its current phase as a solid proof of concept, with the open-source release serving as the primary mechanism for its immediate global scaling. The next steps for the research community involve refining these models for different oceanic regions and exploring the integration of synthetic aperture radar (SAR) to eliminate the “cloud gap.”
We want to hear from you. Do you believe open-source AI is the fastest way to solve environmental crises, or should these tools be managed by centralized international bodies? Share your thoughts in the comments below.
