The EPFL-led ADOPT (AI for Detecting Ocean Plastic Pollution with Tracking) project is integrating AI satellite-image recognition with drift prediction models to enhance ocean plastic debris collection. The technology has completed its proof-of-concept phase and is prepared for field testing.
Project Goals and Development
Identifying and tracking floating debris masses is crucial for ocean clean-up efforts. The ADOPT project, initiated two years ago, aims to address this need by developing two primary systems:
- Debris Identification: Analyzing satellite images to locate garbage patches.
- Drift Prediction: Forecasting the movement of these patches within 24 hours to guide clean-up teams.
The project is a collaboration between EPFL's Environmental Computational Science and Earth Observation Laboratory (ECEO), the Swiss Data Science Center (SDSC), and Wageningen University in the Netherlands.
Technology Details
Initially, the team utilized open-access data from European Space Agency's (ESA) Sentinel-2 satellites. To overcome limitations such as a six-day revisit time and 10-meter per pixel resolution, an AI system was designed to incorporate data from PlanetScope, a constellation of nanosatellites providing daily images at 3-5 meters per pixel. This AI-driven detector combines data from both sources, updating daily with higher-resolution images without requiring manual data annotation.
The system is designed to detect large collections of plastic and debris, such as windrows that can extend hundreds of meters, rather than individual items.
Drift Prediction System
The second system, developed by Christian Donner at the SDSC, predicts debris drift. It uses established models for wind and current forecasting, applying machine learning to correct inherent biases in these models. The machine learning program compiles data from various sources to calibrate biases, improving trajectory predictions. The program was trained using data from GPS-equipped drifters from the Global Drifter Program, acting as a proxy due to limited field data on actual garbage patches.
Challenges and Future
The current optical sensor-based system is ineffective in bad weather due to cloud cover. Incorporating radar images from Sentinel-1 is a potential future solution, as radar signals penetrate clouds and operate day and night. However, radar provides only texture and geometry information, missing the spectral signatures essential for garbage patch detection.
The ADOPT project is scheduled to conclude in the fall when its two-year funding period ends. The team will provide a proof of concept, two publications, and the code for both the debris detection and drift prediction systems.
Dutch NGO The Ocean Cleanup plans to continue comparing the algorithms, and university scientists will proceed with further research based on this work.