Humans produce over 400 million metric tons of plastic annually. Less than 0.5% ends up in the ocean, but that fraction still amounts to over one million metric tons of pollution every year. Of the plastic that floats, a significant fraction disperses across open water, accumulates in offshore garbage patches, and progressively degrades into microplastics.
Tracking floating plastic litter — where it is, how it moves, and where it concentrates — is a prerequisite for any data-driven intervention at scale. That is the issue the Automated Debris Imaging System (ADIS) was built to address.
The ADIS is designed as a distributed sensing platform that deploys edge AI cameras across commercial vessels to capture imagery of the ocean surface and detect floating debris in real time. The system integrates edge inference, large-scale data pipelines, and statistical modelling to transform raw imagery into numerical abundance estimates along vessel tracks.
From an engineering perspective, ADIS is a full-stack distributed sensing architecture: edge hardware operating in maritime environments, cloud pipelines processing large volumes of detections, and modelling layers converting detections into calibrated abundance estimates.
Edge AI architecture at sea
The core sensing unit of ADIS is a compact AI camera platform designed to operate autonomously aboard ships for extended periods.
Each camera is built around a MAIVIN edge compute module equipped with a neural processing unit (NPU) capable of running computer vision inference locally. The onboard model performs object detection on interval-based image captures rather than continuous video streams.
This interval-based design choice reflects several operational constraints:
- Power efficiency: Cameras operate at approximately six watts, allowing them to draw power from standard vessel electrical systems without dedicated infrastructure.
- Data minimisation: Continuous video would generate impractical data volumes in maritime deployments.
- Robustness: Periodic still-image capture simplifies processing and reduces failure modes in constrained edge hardware.
During inference, the system identifies candidate debris objects within captured images and performs bounding box detection directly on the edge device. Rather than transmitting entire frames, the camera extracts cropped detections and attaches metadata describing the observation. The metadata includes the timestamp, geolocation, bounding box information, and model classification output.
By performing detection locally and transmitting only extracted detections and metadata, the system reduces the bandwidth required for large-scale deployments, compared to approaches based on full image transfers.
A data pipeline designed for intermittent connectivity
Operating at sea introduces a constraint uncommon in most distributed AI systems: Connectivity is sporadic.
Ships often lack reliable data links while at sea, so the ADIS data pipeline is designed to function in an offline-first mode. Edge devices store detections locally and synchronise data only when network connectivity becomes available, typically when vessels enter port and reconnect to mobile networks.
The pipeline operates in several stages:
- Edge detection
Captured images are processed onboard the camera using the local AI model. Detected objects are cropped and stored with associated metadata. - Local storage
Detection data are buffered locally on the device until network connectivity becomes available. - Opportunistic upload
When vessels enter port and connect to mobile data networks, stored detections are transmitted to the cloud back end. - Cloud processing
Once uploaded, the data are ingested into a cloud pipeline built on AWS cloud infrastructure, where an open-source Python- and PostgreSQL-based pipeline performs further post-processing.
This architecture allows the system to scale across globally distributed vessels without requiring continuous satellite communication or dedicated networking infrastructure.
Training computer vision models for rare objects
A key challenge for this project is training computer vision models for a domain where positive examples are both sparse and highly variable. Floating debris in open ocean environments exhibits significant visual diversity: lighting conditions vary, wave patterns obscure objects, and debris items vary in size, shape, and material. At present, the classification taxonomy focuses on five primary categories: hard plastics, fibrous plastics (including ropes and fishing nets), buoys, plants, and animals.
This simplified taxonomy reflects a pragmatic design decision: prioritising detection reliability and operational robustness over fine-grained categorisation.
However, continuous offshore data collection offers the opportunity to gather a more refined training dataset. This is currently in progress, based on a subset of cameras specifically configured to collect full image frames in addition to the standard snippets and metadata.
Annotations are generated through a semi-automated workflow. Model detections are reviewed by human experts, allowing the team to refine training data while minimising manual labelling effort.
From detections to ocean-scale intelligence
Understanding detection range is critical for converting raw detections into numerical abundance estimates. Without it, detection counts cannot be normalised relative to the observed ocean surface area. To address this, ADIS incorporates geometric projection techniques that estimate object distance from the camera based on image geometry. By analysing the distance distribution, the system derives a probabilistic estimate of the effective scan width. Combined with the segment length, it provides the area required to determine numerical abundance.
Once detections are validated and detection ranges are established, ADIS observations are aggregated along vessel track segments, producing numerical abundance values expressed in pieces per square kilometre. The resulting data spans major shipping routes across the global ocean.
Beyond direct measurement, these observations are valuable for calibrating numerical ocean circulation models that simulate the transport of floating plastics through advection and ocean currents.
While satellite detection of individual marine debris items remains an open challenge, the time- and location-tagged observations generated by ADIS provide a dataset for validating and constraining large-scale ocean plastic transport models.
Scaling a distributed ocean observation network
The ADIS network currently operates across 18 ships with 33 deployed cameras.
To date, the system has scanned more than 20,000 square kilometres of ocean surface and processed more than 200 million image frames across the global fleet. This has produced one of the largest datasets of ocean-surface imagery for floating debris detection.
However, the architecture was designed with larger deployments in mind. In principle, the platform could scale to thousands of cameras distributed across global shipping fleets, effectively transforming commercial vessels into a distributed ocean sensing network.
A network of approximately 1,000 cameras could provide broader observational coverage, generating continuous data streams across major maritime routes.
Built for the ocean, ready for scale
With over one million metric tons of plastic entering the ocean annually, the monitoring challenge is inseparable from the clean-up challenge. ADIS shows that a distributed edge AI architecture designed for maritime conditions, built around offline-first data pipelines, and refined through operational deployment can generate the observational coverage required for ocean-scale intervention.
As the network scales from 33 cameras to potentially thousands across global shipping fleets, the volume and continuity of data it produces will become an increasingly important input to understanding and addressing a persistent pollution challenge.
