Edge SpAIce Shows How Onboard AI Can Detect Marine Litter From Orbit

The EU-funded Edge SpAIce project has demonstrated an onboard AI approach for detecting marine litter from satellite imagery. Rather than transmitting every captured image to ground stations for analysis, the system processes imagery on the spacecraft and sends down detection findings. That changes the data path for environmental monitoring, where satellite bandwidth and the time required to move imagery to Earth can constrain how quickly analysts receive usable information.

The project carried out an in-orbit data campaign with European nanosatellites, using lightweight, FPGA-based edge AI to identify possible marine pollution in imagery collected over marine-pollution hotspots. According to the official Edge SpAIce results, the operational concept is straightforward: detect in orbit and downlink findings only, rather than relying on full-image downlinks for every observation.

This is a meaningful demonstration of edge computing beyond Earth. It does not eliminate the need for ground-based analysis or validation, but it shows that a compact model can filter and prioritize observations before they leave the satellite. For environmental monitoring teams, that could mean concentrating limited connectivity on the images and locations that warrant attention.

How Edge SpAIce processes satellite imagery in orbit

Edge SpAIce uses FPGA platforms designed for European space applications. Field-programmable gate arrays can be configured to run specialized workloads efficiently, making them suited to constrained environments where power, storage, and communications capacity are limited.

The project's onboard detection package is deliberately compact. It includes an application binary of roughly 350 KB and an AI model of about 700 KB, or approximately 1 MB in total. That small footprint is central to the demonstration: the project is testing whether a practical marine-litter detection workflow can operate on space-grade hardware, not simply whether a large model can identify objects in a ground-based computing environment.

From ground training data to onboard detection

The development process began with satellite imagery from Sentinel-2 and PlanetScope. The project identifies PlanetScope imagery as providing about 3 m resolution data. It then moved toward higher-resolution imagery and refined a semantic-segmentation workflow using approaches including MariNeXt and U-Net variants.

Semantic segmentation is useful here because the task is not merely to classify an entire image as polluted or clear. The system must identify image regions that may correspond to marine debris and distinguish them from visually confusing features such as water, clouds, and natural sea-surface patterns.

The project reports the following performance results for its refined segmentation work:

• A Macro F1 score of about 84.3%.
• A Macro Intersection over Union score of about 75.2%.
• A Marine Debris class F1 score of around 93.1%.

These metrics describe the project's reported model performance and should be read in the context of its datasets and methodology. In operational use, detection quality will still depend on factors such as image resolution, viewing conditions, cloud cover, sea state, and the similarity between deployment imagery and training data.

Balkan-1 and Balkan-2 extend the in-orbit campaign

Balkan-1, launched in January 2025, carried the lightweight edge-AI payload. It supported an in-orbit imaging campaign between September 2025 and April 2026 that included more than 100 imaging sessions across marine-pollution hotspots.

Balkan-2 joined the effort after its May 2026 launch. Its spectral upgrades, including red-edge bands, were intended to improve discrimination between possible marine litter and water, clouds, or natural surface features. The satellite also brought higher spatial resolution, at 1.5 m ground sampling distance (GSD).

Why onboard filtering matters for environmental data services

The important advance is not simply that an AI model runs on a satellite. It is that the model can decide which observations merit transmission. Satellite imagery can generate far more data than a mission can conveniently send to Earth, especially when a workflow needs frequent observation or fast triage. Reducing transmissions to detections can preserve bandwidth for higher-value information and shorten the route from image capture to a potentially actionable finding.

Edge SpAIce also reports substantial onboard throughput at low-bit quantization, with examples in the range of roughly 1,000 to 1,500 frames per second while using around a few watts of power. These figures illustrate why model compression and hardware-aware deployment matter. A model that is accurate in a conventional computing environment is not automatically useful on a satellite if it cannot meet the platform's power and processing constraints.

For organizations building geospatial analytics or environmental-monitoring services, the demonstration points to several practical possibilities:

• Faster alert workflows when systems can transmit prioritized findings instead of waiting for complete image transfers.
• More focused analyst workloads because teams can begin with candidate detections rather than review all captured imagery.
• New specialized data products based on detection events, monitoring areas, or prioritized imagery rather than raw data alone.
• Greater importance of validation because automated detections still need to be assessed against real-world conditions and the limits of the training data.

The project does not establish that every satellite-imagery workflow should move fully onboard. Full imagery remains valuable for analysis, evidence, retraining, and quality assurance. But the Edge SpAIce campaign demonstrates a hybrid model in which edge AI can act as an early filter, while people and ground systems handle deeper assessment.

For businesses, the nearer-term lesson is about system design. AI value often comes from placing inference where it removes a bottleneck, rather than adding a model after data has already traveled through a slow and expensive process. The same principle applies to industrial sensors, cameras, field devices, and remote assets, even when the hardware is far closer to Earth than a nanosatellite.

If your team is assessing where AI can remove delays from data-heavy processes, Scalevise can help turn a promising concept into a practical implementation plan. Our AI implementation consultancy for business workflows helps identify viable use cases, assess technical constraints, and prioritize integrations that reduce manual work without overbuilding. Start by requesting a consultation to map the highest-value AI opportunity in your operation.

What to watch next

The project plans final dissemination activity that includes a dedicated workshop in October 2026. The next useful evidence will be how the higher-resolution, expanded-spectral-coverage work affects detection reliability in varied marine conditions, and how onboard findings can be incorporated into broader monitoring workflows.

Frequently Asked Questions

What is Edge SpAIce?

Edge SpAIce is an EU Horizon Europe project developing lightweight edge AI for onboard processing of satellite imagery. Its demonstrated use case is identifying possible marine litter directly on nanosatellites.

How does Edge SpAIce reduce satellite data transfers?

The system runs detection on the satellite and downlinks detections rather than transmitting all captured imagery. This can focus communications capacity on findings that may need follow-up.

Which satellites were used in the demonstration?

The project used Balkan-1, launched in January 2025, for an in-orbit campaign from September 2025 to April 2026. Balkan-2, launched in May 2026, added higher-resolution imaging and expanded spectral coverage.

Can onboard AI reliably distinguish litter from sea-surface features?

Edge SpAIce reports that Balkan-2's spectral upgrades, including red-edge bands, are intended to improve discrimination between marine litter, water, clouds, and natural sea-surface features. Detection results still depend on imagery conditions and model performance.

Conclusion

Edge SpAIce provides a concrete in-orbit demonstration of a compact AI system that can identify possible marine litter before imagery reaches Earth. By pairing FPGA-based processing with a detection-only downlink approach, the project shows how edge AI can make constrained data workflows more selective. Its broader significance lies in proving that useful AI inference can happen at the point of data capture, where it can reduce transmission demands and speed up environmental monitoring workflows.