Preserving the earth
(Last updated 2026.09.15)
Previous work with MARIDA and MADOS
The Edge SpAIce marine litter detection framework builds on previous research conducted using the MARIDA [1] and MADOS [2] benchmark datasets, which provide observations of marine pollution and sea-surface feataures in Sentinel-2 imagery. MARIDA is a georeferenced dataset specifically designed for marine debris detection and includes annotations for clouds and other relevant surface classes, while MADOS extends the range of annotated marine debris and sea-surface conditions across additional geographic regions. Within Edge SpAIce, MADOS was further enhanced to create MADOS+, incorporating additional cloud- and water-related classes that can otherwise be confused with marine debris or cloud shadows. This extension enables a unified, end-to-end semantic segmentation approach in which marine pollution, water conditions and cloud-related features are identified within the same model. The knowledge gained from these datasets provided the methodological foundation for the subsequent development and adaptation of the Edge SpAIce detection models to higher-resolution satellite imagery.
Use of Sentinel-2 imagery
Sentinel-2 imagery constituted the initial Earth Observation data source for the development of the Edge SpAIce marine litter detection models. The first-stage models were trained and evaluated using multispectral Sentinel-2 observations, taking advantage of the spectral information available in the visible and near-infrared region for distinguishing floating marine debris from water, clouds, vessels, algae and other sea-surface features. Experiments were conducted using Level-1C data and a selected set of seven spectral bands, while the model later used for transfer to higher-resolution imagery had been pre-trained on MADOS+ using Sentinel-2 VNIR data at 10 m spatial resolution. In addition to naturally occurring marine pollution events, data from the Plastic Litter Project (PLP) were also exploited, providing controlled floating targets such as HDPE structures and natural wooden debris observed under known environmental conditions. Sentinel-2 therefore served as the primary basis for learning robust spectral and contextual representations of marine litter before transferring the developed methodology to higher-resolution satellite sensors.
Use of PlanetScope imagery
To move towards the spatial scales required for operational marine litter monitoring, Edge SpAIce subsequently extended its activities to high-resolution PlanetScope imagery. The dataset was developed using the recently published PlanetScope marine-debris dataset of Dalsasso et al. (2026) as a starting point and was enriched with additional acquisitions from marine pollution hotspots, including the Gulf of Honduras, Haiti, Santo Domingo, Ghana [2] and PLP target sites in the Aegean Sea. The PlanetScope dataset includes 45 scenes, 12.63 million labelled pixels, and 1,839 non-overlapping 240 × 240 image patches, providing a large and geographically diverse training resource. PlanetScope provides Blue, Green, Red and NIR observations at approximately 3 m spatial resolution, allowing floating structures and debris accumulations to be represented at a considerably finer spatial scale than in Sentinel-2 imagery. This high-resolution dataset has therefore become the main source for refining and validating the updated DNN and preparing it for the subsequent model optimisation, distillation and edge-AI deployment stages of the project.
Labelling, DNN architecture and performance
A major part of the Edge SpAIce work has focused on developing detailed semantic annotations and identifying an effective DNN architecture for marine litter segmentation. Whereas the original Dalsasso et al. PlanetScope dataset treated floating debris as a single binary class, additional expert photo-interpretation was performed within the project to distinguish different floating-object categories, including marine litter, macroalgae and other relevant sea-surface classes. Two semantic segmentation architectures were investigated: MariNeXt [2], a transformer-based framework derived from SegNeXt [3] and designed to exploit contextual information in multispectral marine imagery, and the widely used U-Net [4] encoder-decoder convolutional neural network. The U-Net architecture was subsequently adapted from the Sentinel-2/MADOS+ domain to the 3 m PlanetScope domain. The adaptation included modification of the output head to match the active PlanetScope classes, a weighted Focal-Dice objective to address strong class imbalance, extensive geometric data augmentation and probability-aware test-time augmentation. These developments resulted in an overall Macro F1-score of 84.30% and Macro IoU of 75.20%, while the latest experiments reported an F1-score of 93.1% for the Marine Debris class, demonstrating the potential of the updated model for high-resolution satellite-based monitoring of floating marine litter.
References
[1] Kikaki K, Kakogeorgiou I, Mikeli P, Raitsos DE, Karantzalos K (2022) MARIDA: A benchmark for Marine Debris detection from Sentinel-2 remote sensing data. PLoS ONE 17(1): e0262247.
[2] Kikaki K, Kakogeorgiou I, Hoteit I, Karantzalos K (2024) Detecting Marine pollutants and Sea Surface features with Deep learning in Sentinel-2 imagery. ISPRS Journal of Photogrammetry and Remote Sensing 210: 39–54.
[3] Guo MH, Lu CZ, Hou Q, Liu Z, Cheng MM, Hu SM (2022) SegNeXt: Rethinking convolutional attention design for semantic segmentation. In: Advances in Neural Information Processing Systems 35 (NeurIPS 2022). arXiv:2209.08575
[4] Ronneberger O, Fischer P, Brox T (2015) U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, pp 234–241.
DNN OPTIMISATION FOR DEPLOYMENT ONBOARD
To determine the positions of plastics on the oceans and seas, one needs to detect them on aerial or satellite imagery. If we do that the tradutionnal way, all acquired imagery has to be downloaded on ground before we can locate the plastics. The imagery itself represents a very high volume of data and would require heavy and coslty ground infrastructures to be processed. Instead, we consider that processing the imagery onboard and downloading only the detected areas with plastics is a much better option from a cost, efficiency, storage and infrastructure point of view.
To be able to run the detection of platics in space, directly on board of the satellite, we must use specific Deep Neural Networks. In fact, the conditions in space are harsh and resources are scarse. So big models are out of the equation, and only much smaller models are relevant. Indeed, they reduce both power consumption and memory footprint. This is precisely one of our big know-how at Agenium Space: to start from a large model on ground and end with a space hardware compatible small model. This know-how covers the design of specific architectures, the quantization of the models to match the native capabilities of the targeted hardware, distillation (both white and black box) and pruning to keep the relevant parts of the DNN.
This know how was fully deployed along 4 different versions of both the dataset and the model used as input that were produced by the NTUA. Models were always larger than 9M while our targets were spanning 50k-100k weights, yielding a reduction factor strictly above 90 times, keeping 1% of the initial weights. In terms of F1 score performances, the best 8bits quantized 100k model reached 85% on the last database. For on board performances on AMD Xilinx Ultrascale device type, the efficiency is given in the table below. Note that these allow processing of 9M pixels per seconds for the lowest and 97M pixels per seconds at the highest.
Model | Deployment | Frames/s | Power(W) | M Pixels/s/w |
50k – 4bit – 256×256 | Hls4ml | 1071.0 | 3.4 | 20.8 |
100k – 4bit – 256×256 | Hls4ml | 1494.8 | 3.4 | 29.0 |
100k – 4bit – 512×512 | Hls4ml | 305.3 | 3.7 | 21.8 |
100k – 8bit – 240×240 | Vitis AI | 164.2 | 11.5 | 0.8 |
European FPGA in space for AI
Globally there are only a few FPGA manufacturers (major being Xilinx, Microchip, Intel), none of which is European. It is not by choice, but rather related to the high cost of silicone wafer production that makes it a rational choice to have only a few manufacturers of FPGAs globally. However, critical dependency on USA components for the whole space industry makes EU’s geopolitical relationships very strained.
In today’s digital age, particularly the wave of AI, having sovereign EU solution for AI processing is critical. For edge-AI ground applications the Dutch company Axelera-AI is gaining momentum, however, there is no VPU/GPU solution for space. Meanwhile, NanoXplore (NX), a French company, is focusing on building space-grade FPGAs, that can efficiently be used as AI accelerators. While NX company is focusing on making FPGAs, Edge-SpAIce consortium has been focusing on making SW that enables executing AI on it.
HLS4ML was developed by CERN, is an open source project that brings AI-acceleration to FPGAs. During Edge-SpAIce this framework was matured, with following major achievements:
- 2026.05 – 4.2k DNN executed on NX NG-Ultra
global marine litter monitoring Service
EnduroSat’s contribution to Edge SpAIce: the satellites that flew the mission, what a plastic-tracking constellation would need, what it would cost, and what Europe needs to fly its own AI hardware.
1. Balkan-1 and Balkan-2 in Edge SpAIce
Edge SpAIce set out to move Earth-observation processing off the ground and onto the satellite — analysing imagery in orbit and downlinking only the result. EnduroSat supplied the flight segment for that demonstration: Balkan-1 and Balkan-2, the first two satellites of the Balkan Constellation. Both are 16U, 30-kilogram nanosatellites in a single sun-synchronous plane at 475–525 km. Each carries a MultiScape200 CIS push-broom imager giving 1.5-metre ground resolution over a 14-kilometre swath, and an FPGA edge-AI payload.
Balkan-1 — the data campaign
Balkan-1, launched in January 2025, carried the project’s real-world data campaign. Between September 2025 and April 2026 it was tasked with 127 imaging sessions over marine-pollution areas of interest across the Mediterranean and Black Sea, the Americas and Caribbean, and South and South-East Asia. A first curated, cloud-screened dataset was delivered to the consortium’s AI partners, NTUA and Agenium Space, for model development. Images can be added here..
Balkan-1 — the in-orbit demonstration
Balkan-1 is also the testbed for the demonstration itself. Two activities are running: uploading and executing the detection package on board with test imagery, and an end-to-end demonstration of image acquisition followed immediately by on-board processing.
What flies is remarkably small — the whole detection package is on the order of one megabyte:
- An application binary of around 350 kB and a model file of around 700 kB.
- A run script and a readme, so the package can be executed on board as a single command.
- Test imagery sized to the memory available: roughly 325 MB for a full uncompressed scene, around 100 MB compressed, or about 500 kB for a sample tile.
Balkan-2 — spectral upgrade and the Greek plastic targets
Balkan-2, launched in May 2026, adds three red-edge bands that Balkan-1 does not carry. This matters for marine litter: the established spectral indices for detecting floating debris, developed on Sentinel-2, use a red-edge measurement to set the baseline against which a near-infrared information is judged. Balkan-2 delivers those bands at 1.5 metres against Sentinel-2’s 20.
Over the summer of 2026 Balkan-2 was tasked to acquire imagery of controlled plastic targets deployed off the Greek coast — giving the project a known reference on the water against which detection can be validated, rather than inferred. The second satellite also halves the wait: a single Balkan satellite revisits a given point on a five-day cycle, two bring average revisit down to around 50 hours. Images can be added here..
2. What it takes to track plastic from orbit
The short answer is that no new satellite configuration is required. The Balkan specification already meets the requirements for spatial and spectral resolution.
Resolution and spectrum
Floating litter gathers into windrows and slicks metres to tens of metres across. Operational detections today use Sentinel-2 at 10 metres, where a whole aggregation may occupy a single mixed pixel; at 1.5 metres it covers around forty pixels and acquires shape and extent.
Spectrally, water absorbs near-infrared almost completely, so anything floating on it brightens sharply in the NIR against a baseline drawn across the red and red-edge. The Balkan band set closely matches the Sentinel-2 channels the detection literature is built on; the one band it lacks is short-wave infrared around 1,600 nm — the most valuable addition to a future block.
Requirement against capability
Parameter | Requirement | Balkan |
Ground resolution | ≤ 2 m to resolve aggregation structure | 1.5 m at nadir |
Swath | ≥ 10 km for useful area per pass | 14 km |
Spectral bands | Blue to NIR, red-edge for the index baseline | 7 multispectral + panchromatic |
Area per pass | Large, to make searching affordable | 182 × 14 km flown during the project |
On-board compute | Detect in orbit, downlink findings only | FPGA edge-AI payload |
How many satellites
Detecting litter needs enough looks to beat the clouds; tracking it needs the satellite back before the patch has drifted somewhere unrecognisable — which means sub-daily. Scaling from Balkan’s documented performance (about 50 hours’ average revisit with two satellites, roughly 15 minutes at the full 120-satellite build-out), eight to sixteen satellites is the workable operational answer:
Satellites | Average revisit | What it supports |
4 | ~24 hours | Daily watch over hotspots: river mouths, harbours, accumulation zones |
8 | ~12 hours | Drifting patches can be re-acquired and linked between observations |
16 | ~6 hours | Operational tracking: trajectories, attribution, clean-up tasking |
3. What such a constellation would cost
EnduroSat builds and operates Balkan satellites for under €5 million each. That figure is the whole cost of the satellite’s existence, not a hardware price: design and engineering, assembly, integration and test, launch, ground segment, and five years of in-orbit mission operations.
Configuration | Satellites | Whole-life cost |
Edge SpAIce demonstration (flying today) | 2 | under €10 M |
Daily hotspot watch | 4 | under €20 M |
Tracking becomes possible | 8 | under €40 M |
Operational tracking service | 16 | under €80 M |
Three points qualify these figures.
- They are not marginal costs. Each satellite carries its full share of design, test, launch, ground segment and operations. A serial production run of a proven design would be expected to fall below the single-unit figure, not above it.
- The constellation is not single-purpose. Balkan is a general-purpose very-high-resolution imaging and screening service. Marine-litter monitoring shares the platform with coastal monitoring, maritime surveillance, emergency management and land mapping.
- On-board processing lowers the ground cost. Downlinking only detections rather than raw imagery means less ground-station time, less bandwidth and less archive, and those savings compound with every satellite added.
An eight-satellite constellation able to detect and track floating plastic at 1.5-metre resolution, with detection running in orbit, therefore costs under €40 million across its entire five-year life.
4. What is needed to fly NanoXplore on NewSpace missions
Edge SpAIce runs its neural network on a Xilinx FPGA. EnduroSat and Agenium Space assessed what it would take to run it instead on NanoXplore’s radiation-tolerant NG-ULTRA — currently the only credible European supplier for this role.
Where European hardware already fits
The finding is more positive than the usual framing allows. NanoXplore devices demonstrate high radiation tolerance, sufficient to serve as a main controller for mission-critical space systems. For continuous semantic segmentation, cloud masking and coarse object filtering over a moderate-rate instrument — a few tens of billions of operations per second, a few watts to low tens of watts — they are a sound match today. For a great many NewSpace missions, that is the workload.
Where the gap is
- Toolchain — the real bottleneck. Much of an FPGA vendor’s value sits in the software that maps a design onto the fabric and in the path from a trained PyTorch or TensorFlow model to a bitstream. The one published demonstration of a neural network on NG-ULTRA relied on a research framework that is not publicly available.
- Fabric and process. 4-input look-up tables on 28 nm against 6-input tables on nodes down to 6 nm. Published benchmarks put NG-ULTRA at roughly 2.8 times lower operating frequency than a Xilinx Zynq-7000 on the same workload — and that is a 2011 part with four newer families behind it.
- Ecosystem maturity. Reference designs, IP cores for common deep-learning operators, and the knowledge base that lets a small team reach first inference in weeks rather than quarters.
What a flight-ready European edge-AI board needs
- A supported model-to-bitstream toolchain: INT8 quantisation, operator fusion, and a deployment path that packages model, bitstream, runtime and configuration as one signed artefact.
- Custom IP cores for the standard deep-learning operators, and a memory hierarchy built for streaming — on-chip SRAM for line buffers, error-corrected external memory for tile staging, and dedicated DMA engines.
- Radiation and fault handling proportionate to the role: configuration scrubbing, selective triple-modular redundancy on critical control logic, ECC, watchdogs, redundant boot images and authenticated update with rollback.
- Power and thermal closure on measured numbers, and a hardware-in-the-loop bench that replays recorded sensor streams at flight rate across temperature.
The number that sizes the problem
Edge SpAIce produced a concrete measure of what “fast enough” means. In the Balkan mission analysis the on-board network takes about 1.04 seconds per kilometre of imaged strip, while the satellite covers that kilometre in about 0.14 seconds — roughly a factor of seven short of real time. True real-time detection over a high-rate multispectral instrument implies 0.6 to 1.7 tera-operations per second at INT8 and 60 to 170 watts, beyond a single device in most smallsat budgets.
The practical route is therefore model compression and quantisation, multiple devices where the power budget allows, and relaxation of the real-time target where the mission does not need it. On that route European silicon is already viable — and the toolchain gap, not the silicon, is the obstacle worth funding.