MEPA Orbital

01MEPA Orbital

A satellite for the Mediterranean.

A long-term concept: a 200 kg satellite, built for seven years or more, that looks at the sea in colour and in heat, listens for ships, and gives MEPA its own view. It is planned, not funded. Scroll to look closer.

Planned. Not funded. Scroll to look closer.

02The eye

A camera that splits light into colours.

Behind one lens: a telescope with eight colour bands (target about 3 m detail) and, beside it, a thermal camera that sees heat (target about 100 m). Oil changes the colour and glint of the sea, and the cameras are designed to catch that.

03What it sees

A dark patch behind a ship.

Two computers in the satellite are designed to screen each image: they throw out cloud, spot ships and candidate slicks, and match a ship to the signal it broadcasts (AIS). A ship with no signal is flagged. They only suggest. A person decides whether anything is a case.

04Night and heat

A second pass in the dark.

The same orbit gives a daylight pass for the colour camera and a night pass for the thermal camera, which sees thick oil and warm discharges against the cool sea. A tasked instrument: twenty minutes a day covers about 8% of the Mediterranean, so it looks where something is suspected.

05The whole

Built to last seven years.

Large solar wings and a deep battery, an electric thruster to hold the orbit and bring it down at the end, and a radio link to ground stations we rent by the minute. An optional laser link to SpaceX’s network, only if offered. It works with free Sentinel-1 radar, not instead of it.

The proposal

What the satellite would carry.

A proposed specification for discussion with partners and funders. Every figure is a target, nothing is built, and nothing here has been agreed with SpaceX or any supplier.

Class: About 200 kg, the first of a series (MEPA-1). Design life of seven years or more
Look: Matte black carbon-fibre composite body with gold-foil accents, deep-blue solar wings and a white radiator: built to be seen, and to be copied
Orbit: Sun-synchronous, about 550 km, 10:30 local time: a daylight pass for the colour camera and a night pass for the thermal camera
Colour camera: Multispectral telescope, about 25 cm aperture, 8 bands from blue to near-infrared, about 3 m detail, swath about 25 km
Thermal camera: Uncooled long-wave infrared, 640 × 512 pixels, about 100 m detail, swath about 60 km, day and night
Ship tracking (AIS): A receiver for ship transponder messages, so a ship seen by the cameras can be matched to its identity, and a ship with no signal is flagged as dark
AI processors: Two space-screened edge AI modules (one spare), each about 100 trillion operations a second (INT8), 10 to 25 W
Memory and storage: 16 GB of fast memory with error correction per module, 8 TB of flash holding about a month of raw imaging
Models in orbit: Small vision models, from tens of millions up to about 8 billion parameters (4-bit, about 5 GB): cloud rejection, ship and candidate-slick detection, download ranking
Flight computer: A separate, simpler radiation-tolerant computer for housekeeping and safe mode, with a watchdog that can reset the AI modules
Power: About 3 m² of solar wings (about 700 W peak, 300 to 350 W average) and a battery of about 1 kWh, oversized for seven years of cell ageing
Propulsion: A small electric (Hall-effect) thruster to hold the orbit against drag for seven years and to bring the satellite down at the end of its life
Downlink: X-band, up to about 600 megabits a second (about 45 GB in a ten-minute pass), via commercial ground stations
Commands: S-band or UHF, encrypted, with signed software and model updates
Relay (optional): SpaceX Starlink laser terminal, reported at 25 gigabits a second over up to 4,000 km, for alerts within minutes. Not agreed, not required
Pointing: Star trackers, four reaction wheels and magnetic torquers
Launch: SpaceX Transporter rideshare, booked through an integrator about 12 to 18 months ahead

Where the thinking happens

In orbit: A small model does the first pass: reject cloud, find ships and candidate slicks, rank the download queue. It runs in seconds on a few watts.
On the ground: The largest frontier AI models, far too big for any satellite, read the candidates with every other source (radar, ship tracks, wind, currents), write up the evidence and flag what a person must check.
Why split it: A model of a trillion parameters needs hundreds of gigabytes of memory and kilowatts of power. A satellite of this class has about 200 watts. So the satellite sends small, ranked packages and the ground does the thinking.

How fast this has to move

A rideshare slot has to be booked 12 to 18 months before launch, and sponsors take six to twelve months to decide. To fly around 2030 the proposal has to be selling now.

Months 0 to 3: The proposal package: this page, a specification, a budget, a risk list. Pitch ready.
Months 3 to 9: Outreach: SpaceX Starlink business team, bus makers, university partners, ESA and philanthropic funders. First free and bought imagery to prove the need.
Months 9 to 18: Three written quotes, a hosting or bus partner, a lawyer’s permit pathway. A funding decision.
Months 18 to 30: Design review, build and environmental testing. The launch slot is booked 12 to 18 months before launch.
About years 3 to 4: Launch and the first images. Realistically around 2030 or later.

For the engineers

The numbers behind the targets, worked out so that they agree with each other. All of them are open to challenge.

Orbit and coverage

Orbit: 550 km sun-synchronous: a period of about 95.5 minutes, about 15 orbits a day, 7.6 km/s along the orbit and about 7.0 km/s over the ground
Light: 10:30 local time on the way down (daylight, good sun angle for the colour camera) and 22:30 on the way up (night, for the thermal camera). Up to about 35 minutes of shadow per orbit
Area imaged: A 25 km swath sweeps about 175 km² every second. Twenty minutes of imaging a day covers about 210,000 km², about 8% of the Mediterranean (about 2.5 million km²). So it is a tasked instrument: it looks where something is suspected, not everywhere
Revisit: A given point is reachable roughly once or twice a day by tilting up to about 30° off the ground track. Three satellites would make that near-daily across the basin

Data budget

Colour camera: At 3 m, 25 km is about 8,300 pixels across and about 2,300 lines a second. Eight bands at 12 bits is about 1.9 gigabits a second raw: about 280 GB for twenty minutes of imaging
After compression: Near-lossless compression at about 3 to 1 gives about 90 GB a day
Storage: 8 TB of flash holds about a month of raw imaging, so nothing is lost if a ground station is out of reach for days
Downlink capacity: At one Mediterranean ground station, typically 4 to 6 useful passes a day of 6 to 9 minutes each. At 600 megabits a second that is about 150 GB a day of capacity, about 1.7 times what the imager produces

Power

Generation: About 3 m² of triple-junction cells: about 700 W at the peak, about 300 to 350 W averaged over the orbit, with margin for about 2% ageing a year
Use while imaging: About 200 to 230 W: computers 25, pointing 20, colour camera electronics 45, thermal camera 8, AIS 5, AI modules up to 50, heaters 30, and margin
Downlink and thrust: The X-band transmitter adds about 60 W for a pass, and the electric thruster about 250 W while it fires, in sunlight only
Battery: About 1 kWh. A 35-minute eclipse at 220 W takes about 130 Wh, a depth of discharge near 13%, gentle enough for seven years

Mass

Structure, power, pointing: About 90 kg: carbon-fibre structure and harness 38, battery and power electronics 18, solar wings 16, star trackers, wheels and torquers 18
Payloads: About 62 kg: telescope 45, thermal camera 4, AIS receiver 3, AI and flight computers 10
Radios and propulsion: About 28 kg: radios 10, electric thruster, tank and xenon 18
Margin: About 10%. Total about 200 kg, so a launch at about $7,000 a kilogram is about $1.4M

Compute

AI modules: Two modules, one spare, each about 100 trillion INT8 operations a second at 10 to 25 W, so 4 to 10 trillion per watt. 16 GB of memory with about 100 GB/s of bandwidth
A slick screen: A segmentation model of tens of millions of parameters, run over 512-pixel tiles (1.5 km at 3 m): a 25 × 25 km scene is about 280 tiles and takes a couple of seconds [est.]
A language model in orbit: A 4-bit model of 8 billion parameters fills about 5 GB. Memory bandwidth limits it to roughly 10 to 20 words a second, plenty to write a one-line note about a candidate
Fault protection: ECC memory, latch-up-protected power rails, periodic memory scrubbing, and a separate flight computer with a watchdog that can power-cycle the AI modules. We expect single-event upsets about daily and design for them

Ship tracking

What it does: The receiver collects ship transponder (AIS) messages in the area the cameras are looking at. A detection is matched to the nearest message in time and place. No match is a dark vessel
Limits: In busy water many ships talk at once and some messages are lost, so a missing signal is a lead and not proof. Ships can also switch off or spoof their signal, which is why it is one source among several
Cost and size: About 3 kg and 5 W. The antenna is the thin mast on the top of the satellite

Seven years in orbit

Drag: At 550 km the atmosphere is thin but not absent, and it thickens with solar activity. The electric thruster, with about 8 to 12 kg of xenon, holds the altitude and keeps the camera pointing where it should
End of life: The same thruster lowers the orbit at the end, so the satellite falls back and burns up within a few years instead of staying as debris
Wear: Solar cells lose about 2% a year, so the wings are oversized. The battery is run shallow. Wheels and star trackers are duplicated. Software is updated in orbit, with signed updates only
Dose: In this orbit, behind about 3 mm of aluminium, on the order of 15 kilorad over seven years [est.]. Commercial parts are acceptable with screening, shielding and mitigation; the flight computer is chosen for tolerance

Pointing and measurement

Pointing knowledge: A 3 m pixel at 550 km is about 1 arcsecond across. Star trackers give a few arcseconds, so we correct on the ground against known points and aim for 10 to 15 m of location accuracy
Calibration: Dark frames, a lunar view once a month, and ground targets as references. The calibration log goes with the data. The thermal camera sits on its own radiator and is recalibrated often

Evidence on board

Signed at capture: Each frame gets a GNSS timestamp, its position and pointing, the AIS messages heard, and a hash. A secure element signs the record before it is stored, so the file can be shown to be unaltered since orbit
Uplinks: Commands, software and model updates are signed and encrypted, with keys held by MEPA and rotated. No update runs without a signature
Ground: The original frames and the signed records go into the same sealed archive as every other MEPA source, with the processing steps logged

Where this stands

An idea we are taking seriously, and not yet a mission.

What it is

A concept for a 200 kg multispectral and thermal satellite with a ship-tracking receiver, designed for seven years or more, with a target budget of about €20 million (a range of €15 to 25 million), to be confirmed by independent quotes. Nothing is funded, built or approved, and we are not asking for money for it. We plan to use AI heavily in the design, the software and the operations, with named people accountable for every decision.

What it is not

It is not a radar satellite, which would cost several times more. It cannot see through cloud, or the colour camera in the dark, and one satellite cannot watch the whole Mediterranean. Free radar from Sentinel-1 stays the backbone.

What comes first

We prove the need with free imagery and a few bought scenes, measure where they fall short, and only then decide whether to rent a place on someone else’s satellite or build our own. A satellite is a decision we make with evidence in hand.

The gates

Before real money is committed: a measured gap, full funding for the whole life of the mission, written quotes, a lawyer’s pathway for the permits and radio licences, a funded operations team, and a result that does not depend on any one company.

What we need today How the system works

The satellite shown is an illustration of a concept, not a design drawing. Targets are goals, not measured results. Questions or ideas: admin@mepa.ngo