It depends on how much power you're limited to in the jurisdiction where you're operating those radios.

In the US, a point-to-point link can use as much as 45dB in certain portions of the band. The flip side is point-to-multipoint (which is how you're technically operating if you're using omni's) is limited to as low as 30dB in some portions of the band. And CPE are allowed slightly higher TX power than AP's in some configurations.

Assuming worst-case scenario, the 30dB limit with 20dBi antennas gets you -76dBm and maybe 30Mbps in each direction on 5GHz at 8km on a 20MHz channel. If you have the full 45dB to play with, you could get more bandwidth or use smaller antennas (15dBi) to achieve a similar (or possibly better) result.

With multiple radios at each end, you're also looking at some kind of way to adapt to the changing link. I would go with routers doing OSPF (+BFD if using v6 routers).

For that distance, this gets elaborate pretty quickly.

We're under European regulations.
Can those multiple radios be merged? i.e. can you have 3 directional antennas and a single common output?

We're under European regulations.
Can those multiple radios be merged? i.e. can you have 3 directional antennas and a single common output?

You could, but you lose 3dB for each split. With the minimum required two splitters, one antenna would be 3dB down and two of the three antennas would be 6dB down. If you used three splitters (to make the third one match), all of them would be 6dB down.

The highest gain dual-polarity 5GHz omni's I've found with just a quick Google search are 13dBi. High-gain 120° sectors run around 16-19dBi. Take 6dB off that, you're down to 10-13dBi anyway. The splitters undo any benefit you'd gain from using sectors with a single radio.

Ultimately, you'd need an array of radios with high-gain, narrow-beam antennas. The highest gain, narrow-beam sectors I've used are Ubiquiti's 22dBi 45° sectors. You'd need eight of those for 360° coverage, and they're not small.

Horns at best will get you 18-20dB over 30°; you'd need 10 of those. For anything over 22dB, you start moving into dish territory, and 23dBi dishes average about 10-13° wide. (20-30 radios doesn't seem very practical.)

Most sectors have an electrical down tilt and a vertical beamwidth of 3°. You'd have to tilt them back to compensate. On rough seas, the antennas would go in and out of vertical alignment constantly. Horns would be a better fit if the water isn't guaranteed to be still the majority of the operating time.

The other factor is budget. Radios will run you anywhere from US$50 to US$250, and antennas from US$250 to $US500. Custom mounting hardware, labor, etc. will bring the price up to thousands of euros/dollars per watercraft.

Don't get me wrong. It would be a fantastic solution to engineer, and actually getting it to work would be exciting. I think understanding the rest of the use case would help to see if alternate solutions would fit the need.

If it's just two points you need to connect (not multiple stations) then use dish antennas and antenna rotators to steer them accordingly. You may only need to do the azimuth if it's a ship on the sea (difference in elevation will be minimal, probably within beamwidth). You would need some another low rate link to exchange positions, but that could be VHF or UHF and very low rate so it works with just omni antennas (could be as slow as 100bps, doesn't really need more than like 1 update per minute).
Some inspiration could be steered antenas used by drone operators for live video or HAM radio community. Also I remembered similar project done by re-purposing Diseqc satellite dish motors. They are "smart" in a way you send them just command "go to +30degress" and it turns. Or CCTV camera mounts.
All this will require some custom arduino programming and HW assembly... but it looks like fun and interesting project.

Well it seems omnis is the least bad choice. We also thinking about stabilizing the antennas, maybe with gimbals, but that's a very out of the box idea. We'll see. Thanks for the informations.

I was thinking about an automatic steering system. The only problem is indeed the positioning. We're thinking for a stabilizing algorithm that "seeks for a certain threshold in signal strength and steer acordingly. Could you elaborate on the VHF solution?

I just remembered that omnidirectional antennas usually have the same narrow vertical beamwidth (3-5°) and electrical down tilt (usually 2°) that most sectors do. You'll need to factor that in.

"omnidirectional" in antenna description most often describes horizontal radiation pattern. As you say it doesn't necessarily describe also vertical radiation pattern, it very much depends on omni antenna construction. However the narrow vertical beam width mentioned is quite unusual and is usually result of using very long antenna (i.e. multiple tens of wavelengths) to achieve higher gain. More oftenly omni antennae feature vertical beamwidths of around 60° - +-30° around horizontal plane . Also keep in mind that realistic omidirectional antenna can never be truly omnidirectional in all 3 dimensions, even simplest dipole has a gap in directions along the dipole. And mind that antenna gain is (generally) inversely proportional to both vertical and horizontal beamwidths, so high gain antennas can never be omnidirectional, they have to be directional (to certain degree) in one or both directions.

... it is assumed that we want to flatten the pattern from a standard dipole doughnut shape (with a hole the size of the radiating element) to something resembling a pancake ...

Mount the antenna 'hanging' (with some extra weight) so gravity (theoretically) evens out the waves.