Pre merging seems a good idea as the textures would already be 'set' if you know what I mean? To me it sounds good but I am not the expert
Do the numbers... A texture sheet is 2x2 km and takes about 0.5 MB on the harddisk in compressed form, judging by the dds texture size perhaps 2 MB when decompressed and mipmapped in the GPU memory.
With a visibility range of 50 km, we see about 2000 such texture sheets in the scene. If we pre-compute them, each of them is different, so they need 2 MB each of graphic memory, i.e. we need 4 GB memory just to store the textures (!) - that's where pre-computing starts to look a bad idea.
So trying to do this on the fly at 10 cms is very complex?
Moderately so for natural landclasses - there it's just mixing based on noise terms and terrain gradient (I think the shader executes ~400 lines of instructions per pixel, but a lot of this is fog and light computation, the hires texturing itself is just 30% of the computation load).
If I am getting this correct, we could use your Shader based LOD system to get that grass effect?
We'd do this the other way round: For natural terrain which you mostly see from high altitude, we want the base pattern to be supplied by the base texture (forest clearings, different forest types, see a mixedforest texture). On top of that we want to dynamically add detail by mixing smaller patches of hires textures in.
For airport grass, the highest detail would be in the base texture already, and what we want to do is to add larger scale patterns on top of that so that it doesn't look like a flat green surface from the distance when the hires detail can no longer be seen. This wouldn't work so well for other landclasses - one has to custom-design it in many cases.
I could make the grass and the sand texture if there was going to be something like this in FlightGear.
If you can come up with good hires GPL-compatible textures, I'll see what I can do with it (might not be immediately, though...).