I'd use a morphing LOD solution with tiles, e.g. vertex indices per grid point:

LOD 0:

01

23

LOD 1:

012

345

678

...

To fit LOD 1 tile top edge to LOD 0 neighbour bottom edge, you lerp LOD 1 vertex 1 to LOD 0 vertex 2 (or 3), depending on view distance.

The good is this works with any geometry, so caves, buildings, whatever. And it solves the popping problem.

The bad is you need to load 2 vertices instead one, at least near the border - quite acceptable if your tesselation is not too fine.

(The technique has a proper name but i don't remember, and i don't know if hardware tesselation makes sense here but guess not.)

It sounds like a quad-tree model essentially, but with a lower bound, if that makes sense

7 minutes ago, JoeJ said:

The technique has a proper name but i don't remember, and i don't know if hardware tesselation makes sense here but guess not.

I think it would make sense. CPU tessellation for the low-frequency and GPU tessellation for the high-frequency data.

EDIT: Also, I'm still wondering how that compares to geoclipmapping

'Geomorph' is the name i was looking for. Here is a picture if you scroll down: https://ig88.azureedge.net/8-2htmldocs/index.html

The idea is neither new nor much different to typical terrain LOD approaches, but it is more flexible, lifts height map limitation, and most work is done by a preprocessing tool.

Edit:

A simpler idea would be to use tiles without any morphing but extrude skirts from their border edges downwards, so the skirts fill the cracks at LOD mismatch.