I'm still at the design stage for my terrain system. A lot of the other work on my flight sim has gone very well, but I've been battering around ideas for the terrain for months now without coming up with a design I like.

I can generate whole planets worth of random terrain no problem, but generating terrain based on real world locations is another whole barrel of triangles.

The input data I already have is in the form of two textures per map, Each map is something like 1000 miles by 1000 miles.

The first texture is a height map. Standard stuff, but at a resolution of 1 pixel = 400 square meters .... not very useful as it stands.

The second has type information in it. it is broken down into 16 terrain types,(4 bits) (water, grass, mountain, etc.) and two bits for roads and railways.

My original plan was to generate a terrain patch for each pixel of the source data, but when you get to above 35,000 feet the number of terrain patches visible becomes ridiculous.

So I'm thinking I am going to need a terrain patch to be about 10 miles square. These will have to be generated from the original data and include a really good CLOD system. Which one is up for debate.

To get the level of detail I need from this is going to be a major challenge. I need to have terrain that looks good when you are stood next to your aircraft, and still looks good at 35,000 feet. Not easy.

So I am looking for guidance from the folks here.

Anybody seen any good papers worth reading?

Anybody got experience writing terrain systems for flight simulators and is willing to share knowledge?

At the moment I have a dozen possible solutions in my head, none of which solve all the problems, all of which solve some of the problems.

I keep starting writing code, and then stopping what I am doing because I realise it won't work.

http://vterrain.org/LOD/Implementations/ is a great site for terrain rendering stuffs. Also since much of the work is academic, you can find good links on citeseer.

Having just watched "The code" on netflix, I can offer what was done by boeing in their sims(In the 80's mind you). They used random heights, and what appeared to be midpoint displacement or another fractal function to create the fine detail. The run-time looked like ROAM, He gave some detail as to the vertex count, but i can't recall of the top of my head. You might try this to fake increased heightmap resolution when creating your patches. Or get a better data set. I'm currently learning how to stream patches myself and using the SRTM 1 arc minute data.

What do you expect the visible distance to be? knowing this would let you figure out what the max area that would need to be in memory at once. From there you could start to figure out where the trade off goes for increased number of patches ( with less LODS) vs increased patch size( with more index buffers) but my instincts tell me that 10 miles x 10 miles is large for a single patch.

The size of the terrain patch is not so much of an issue. The key things you want to look at is the LOD you apply to each patch.

You have data points every 20m (Using your 1 pixel per 400m2). You can use this resolution close up but at 35000ft do you need to use one point per 20m, the further your patch is away the less detail you need.

If the terrain patchs carry LOD details you can then use whatever LOD calculations you want to apply to know if the patch needs a rebuild at a different LOD than it currently has. This way you can carry a massive number of terrain patches and reduce the render workload.

Yes guys, my initial plan was to have each terrain patch be a single quad at the lowest resolution. However after some experimentation, I think the best I can get away with is 4 triangles, so I can give the patch some form.

The reference LOD will be one of the higher resolution LOD's, but which one depends on my initial patch size.

So my LOD design looks something like this.

Lowest res : 4 triangles

....... N LODS down sampled from reference data

Reference : Real world data

........ N LODS up sampled from reference data (standard CLOD of some form)

Best : Up sampled and mixed with noise

Sounds simple when you write it like that

there should be a way to look this up or figure it out.

Yes, I remember looking this up some time ago. It's based on the curvature of the Earth and altitude of eye point. It's vast though. If I am going to support modern planes I need to be able to go up to 60,000 feet.

step 1: figure out how to do it. step 2: implement. imagine writing every line of code in your head before you sit down at the computer:

Do that all the time. Even dream about code.

It's things you don't expect that bite me. Usually in the graphics card. I'll use this technique, only supported on Nvidia cards..... ballacks..... I'll use this one , not supported on Nvidia cards .....

Wall. Meet head.

What I did with random planets is use latitude and longitude as inputs to a noise function. This gives you completely seemless terrain even at the poles.

As you decrease in altitude I just increased the resolution of the noise function, so you get more detail but without an increased vertex count.

It's like sliding one of those frames full of nails across a terrain.

This turned out to be ok at the time, but as you get nearer the base frequency of the noise functions, things get very bad. Verts dance around with even the smallest change in view vector. Not good.