Flood-filled elevation

I sat down to eat in the Seattle airport last Thursday with a guy who unsurprisingly turned out to be a programmer. We talked a bit about Gaussian blurring; he was saying that for filter sizes much above 5x5 it was quicker to do the Fast Fourier Transform and do the blurring in frequency space. I spent some time trying to figure out how I might be able to do that. The pegging of the river to zero complicates things. Since I'm doing a bunch of steps it ought to be possible to replace them all with a single step with a really big kernel, but since I'm re-pegging the river in between each step I am not sure that would translate to the same thing at all. Maybe I could do half as many steps with a slightly larger kernel (in frequency space) instead.

On the plane back from Baltimore yesterday I tried a quick experiment. Because the Gaussian blurring is so slow I thought that if I could come up with a better first approximation I could get away with doing less of it. I tried flood-filling initial elevation data, working outward from the river. Each node is slightly higher than the highest of its neighbors that have been defined so far.

Unfortunately the quality of this is pretty bad:



Here are some resulting roads:



Overall I'm not sure that having roads primarily follow or cross the gradient is what I want. I may dive in and try to understand the L-system-based approach that Mueller's original paper describes. The writing is impenetrably bad, as is typical for academic papers.

I'm starting to feel like the terrain generation is eating up too much time. I'll probably just settle on something for now so I can move on to other aspects. I need to figure out if and how having a large city will be fun, and I need to work on plugging buildings into lots, generating patrol routes, street lights, and so forth.