jbednar
commented
6 years ago I'm sure that something like that could be done, and probably has been done by someone. To decide which of the two main groups is water, you can probably use some measure of variance -- sometimes the water is darker than the land, sometimes lighter (as in this middle picture), but in each case (due to mixing) the water is more spatially uniform. So once there are two groups, you could run a local spatial homogeneity index on it to guess which is water. If both groups are homogenous, as in some desert areas, I guess you could pick the group closer to blue as the water one.
In any case, is getting this initial guess the limiting step in your workflow? For hydrology modelling it seems like later steps still need a good bit of manual cleanup and tweaking, so getting a good initial guess may not be worth going to a lot of effort and complexity.
kcpevey
I was just brainstorming this morning and it seems like there might be a way to let GrabCuts make a semi-intelligent automated first guess for coastline delineation. I think one reason that GrabCuts works so well is there tends to be a fairly dramatic distinction between land and water. Admittedly, I have minimal knowledge of image processing and how GrabCuts actually works, but... is there a way that we could sample the RGB values for the whole image, create a spectrum of the results, grab the two main peaks and feed the RGB values representing those peaks to grabcuts? This would only work if the image has a large distinction between land and water and if the colors are fairly uniform across each type.
Examples where I think that would produce reasonable first guesses:
Might work less well for color-varying waters:
