People have sort of brushed on this idea before at http://www2.cr.chiba-u.jp/symp2005/documents/Postersession/p002_Ketutwikantika_paper.pdf, but it doesn’t look like there’s any significant progress towards actually having done it yet. I bet you can use vector quantization to do this – feed in some urban, suburban, and rural areas, extract texture descriptors using VQ, and feed them into a classifier.

This is one I’m going to pursue. The barrier to entry is very low and fits nicely into my area of expertise.