I've been working on a revised version of the development projection images that I did a while back. I wrote a long post about it on Stat Chat that has the full images and some commentary on them. Here I wanted to explain a little more of the technical background on how I made them.
How Is This Constructed?
The model starts with a random value in a Poisson distribution that simulates the likelihood of development. This distribution captures the fact that some parcels will develop seemingly at random. As different circumstances change, more and more areas will develop until finally the vast majority of land in an urban area will develop. After that, some undeveloped plots will remain for a while and a few will simply never develop.
The values in this raster are then raised by adding a score based on the driving time to major employment centers. This is because most development is and has been automobile-driven. The shape of an urban area is highly predictable based on the driving time to an employment center. Lastly, I reduced the likelihood of development based on the slope of the terrain and subtracted all national and state parks, wetlands, military bases, conservation easements, and local parks that were in some way preserved. I then split this model up into planning district commissions because they roughly encompass metro areas that expand outward from a core.
The density that I used was the same density of development in the district that I was adjusting. So if the developed areas of a planning district had a density of 1000 persons per acre, that's the density I used for the predicted new residents also. These densities are likely to be off because of several factors. They could be too high because, while areas have been gaining population over the last 50 years, the existing population has also been decentralizing and new residents are likely to buy the lowest density homes on the periphery. On the other hand, they could be too low because, as areas grow in size, they become denser and each new person's marginal amount of developed area is a little smaller. Additionally, the trend in recent years has swung the other way, as I pointed out earlier in the post. For lack of a better model, I considered it a wash and stuck with the existing density.
Cities and driving distances:
Slope:
Conserved land:
Growth likelihood raster and planning district boundaries. Some counties are shared by planning districts - I had to assign these to one district and deduct their population numbers from the other.
I'd be very interested to hear of ways you think this could be improved or additional factors that could be added in. One I thought of was buffering around the Chesapeake Bay and other bodies of water to account for the attraction of living by the water. Another is doing a more comprehensive service area just around the roads to get more of the development aligned with roads rather than "speckled" about. The problem with that is that new developments come with new roads and you don't know where they're going to be.
