CNN where pixels are constituted by large, potentially sparse vectors I'd like to apply a CNN to a problem where the image is essentially a matrix representation of a geographical map where matrix indices correspond to the locations of buildings and roads.  Each location will have various features associated with it which would be represented by a sparse vector corresponding to a pixel.  The reason I am applying a CNN to this dataset is that I think shape data is important to phenomenon I am trying to learn, however, I've only ever encountered vectorized pixels in the context of representing rgb values for colorized images.  
My questions are:
1) Is this a sensible approach to the type of data I am trying to learn from?  Are CNNs a good fit or are there better options?  
2) Are there any good examples of this sort of semi-pictoral approach working?  Are there existing applications of CNNs where each training datum consists of a matrix of spatially arranged locations associated with a set of non-pictoral information, such as location-type (park, road, building, etc.)?  
 A: Yes, this is a sensible approach. 
I have a friend who researches search using deep networks, and this is part of what they do. Their lab has a few papers, such as this one on per-map algorithm selection:

We attempted to select the best algorithm for a problem
  by finding problems with similar start and goal states on the
  same map and then combining the algorithms found to perform
  well on them. The results were worse than the those
  of the “prior” method likely due to the poor octile-distancebased
  similarity metric (e.g., close-by start states separated
  by a wall would be considered similar). Future work will
  investigate better similarity metrics.
We also represented a problem as a bitmap image of the
  video-game map with the start and goal states shown on
  it in color. We then trained deep convolutional networks
  AlexNet and GoogleNet as included in MatConvNet framework
  (Vedaldi and Lenc 2015) to predict the six parameters
  of a suitable algorithm for it. The results were slightly
  worse than those of the “prior” method. We are currently
  investigating using convolutional deep networks to predict
  algorithm suboptimality on a given problem, represented by
  a bitmap image. 

