Capturing visual patterns for each segment on a segmented site

Hello,

I'm a high school student working on a project at my local university. Our project has some computer vision aspects. Unfortunately, the professor I'm working with is not an expert in computer vision which has kind of left me in the dark on those parts of the project. I'm just looking to get a small start and some suggestions.

I'll start by explaining the problem.

We're working on improving the landform classification accuracy of terrestrial surfaces. Essentially, our sites will consist of many different landforms such as crater walls, ridges, plateaus, etc. Here's a ground truth for what one of our sites looks like: http://d.pr/i/Z4sI . We're trying to do this on a segmented map of that site. The segmented map was done by a k-means algorithm and consists of about 7000 segments. Each landform that you see is made up of many many segments.

What we want to do is exploit the fact that different landforms have different shapes. A crater wall is circular, ridges are diagonal and linear, etc etc. The best way to do this, we thought, was capture visual patterns around each segment. We need to extract a feature on the segment level. We hope that the visual patterns around, for example, all crater wall segments would be similar when compared to the visual patterns around all ridge segments.

The whole point is to relate the segment back to the shape of the landform.

Do you have any idea, suggestions, or papers I could look at for this.

Thanks

EDIT: Just wanted to include this to ward off any confusion. The kind of data we're working with is called a DEM, digital elevation model. The data stores the elevation of every pixel at that specific point. If I computer the magnitude of the gradient vectors of each pixel, this is what it looks like: http://d.pr/i/TfXF . Also, here's an RGB image where the 3 channels are 3 different pixel based features calculated off of the DEM: http://d.pr/i/6n8K . The k-means algorithm was done on the RGB image to produce the segments. You can contrast that with what the ground truth is supposed to look like.

EDIT: This is what the segmented map looks like: http://d.pr/i/FR1O