Texture descriptors based on oriented gradients?

I asked a previous question yesterday. The post can be found here. The post has many details that I'm going to reference below.

I have a k-means segmented image that represents a terrestrial site. The site consists of different landforms. The details of this can be found in the above post.

I made an interesting discovery yesterday when playing around with some of the data we have. I took the RGB image (which I explain in detail in the above link), where the 3 channels are pixel based features, and calculated the magnitude of the oriented gradient at each pixel in each channel. I mapped the maximum magnitude in all three channels and here is what it looked like: http://d.pr/i/fS8f . As you can see, the gradients of the craters are much stronger and thicker as compared to the ridges. I mapped their corresponding angles and this is what I got: http://d.pr/i/4C1L .

My goal is to essentially calculate a feature vector for each segment in the image so I can put it through a classifier. It seems to me that texture analysis would be apt for this, but I can't seem to find any descriptors that use oriented gradients.

I am familiar with HOG and some of its variants. Obviously, HOG wouldn't work for this problem, but I'm looking to do something similar.

I am also familiar with the Haralick features and their implementations. I'm not sure if they would be ideal in this scenario however. The reason why I'm keen on using oriented gradients is because they allow for distinction between landform boundaries.

I'd appreciate any thoughts or inputs on this.

EDIT: I includes this in my previous post, but for easier access: here is the ground truth: http://d.pr/i/Z4sI . Here is the RGB image, where the three channels are pixel based features: http://d.pr/i/6n8K . Here is what the k-means segmented map looks like: http://d.pr/i/FR1O .