K-Means Clustering with Spatial Correlation

Hi,

I'm wondering if there is any way to incorporate spatial correlation into a K-Means clustering algorithm? I'm working with a mining dataset that is made up of drill holes (strings of data) with 1.5 meter sample intervals. Each sample has measurements for 63 elements. Of those, about 8 are of interest to me. Regardless, it's still a highly multivariate problem.

Each sample looks like this:

1. x y z au ag c as v u ti ... 2. x y z au ag c as v u ti ... 3. ...

I have over 200,000 samples that I can use.

We have several different rock types on the property and each has a "somewhat" unique geochemistry signature. For example:

Rock Type 1: depleted in U and depleted in V Rock Type 2: depleted in U but enriched in V Rock Type 3: enriched Ti and enriched V

I've been playing around with the algorithm and it does a good job of correctly classifying rock types. The problem is that I'm completely ignoring spatial correlation at this point.

Can spatial correlation be incorporated into k-means?

Is there a better clustering algorithm for what I want?

Thanks,

Cole

Answer 1

[Jeremy Kemmerer]

It sounds like you would like mining samples that are physically collected closer together to be clustered together in your analysis. If so, can you include the spatial position where the samples were collected as features in your clustering, along with chemistry information?

Using “ kmeans ”, this would involve adding an additional few columns to your data matrix X .

Also, you may need to scale your features so that none of them dominate your clustering distances. One possible way to do this is using the “ zscore ” function.

For more information on this function, please refer to our documentation:

http://www.mathworks.com/help/stats/zscore.html