I am not sure if I can formulate this such that it is clear. :)

I have around 700 80x80 matrices, where each matrix shows some weather event (a matrix has continuous entries from 0 to 60). Now I would like to cluster these events into several event types.

I already used a very simple method where I transformed each matrix into a vector, and did kmeans on these vectors with euclidean distance.

But is there no better method, that uses somehow more information (e.g. neighboring pixels)? It seems to me, that I loose alot of information if I cluster in such a fashion.

So maybe it is a general question:

How to cluster matrices, where each entry has a "meaning" in 2 dimensions.


1 Answer 1

  1. specify a distance or similarity measure that takes neighbor pixels into account. For example, a quadratic form may be helpful.

  2. compute a distance matrix, 700x700

  3. run hierarchical clustering

  • $\begingroup$ Thank you for this suggestion. Do you know of good distance/ similarity measures for example radar images? $\endgroup$
    – thezebra
    Jun 5, 2015 at 12:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.