I am using kmeans clustering (ee.Clusterer.wekaKMeans) to classify a Landsat8 image using all 11 bands. I was wondering if it is possible to see what bands are relied on most to create the clusters (i.e., in what bands are the greatest differences, that have been used to differentiate clusters?) Because I am using unsupervised classification, I want to give the kmeans clusterer all the information possible (all 11 bands), but I also would like to be able to dig into the clustering process to see how those bands were used to create the resultant clusters.

Specifically, I want to identify water. Water is often identified using the NDWI ratio (green and NIR bands), but kmeans using these two bands produces a different result than kmeans using all 11 bands. So I want to know which other bands are influencing the clustering!

Is this possible using GEE's clusterer tool?


The nice thing about k-means is that the results are linear. That is, you can simply take the mean of all the values in each cluster to find the cluster centers that were (ultimately) used. Assuming all your bands are scaled the same, you can then just find the distances between all the cluster means.

I'm not sure that answer is actually going to help you though. k-means doesn't make a "decision", it just puts things into the closest cluster and recomputes new centers from the resulting groups.

See this image and site for more information:

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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