I'm using Google Earth Engine to classify mangrove forests from Landsat imagery to calculate extent. Let's say I perform an unsupervised K-means cluster analysis on a Landsat scene to identify general patterns and themes in the spectral bands.

After grouping the clusters together into several relevant classes (e.g. mangrove, other_vegetation, soil, urban, water), would it then be possible to use this classified image for training a supervised classification algorithm such as random forest?

The idea is to use a stratified random sampling technique (~5000 points) to assign the training points within the classes obtained through the cluster analysis.

Can anyone see a flaw in this method and would it increase the accuracy of the final classification?

TLDR - is it possible (or advisable) to use the output from an unsupervised classification technique to train a supervised classification algorithm?

  • Well, in theory you can mathematically decompose k-means into random forests so, one would wonder what the point would be. – Jeffrey Evans Oct 7 '18 at 2:37
  • Thanks for the reply. Can I ask; what do you mean by 'decompose k-means into random forests'? The reason I asked the question is there are some references in the primary literature that use a hybrid approach to classifying mangroves using the methods I outlined above, but I'm at a loss as to how to apply them as they don't go into enough detail. I suspect my lack of remote sensing knowledge is the issue here. – Samual Chance Oct 7 '18 at 4:18

Your Answer

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

Browse other questions tagged or ask your own question.