Hello,
I am starting to use the RF supervised classification offered by SNAP and I would like to understand the exact meaning of “the feature importance score” that is displayed at the end of the cross validation file.
Thank you!
feature importance characterizes how often a raster layer was used to split the single trees in your random forest classifier. So this means which ones had the most important information which was used to discriminate your training samples. Rasters with low feature importance can not (or only to a small degree) be used for the discrimination of the training classes.
On the estimation of good and bad features (=input rasters) of RF classifiers I can recommend this page:
http://alexperrier.github.io/jekyll/update/2015/08/27/feature-importance-random-forests-gini-accuracy.html
Or a bit more statistical:
Thank you vey much.