Hi! So one interesting thing that I noticed today. 

I have a rule-set for applying the RF algorithm based on input Sentinel 2 imagery. 

I load R G B NIR and SWIR channel, calculate NDVI, NDWI, SAVI, NDSI, NDBI and then segment the scene, apply samples, and start the training process...

Here is the problem. For some reason, the generated index layers, at some spots, have "nan" values even though I placed the "-1" for undefined areas. Due to this, the algorithm refuses to train the RF. Any idea how to void this?