Hello. If we present all the data in tabular form (in rows are pixels, in columns are channels), then the k-means algorithm works quite well. You can use the F-mask for images of Sentinel-2, or use the library of Sentinel2hub. These algorithms create masks with which you can cut out unnecessary areas.

http://pythonfmask.org/en/latest/index.html#downloads

https://github.com/sentinel-hub/sentinel2-cloud-detector

https://medium.com/sentinel-hub/improving-cloud-detection-with-machine-learning-c09dc5d7cf13