We propose a simple filtering method to improve classification precision of images.

It focuses on classification probabilities and discards low confidence objects.

It successfully resolved the in situ fine scale distribution patterns of plankton.

It requires very limited manual identification work.

It is applicable to all machine learning methods.