Agree with @Nicholas_Bokulich on the key questions: Are your primers on your reads? And what is the length of the region you are amplifying?
This is because the cause is probably:
(1) Unremoved primers on reads -> spurious chimera detection -> few reads making it through pipeline.
OR
(2) Truncated reads not long enough to overlap in middle -> failed merging -> few reads…
See also the discussion here: Lost of data with dada2