Normally, this should not happen, but it is not uncommon to see very small negative numbers due to numerical errors and rounding off.
This is further made worse if the number of samples is too small. Usually, when using sobol, you need thousands of parameter-sets for your sample size. so your sample size should be very large, a quick check will tell you that you need at least 1500 samples or more for your run. If you have a slow running model and cannot run all that, then you may want to run using the LH-OAT with at least 400.
This page lists references you may find useful:
https://celray.github.io/SWATPlusToolbox/runSensitivityAnalysis/#sobolWith Kind Regards,
Celray James CHAWANDA