After some discussions here. The zeros will not have an impact. But the large number of observations might lead to ensemble collapse. The resampling as used before might solve this.

However, there is a more sophisticated way available in ERT that we could test:
I got the suggestion to use: https://fmu-docs.equinor.com/docs/ert/reference/configuration/keywords.html#std-scale-correlated-obs (only accessible inside equinor - but the documentation is open, just check the github pages for ert).

Just came a release message for ERT 2.15 where the STD_SCALE_CORRELATED_OBS workflow is said to be deprecated and that it is recommended to use the MISFIT_PREPOCESSOR instead, see https://fmu-docs.equinor.com/docs/ert/reference/workflow_jobs.html?highlight=misfit_preprocessor.

The manual doesn't mention the MISFIT_PREPROCESSOR workflow, so I'm not sure yet what exactly to use...

Update: Just realized that ERT 2.15 is in testing at the moment and is going to be released soon. Forefront of technology @anders-kiaer - as usual.

Due to the large number of observations the new ERT storage doesn't work very well either. This issue therefore has an impact on #265

@edubarrosTNO Will look at resampling options for the obs file creation.