The importance of the warning depends on the data in each volume
you care about. The voxel values that contribute to a high misfit
average are those very close to zero, which are often of the least
interest.

But if you have a single huge value in a volume, it can throw off
the numerical resolution of the rest of the volume. Keep in mind
that signed shorts have only 15 bits (for the unsigned part).
There are only 32,767 possible numbers (call it 30000), plus a
scalar.

So if the largest absolute value in your dataset were 300, then the
scalar would be 1/100 (so 300.0 would be stored as 30000 times the
scalar). The smallest non-zero value one could represent like this
is 0.01 (1 times the scalar). So numbers less than 0.01 would be
truncated to zero, and those affect the misfit the most.

Now suppose there is one really large value, like 60000. So the
scalar would be 1/2, meaning one could not tell the difference
between 7 and 8. Anything less than 2 would affect the misfit most
significantly.

- rick