Is POISSON the correct family for my model? my model is:
model1<-glmer(meanNOspp~VCI+treatment*size+(1|site), family=poisson) 

I get more than 50 warning messages like this:
In (function (fr, X, reTrms, family, nAGQ = 1L, verbose = 0L,  ... :
  non-integer x = 0.500000

meanNOspp is mean number of species, am I getting this error because poisson is not the correct family for this type of data? can someone help me? 
 A: Poisson regression is appropriate if the dependent variable is a count (i.e., $y \in \mathbb N_0$). However, the error message

non-integer x = 0.500000

indicates that your dependent variable contains real numbers.
You might consider using the total number of species instead (not the mean).
A: I assume that the mean number of species cannot be an integer. Poisson regression is used to model count data and it is maybe your problem.
Maybe, a log-Normal regression can be more adequate instead of your Poisson regression:
model1<-lmer(log(meanNOspp) ~ VCI+treatment*size+(1|site)) 

One can have two equal means, for two different totals. I think to model the mean and the total are two different things, especially if it is not the same number of staff for each individual mean meanNOspp in the regression. So it depends on what you want to do.
A: Another solution is to use a modified Poisson regression for the totals. This will be correct if a Poisson model is correct for the individual counts. It would probably be better then to model the individual counts, if you have them). Suppose the $N$ (number of obs) behind your mean (meanNOspp). Multiply that with its $N$ to get the total, and then include $\log N$ as an offset in the model equation. Then it should work. 
Another advantage with this approach is that, if it turns out that
the poisson assumption do not work, maybe you have overdispersion, then you can just use the same mean specification with another error distribution (such as negative binomial), as long as you keep the log link function (or you could use a quasipoisson family).
