Hi, you can try to remove the replicate from the analysis and verify if the problem is solved. However, an analysis with only 2 replicates could be insufficient, depending on remaining df.  Then if you would like to save some data of the deviating replicate, you can try to remove the aberrant data and try again.  Another solution could be to perform arcsine or Log or Square root transformations to the entire block of data.  If you use ARM software, this is automatically done.

I think 20% cv under field condition is not too bad. But told us the third replication is not representative. If you have more site, you drop the third replication that has problem. If it is only 2 replication, it is not possible to make ANOVA. Please check cv for each treatment. However, if your research is for validation of models , two replications will be sufficient!

I think first check if there are outlayers, and then you can transform your data; but generally, 20% CV is not a serious problem in field experiments.

Go for the out-layers, through statistics techniques you can delete upto 20% of the dispersed data point that will help in minimizing the CV.