mixed effects model in lme

I believe you have three levels of variation:

between plots assigned to the same fire/water treatment
between subplots assigned to the same fire/water/seed treatment
and between measurements (fire/water/seed/year)

Your models (and the previous replies) appear to be ignoring the repeated 
measurements on the same subplot.  You will probably need to explore various 
choices of correlation structure among years (e.g. Compound Symmetry = 
split/split plot, ar(1), ...)

One key is that tests of split plot factors have 270 error DF (denDF).  You 
only have 120 subplots.
  
Also, I suggest you retain at least all the fire/water/seed interactions in 
the model.   Each combination of fire/water/seed represents a specific "thing" 
(some call it treatment, but treatment has many different meanings) done to a 
subplot.   The full model with all these interactions corresponds to a cell 
means model.  Marginal means (i.e. main effects) correspond to averages of 
cell means.  If you drop those interactions, you are assuming that the true 
interaction is zero, so those dropped interactions only inform you about the 
error variation.  I believe this is somewhat dangerous because tests of 
interactions have low power (i.e. lower than tests of main effects).
I recognize that there are many other opinions here.  

The above argument does not apply to interactions with year because year is 
not randomly assigned to a subplot or plot.  Many different opinions.

I prefer to fit models like this in SAS, because it is much easier to get 
meaningful estimates (i.e. not just tests) in SAS than in lme.

Best wishes,
Philip Dixon