I began with 24 variables for each of the 4 veg types.  Since we have
maps of both mean and std dev, this is 48 maps.  With all 4 veg types,
that is 192 variables.

Then I realized that some of the driver variables are exactly the same
for all four veg types.  I identified all of the variables (mean and
stddev) that were exactly the same over all 4 veg types, and dropped
all but one copy of them.  Ppt, tmax, tmin, tday, tnight, vpd swavgfd,
and par are the same for all veg types.  This got us down to 8 x 2 = 16
x 4 = 64,  64 - 16 = 48, 192 - 48 = 144 variables.

Then I found that dday, srad, and rhum are also the same for all veg
types.  Dropping all but one copy of mn and std for these resulted in 3
x 2 = 6 x 4 = 24, 24 - 6 = 18, 144 - 18 = 126 variables.

Then, I identified all of the variables (mean and stddev) that were
exactly the same over more than one veg type, and dropped all but one
copy of them.  Par turned out to be the same as sfwd, so 2 more were
dropped, (124).  Potential evapotranspiration mean and stddev were
identical for C3 and C4 grasses, but different for the two tree types,
so I dropped 2 more (122) and retained a pet-grasses variable.  Coarse
woody debris mean and stdev were all zeros for C3 and C4 grasses, so I
dropped both of them - 122 - (2 x 2) = 118 variables remaining.

So the elimination of exact duplicates reduced the number of variables
from 192 to 118 variables.  Considering that each of these represents a
1 sq km map of the US with 7.8M cells, we wasted a lot of disk space
computing and storing multiple copies of all of these.

Even after dropping all of the exact duplicate variable sets, there was
still so much multicollinearity among the remaining variables that SAS
retained only 113 positive pca factors from 118 variables.  I
considered doing a simple pairwise Pearson correlation-based
elimination, but that would be subjective in terms of the selection of
which of the paired variables to eliminate, and would ignore
multi-collinearity across more than two variables.  But I couldn't
figure out how to use the PCA factor loadings themselves as an
objective basis for reducing variables to a parsimonious subset.

I fooled SAS proc reg into doing a multicollinearity diagnostic by
setting up a "dummy" linear regression model, which I ignored.  I did
this multicollinearity analysis in two steps: first for the common
variables and each of the four veg types separately (5 analyses), and
then once with all remaining variables together, in order to detect
multicollinearity across the common and veg type sets.

From the first set of separate multicollinearity analyses, I dropped
daytime temp (correlated w max temp), and night temp (corr w min
temp).  From C3 grasses, I dropped gpp (corr w nppmn, nppsd, pLAImn)
From C4 grasses, I dropped gppmn, gppsd (corr w nppmn, nppsd, pLAImn).
From decid broadleaf, I dropped litter carbon mn and sd (corr w veg
carbon mn), and I dropped gpp mn and sd, and pLAI mn and sd (corr w
nppsd).  From evergreen needleleaf, I dropped litter carbon mn and sd
(corr w coarse woody debris carbon mn), I dropped pLAI mn and sd (corr
w npp mn).

A total of 18 variables were dropped in the first phase of
multicollinearity analysis, taking us from 118 down to 100 variables.
I tried to drop variables symmetrically, removing the mean and std dev
pairwise.  I also tried to be symmetric across veg types, but this was
not always possible.  For example, evergreen needleleaf gpp values were
not cross-correlated, but gpp values from the other three veg types
were, and were removed.

In the second round of multicollinearity diagnostics, all 100 remaining
variables were included in a single analysis to check for excessive
correlations across the common variables and the four veg types.  I
removed C4 days since rain mn and sd (corr w C3 days since rain mn and
sd, respectively).  I also removed decid broadleaf leaf water potential
(psi) mn and sd (corr w C3 psi mn and sd, respectively).  Other
cross-correlations still existed, but were not as severe as these.

This brought the number of variables down from 100 to 96 variables
remaining in the parsimonius set.