How best to do CCA between gridded observations and model output with multiple predictors with disparate numbers of grid points?
I'm doing CCA between observed tropical SST and the T42 observations. I'm comparing to CCA with model outputs and observations. First, I'm taking model precipitation to see if it does as well as SST. It seems not.
I'm also trying point-wise variance inflation with both SST and model precipitation.
I'm looking at the impact of using tropical model precip., regional precip and vertical shear. Each separately seems to effect a different region but if I stack them, the results don't improve as I expect.
Possible reasons: (1) using too few EOFs. Some of the "optimal" truncations are at my upper limit of 8. (2) Very different numbers of grid points. The global tropical pacific seems to dominate the others. Weight? Do stacking after EOF decomposition?
I'm trying more EOFs (16) and weighting by the number of grid points.
I was focusing on truncation of predictor EOFs but predictand EOFs may be and issue too. Suppose the two predictors have skill to two distinct areas but to separate them requires many predictand EOFs. Would not be an issue with single predictand.
Another approach is to to pre-filter, do PCA on the predictor fields separately. That way the number of degrees of freedom of the different predictors is more likley to be comparable. Cross-validation becomes more complicated.
Another alternative is to do CCA on each field separately, and then put them together again, "ensemble CCA."
Posted by mktippett at September 25, 2003 09:34 AM