Hi Alex - the scanne/site differences would be a problem for any analysis, not just PLS. Having said that, if you processing steps correct for it, you should give it a try and definietly include "site" as a factor (covariate) to see if it plays out still.
However, do keep in mind that the values in your T1w images mean nothing - so you should not submit T1w images to PLS (or any analysis).
You can submit GM probablility maps to PLS, with the caveat that the preprocessing pipelines will have different effects on the GMprob calculations.
If you have quantitative T1 measures, the scanner differences may be less problematic, but it doesn't sound like that is what you have.