This is little more than a comment on the other two answers, but I'd like to emphasize that pinning root certificates is also tricky and dangerous, if less dangerous than pinning intermediate certificates.
As mentioned, CAs are likely to switch intermediate certificates unpredictably and without notice, as they expire, or for other operational reasons.
Less often but for the same reasons, CAs will also switch which roots they use for new certificates or in the default chains provided for download. They are likely to give advance notice, but even if they do you probably have to regularly check their news and support websites to find it.
Older CAs have accumulated eclectic tangles of root certificates that were obtained through mergers and acquisitions, or issued to upgrade to more modern cryptography, to replace older certificates that will expire, to use for different categories of certificates (EV, S/MIME, government, etc.), or for other reasons.
Newer CAs are likely to have a few roots generated on their own, cross-signatures from one of the aforementioned older CAs, or maybe an entire root acquired from an older CA.
It can be difficult to make sense of which roots are relevant to you, which TLS clients and versions your clients use, and how their path building, pinning and validation algorithms work.
Your CA may have detailed documentation about how their graph of trust anchors fits together. They may provide advice about pinning. They may provide promises. They may not be able to keep them. They may work with you if you get backed into a corner by your choices or theirs.
As an example, Amazon lists 5 root certificates (1 acquired from another company). There are cross-signatures. Building a graph would probably find about 10 roots involved total. A misguided AWS certificate customer could probably pin 1 of about 5 different certificates and have something that would work today and explode later. I have not checked their documentation. (Also, some of Amazon's own services use other CAs.)
As further context, Ryan Sleevi wrote two posts about the complexities of path building.