I can't speak definitively for any specific project, but I gather much of this is due situations like this:

SoC Mfg designs a new chip, heavily leveraging IP from an old chip. Since some old IP is still around, and already (mostly) has kernel support SoC Mfg, patches the existing code to work on the new SoC, publishes the bin/source as an SDK, and declares their work done. (The linux-meson/sunxi geeks will eventually clean this up and mainline it anyway).

In the mean time, Board Mfg needs to get a product out the door, so they start with the SoC and SDK, and find that a bunch of the work done in the SDK is really specific to SoC Mfg's reference board, and requires a few tweaks, so they hack on the SoC Mfg kernel tree until the device mostly works, declare a success and publish a board SDK. (Those same geeks from above will handle the mainline work if they want to run a mainline kernel anyway, so why should Board Mfg bother?)

In some cases, this cycle gets to repeat yet again when the board gets OEM'd into a half dozen devices, each requiring slightly different kernel tweaks.