It is worth noting that checkpoints in Plasma XT: Plasma Cash with much less per-user data checking can be thought of as state roots, which are tied to the transaction roots cryptoeconomically.
One problem with using only state roots is that you no longer have an O(1) exit game. The plasma operator could commit N state roots, 1 of which is an invalid state transition (with respect to the previous state root) but withhold all the data, and then exit with the last state root. The proper coin owner cannot construct a fraud proof because the necessary data is unavailable; hence the exit game must take O(\log N) steps. I believe this is necessarily true in your design, but it seems like a description of the exit game has been omitted.
Also, small nitpick: I think you mean “sparse merkle tree” since “merkle tree” as the term is commonly used doesn’t support succint non-inclusion proofs.