Sampling at 48kHz seems like overkill. If you record at a lower sample rate (e.g. 22k) you will still be able to detect the bubbles, but the files will be about 4 times smaller, and as an added bonus it will function as a sort of high cut filter, which may help with stray high frequency noises or thermal noise (to some extent). Additionally, you can apply a band pass filter around the frequency of the popping, if you aren't already cleaning it up before processing. Given a clean recording, you can trim the silence and still have full sample data for re-running algorithms on later if you preserve the timing metadata.

You could match the spectral signature to a known sample to distinguish between bubbles popping and other types of noises, if there were any, but the recording conditions seem good enough not to warrant it.

You could also possibly use an array of two, three or more mics and do some fancy triangulation to get a 3D map of the pop locations and better distinguish between individual pops, but I doubt how useful that would be besides a few cool visualizations.