I think it'd be much easier to use spread-spectrum techniques. With enough coding gain the signal can even be well below the noise floor and still reliably reco

I think it'd be much easier to use spread-spectrum techniques. With enough coding gain the signal can even be well below the noise floor and still reliably recoverable. (This is the spread-spectrum equivalent to sending a very narrowband signal and receiving it through a very narrowband filter.) Generating a DSSS waveform is extremely computationally cheap; receiving one is less so, but probably still negligible amount of cpu time on a modern processor.