We discuss the use of reverse time migration (RTM) with dense seismic networks for the detection and location of sources of atmospheric infrasound. Seismometers measure the response of the Earth's surface to infrasound through acoustic-to-seismic coupling. RTM has recently been applied to data from the USArray network to create a catalogue of infrasonic sources in the western US. Specifically, several hundred sources were detected in 2007-2008, many of which were not observed by regional infrasonic arrays. The influence of the east-west stratospheric zonal winds is clearly seen in the seismic data with most detections made downwind of the source. We study this large-scale anisotropy of infrasonic propagation, using a winter and summer source in Idaho. The bandpass-filtered (1-5 Hz) seismic waveforms reveal in detail the two-dimensional spread of the infrasonic wavefield across the Earth's surface within approximately 800 km of the source. Using three-dimensional ray tracing, we find that the stratospheric winds above 30 km altitude in the ground-to-space (G2S) atmospheric model explain well the observed anisotropy pattern. We also analyse infrasound from well-constrained explosions in northern Utah with a denser IRIS PASSCAL seismic network. The standard G2S model correctly predicts the anisotropy of the stratospheric duct, but it incorrectly predicts the dimensions of the shadow zones in the downwind direction. We show that the inclusion of finer-scale structure owing to internal gravity waves infills the shadow zones and predicts the observed time durations of the signals. From the success of this method in predicting the observations, we propose that multipathing owing to fine scale, layer-cake structure is the primary mechanism governing propagation for frequencies above approximately 1 Hz and infer that stochastic approaches incorporating internal gravity waves are a useful improvement to the standard G2S model for infrasonic propagation modelling.