Analog network clipping

[gthompson]

Q: There is also another seismic network (the “analog” network), which had much less dynamic range, but captured the only data from the first 20 months of the eruption, so these signals are an important target. The result is that many signals are “clipped” (saturated).

A: Marielle agreed it makes sense to train on the digital network first, and then just try to apply this to analog network and see how it does. And then iterate from there, coming up with strategies to train a model and/or correct saturated data.

M: Also this might be interesting from a “can our model train on some data generalize to different data” point of view. Would be interesting in a publication!