instrument response removal for very low frequencies

[carltape]

I'm not sure if this is a pysep issue or obspy issue, since we are using obspy functions within pysep. The testing notebook is response_removal_test.ipynb. This contains more than is needed to demonstrate the issue, which is this: When testing a range of options for removing the instrument response, we cannot seem to achieve the option to directly divide FFT(c(t)) by Ia(f) to get a(f), where c(t) is the time series in counts, Ia(f) is the instrument response to acceleration, and a(f) is the acceleration in the frequency domain.

Here are four subplots (from the notebook above) showing the 0.2-1.0 mHz spectra for II.NNA.00.LHZ derived from a 10-day time series of the 2004 Sumatra earthquake. From the top, they are: 1) instrument response to acceleration (simple flat line) 2) raw 3) acceleration from manual removal of response 4) acceleration via obspy remove_response (via pysep) using the best input options we found (water_level = None, pre_filt = default).

Here is II.SUR.00.BHZ:

For this class problem, we would like to achieve the cleanest results for instrument response removal, which seem to be the direct division. We hope that this is possible by using the remove_response function in obspy (and, in turn, pysep), but we can't see how. Perhaps there is some taper or operation that is happening in obspy that is causing the issue. Or perhaps pysep is applying some additional operation. Thanks.

[carltape]

Offering up some improved results (from Aakash) by using the taper_percentage option in pysep.

Here is NNA for using a 20 percent taper (taper_percentage = 0.2):

Here is NNA for using a 50 percent taper (taper_percentage = 0.5):

Still, the pysep version is slightly worse than the manual removal, so there must be something else happening inside pysep (or remove_response).

[carltape]

Attaching a test script here, since the notebook at the top of this thread is no longer available. test_response_removal.txt