search

AI4EPS / PhaseNet

PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method
MIT License
268 stars 111 forks source link

Getting different time and probability when changing the input stream length. #74

Closed saeedsltm closed 1 year ago

saeedsltm commented 1 year ago

Hi, When trying to change the input stream length (from 10 min to 1 houre), I'm getting different time and probability of P and S phases. That causes when assocciating phases, in some cases we missed events. I can't undrestand why changing the input stream length would affect picking phases?

zhuwq0 commented 1 year ago

Hi, this is mainly because of the data normalization: https://github.com/AI4EPS/PhaseNet/blob/1d7ea66a8f1e68b1284e39e0bf8524079790f1ed/phasenet/data_reader.py#L64

saeedsltm commented 1 year ago

Hi @zhuwq0, Ok, so is there any preferred time window? Should we avoid passing long streams like a day (86400s)?

zhuwq0 commented 1 year ago

Hi, the moving window normalization should not be strongly dependent on the window size. The location of the event within the moving window will cause the difference. You can see this paper: https://pubs.geoscienceworld.org/ssa/srl/article-abstract/doi/10.1785/0220230003/620893/A-Mitigation-Strategy-for-the-Prediction

saeedsltm commented 1 year ago

Hi, Thank you very much. I will go through the paper, and so I close this issue.