Deep learning for quality control of receiver functions

Receiver function has been routinely used for studying the discontinuity structure in crust and upper mantle. The manual quality control of receiver functions, which plays a key role high-quality data selection accurate structural imaging, challenged by today’s booming volumes. Traditional automatic methods usually require tuning hyperparameters fail to generalize low signal-to-noise ratio data. Deep learning increasingly deal with extensive seismic However, it generally requires manually labeled dataset, its performance is highly related network design. In this study, we develop compare four different deep designs traditional using 53293 functions from three broadband stations. Our results show that combination convolutional long-short memory layers achieves best ∼91% accuracy. We also propose fully training schema zero yet similar carefully ones. Compared method, our model retrieves ∼5 times more reliable relatively small earthquakes magnitudes between 5.0 5.5. average waveforms H-κ stacking these are comparable those obtained larger than 5.5, further demonstrates validity method indicates potential making use smaller analysis.