Seismic Signal Denoising with Attention U-Net

Seismic stations record superpositions of the seismic signals generated by all kinds of seismic sources. In earthquake seismology, seismic noise sources can be natural events such as wind or anthropogenic events such as cars. In this study, we developed a machine learning (ML) based algorithm to remove the noise from earthquake data. This is important since the information related with the features of the seismic event may be overlapped by the noise. The presence of noise in the recordings can affect the performance of the seismic network, lowering its sensibility and increasing the magnitude of completeness of the seismic catalogue. To train ML model, 10000 thousand earthquake records with relatively low signal to noise ratio (SNR) are selected and contaminated by 25 noise records that are magnified up to 50% of peak amplitude of the earthquake signal and frequency content of those signals are stored as three component traces. The architecture used consists of an Attention U-Net, i.e. an encoder-decoder model using an attention gate within the skip connections: the encoder maps samples from input space (the waveform STFTs) to a latent space while the decoder maps the latent space to the output space (the signal-noise mask). Skip connections are introduced to recover, from previous layers, fine details lost in the encoding-decoding process. Attention gates identify salient regions and prune inputs to preserve only the ones relevant to the specific task. The use of attention gates in skip connections allows to pass "fine-detailed" information to high levels of the decoder that the model itself considers useful to the waveform segmentation. Trained model is tested with a new set of data to understand its capabilities. It is found that trained model can significantly improve the SNR of noise signals with respect to standard filtering methods. Hence, it can be considered as a strong candidate for seismic data filtering method.