Seismic Signal Denoising using U-Net in the Time-Frequency Domain

Signal denoising is one of the main routines comprised in the seismic data processing chain in order to improve the signal-to-noise ratio (SNR) of registered signals. In this paper, we propose a new method for seismic signal denoising based on a U-Net convolutional neural network architecture. The model is able to learn a decomposition of the noisy seismic signal into the denoised version of the signal and noise. This decomposition is performed in the time-frequency domain, by learning masks to extract both the denoised seismic signal and the corresponding noise. In order to prove the effectiveness of the proposed approach, we use a publicly available dataset, namely the Stanford Earthquake Dataset (STEAD).