Noise suppression for DAS seismic data with attention-aided generative adversarial network

The fiber-optic distributed acoustic sensing (DAS) technology provides a highly efficient method for geophysical exploration, in which a fiber cable is equivalent to thousands of receivers with high density acquisition in space. However, the data interpretation for the vertical seismic profile (VSP) obtained from DAS is deteriorated by noises. Therefore, in this paper a noise reduction method with attention-aided generative adversarial network (GAN) is proposed. It integrates three sub networks: feature extractor, generator and discriminator. Specifically, in the feature extractor, the multi-head self-attention mechanism is used to generate a spatial attention weight matrix to extract the key information of the noises quickly. Then the original DAS-VSP signal and the spatial attention weight matrix are fed into the generator, and the noise reduction of original DAS-VSP signal is realized by the adversarial mechanism between the generator and the discriminator. A total of 80 data groups were divided into training set and test set according to the ratio of 7:3. Finally, on the test set, the average duration, signal-to-noise ratio (SNR) and structural similarity (SSIM) were 3.5s, 17.85 dB and 0.89 respectively.