Evaluating CNN-based models for seismic data denoising

Noise suppression is an important research topic in seismology and seismic signal processing. Accurately suppressing seismic noises and extracting effective signals is a key step in seismological research and seismic monitoring. Traditional denoising methods have some shortcomings, such as insufficient flexibility, difficulty in dealing with complex noises, information loss, and dependence on manual feature extraction. In order to overcome these shortcomings, this paper probes in a method of time-frequency domain transform combined with deep learning and its application to noise reduction. Five neural network models (including FCN, Unet, CBDNet, SwinUnet and TransUnet) are constructed for noise suppression after the time-frequency transformation of seismic data. This paper introduces three indicators: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and root mean square error (RMSE) for quantitative evaluation of denoising performance. Numerical experiments show that the convolutional neural network(CNN) method based on time-frequency transform can effectively suppress common noise types (including random noises, swell noises and surface waves) and improve the signal-to-noise ratio of seismic data. The introduction of the Transformer module can further reduce above-mentioned noises and enhance the denoising performance of the CNN model. Further research will focus on an improved network structure for more complex seismic signals and the combination with other advanced techniques to improve denoising performance.