Desert mixed seismic noise suppression by using multiple forward models and a supervised deep-learning method

Abstract

Desert seismic data pose a particular problem due to its low signal-to-noise ratio and severe spectral overlapping. Most of the existing denoising methods have negative suppression effect on the desert mixed seismic noise, which mainly includes random noise and surface waves. The convolutional neural network (CNN) is a classical supervised deep-learning method that has proven to be an effective tool for noise suppression. The drawback is that CNN usually needs a large amount of ideal noise-free data to complete network training, which largely limits its applications to seismic data denoising. In this paper, we combine forward modelling with CNN to overcome this technical problem. Specifically, we construct multiple forward models from different physical parameters to obtain a noise-free seismic dataset by finite difference method. After adding noise data taken from a real desert noise record, we use the above data as the input for CNN training. Then, the MSE loss function allows us to obtain an optimal denoising model for the mixed noise suppression of desert seismic data. Continuing with this optimal model, we can predict mixed noise, which, once subtracted from the desert original seismic data, provide the denoised desert seismic data. Experiments with synthetic and real data demonstrate that this CNN-based denoising method can suppress the desert mixed seismic noise successfully. Several tests using artificially generated models with a different number of signal patches prove that the forward modelling can significantly improve the continuity of events and prevent the appearance of false events. In addition, analysis of the effect of network depth supports the effectiveness of our method through a trade-off among higher signal-to-noise ratio, lower root mean square error and shorter training time.