Desert Noise Suppression for Seismic Data Based on Feature Enhancement Denoising Network

Increasing the quality of seismic data is the basis for subsequent seismic data processing and interpretation. In the acquisition of seismic records, the introduction of noise is the main factor that affects the quality of seismic data. Especially in the desert areas, the complex desert noise has the characteristics of high energy, nonstationary, nonlinear, low frequency, non-Gaussian, which brings great difficulties to the conventional denoising methods. Due to the powerful learning ability, we have been devoted to use convolutional neural networks (CNN) to suppress the desert noise of seismic data. However, as the depth increases, network might be faced with the problem that the influence of shallow layers on deep layers is reduced, which is not conducive to the suppression of complex desert noise. This paper raises a new network named a feature enhancement denoising network (FEDnet). First, the proposed model increase the width of the network and fuse the feature information of several different layers to fully utilize the influence of the whole network. It is beneficial to capture more desert noise features concealed in the complicated background. Second, we integrate the hybrid dilated convolution design into our model to improve the receptive field, which plays an important role for getting more context information in the denoising task. Finally, we also adopt residual learning to facilitate the network training. The experimental results demonstrate that the proposed network can obtain superior performance for the denoising task of seismic data compared with existing denoisers.