Deep Learning with Fully Convolutional and Dense Connection Framework for Ground Roll Attenuation

Abstract

Ground roll could seriously mask the useful reflection signals and decrease the signal-to-noise ratio (S/N) of seismic data, thereby affecting the subsequent seismic data processing. It is challenging for traditional methods to effectively extract high-fidelity reflection signals when ground roll noise and low-frequency reflection signals overlap in the frequency domain. We propose a fully convolutional framework with dense connections to attenuate ground roll (GRDNet) in land seismic data. GRDNet mainly consists of four blocks, which are convolutional, dense, transition down, and transition up blocks. The dense block consists of several convolution blocks to extract the waveform features of the seismic data. The short-long connection in the dense block and the skip connection in the encoder-decoder not only reuses the features extracted by the previous layer but also adds constraints other than the loss function to each convolution block. The well-trained network is tested on one synthetic data and two real land seismic datasets containing strong ground roll with linear and hyperbolic moveouts, respectively. Three traditional and two state-of-the-art deep learning (DL) methods are used as benchmarks to compare denoising performance with GRDNet. The testing results show that the proposed method can effectively attenuate the ground roll in seismic data and preserve useful reflection signals.