Abstract:  Due to the limitation of acquisition conditions and costs, the pre-stack seismic data may be irregularly distributed or incomplete in space, which  brings difficulties to the subsequent processing and interpretation of seismic data. In recent years, the convolution neural network method widely used in the reconstruction of missing seismic data lacks attention to the global information, while the network model with multiple downsampling brings low-frequency signal loss, and the reconstruction results of the low-amplitude missing part still need to be further improved. Therefore, this paper proposes a multi-scale feature self-attention model. A multi-scale wavelet fusion block based on the self-attention mechanism is designed at the bottleneck of the U-Net backbone network, and the outputs of all encoders are fused by discrete wavelet transform and self-attention mechanism, which effectively balances the global and local feature processing and reduces the signal loss caused by downsampling. A multi-scale receptive field is inserted into the network to improve performance and enhance spectral learning of different frequencies by learning multi-scale features for different degraded data. Compared with the classical reconstruction methods for seismic data, the reconstruction results of the algorithm in this paper are improved in both qualitative and quantitative assessments. On the synthetic dataset and the real dataset with 30% continuous missing  data,  the signal-to-noise ratios  of the reconstruction results   are 21.748 7  and 14.954 0 dB  respectively; On the synthetic dataset with 50% random missing and regular missing data, the signal-to-noise ratios of the reconstruction results are 28.832 0  and 37.724 2 dB  respectively.

Key words: self-attention mechanism, wavelet fusion, multi-scale receptive field, seismic data reconstruction