Seismic Data Reconstruction by the Residual U-Net Network Based on Attention Mechanism

The geological conditions and collection environment and other factors make it impossible to acquire complete seismic data during geological exploration. Deep learning’s fast growth has resulted in a growing number of useful algorithms being used to rebuild seismic data. In this research, we offer a residual U-Net network model with an attention mechanism (A-Res-Unet). The traditional U-Net network has the problem of losing the extracted features due to the down sampling method, and we use hybrid pooling instead of the maximum pooling to retain more feature information. Also, by introducing residual blocks as part of the model, we increase the network depth, improve the precision of training and prevent the over fitting problem. In order to successfully and exactly recover what is lacking seismic data, an improved channel attention mechanism (ECA) is added to the residual block, and a spatial attention mechanism is added between the encoder and decoder to further mine the spatial and channel information of the seismic data. Experiments on synthetic and field data demonstrate that for missing seismic data, the A-Res-Unet model can produce impressive reconstruction results that ultimately present better accuracy than the U-Net and ResNet models that have been used previously.