Acoustic Impedance Inversion from Seismic Imaging Profiles Using Self Attention U-Net

Seismic impedance inversion is a vital way of geological interpretation and reservoir investigation from geophysical perspective. However, it inevitably an ill-posed problem due to the noise or band-limited characteristic seismic data. Artificial neural network have been used solve nonlinear inverse problems in recent years. This research obtained acoustic profile by feeding background into well-trained self-attention U-Net. The U-Net got convergence appropriate iteration, output predicted profiles test. To value quality different perspectives, e.g., correlation, regression, similarity, we four kinds indexes. At same time, our results were conventional methods (e.g., deconvolution with recursive inversion, TV regularization) 1D was calculated contrast. Self-attention showed be robust does not require prior knowledge. Furthermore, spatial continuity also better than deconvolution, regularization, deep learning this paper type full convolutional network, so there are no limits shape input. Based on this, large can U-Net, which trained patchy training dataset. In addition, applied proposed method field data Ceduna survey without any label. predictions prove that could generalized synthetic