Blocky Impedance Inversion of Seismic Data Based on Semi-Supervised Learning Scheme

Semi-supervised learning scheme for seismic inversion uses less labeled data to obtain good inversion results of impedance values by using the forward process of seismic data as constraints. Considering the impedance with a blocky structure is helpful for subsequent interpretation and reservoir characterization, an adaptive edge-preserving smoothing (AEPS) filter was introduced into a closed-loop deep residual network (ResNet) structure. In the training process of the inverse subnet, by constructing a loss function of blocky-structure constraint using the filtered impedance by AEPS filter, the inverse subnet is updated to make the inverted impedance tend to be blocky. Both numerical tests and real data applications demonstrate that the semi-supervised learning scheme combined with AEPS filter can obtain ideal inversion results for impedance.