Seimic impedance inversion based on semi-supervised learning

In coalfield seismic exploration, seismic impedance inversion can infer subsurface physical properties through seismic data, allowing the identification of subsurface strata. Deep learning has shown great potential in the field of seismic impedance inversion. Typically, the quantity of labeled data during training significantly impacts model validity. To reduce model dependence on labels, we propose a semi-supervised learning workflow for impedance inversion. By using a small number of seismic and well logging labels, a closed-loop mapping is established from seismic data to impedance data, impedance data to generated seismic data. In addition, non near-well seismic data are introduced to participate in network training during each iteration update process of the network. The waveform matching of non near-well seismic data can ensure the prediction accuracy of non near-well impedance to a certain extent. We evaluated the algorithm using a dataset from the Xinjing coal mining area near the city of Yangquan in Shanxi Province, China. The results showed that the proposed semi-supervised inversion method performed better and demonstrated improved inversion results relative to the conventional single-supervised method.