Lithology Prediction of One-dimensional Residual Network Based on Regularization Constraints

Lithology prediction is an important work in seismic reservoir prediction. Deep learning can explore the nonlinear mapping relationship between lithology and seismic properties, and achieve efficient and accurate lithology prediction. On the one hand, when the depth of the network increases, the problem of model degradation is prone to occur. On the other hand, due to the small sample size of logging data, overfitting is common when deep learning methods are used for lithology prediction. We apply a one-dimensional residual network to lithology prediction with regularization constraints on the overfitting phenomenon of the model. According to the change of loss function under different regularization constraint methods, the influence of regularization constraints on model overfitting is analyzed. Compared with the initial model, the prediction accuracy of the model with regularization constraints in the validation set is improved from 48.81% to 59.87%. When considering adjacent lithology, the validation set accuracy improves from 89.37% to 91.54%. The proposed model achieves 92.65% accuracy on the test set. Applying a regularized residual network model to seismic data prediction can effectively indicate the distribution of subsurface lithology.