Comparison of Convolutional Neural Networks for Dividing Seismic Sequences

Abstract

Application of the machine learning technique is expanding to the field of seismic exploration. For the purpose of feasibility assessment, we applied the machine learning technique to seismic sequence identification, which is important in seismic interpretation. From among the convolutional neural network techniques used in image analysis, we applied four models to seismic data obtained in the F3 block, offshore Netherlands, which have yielded remarkable results. One of the four models was ResNet34. The others were encoderdecoder models: U-Net, Residual U-Net, and FD U-Net. We first performed a qualitative analysis of the predicted images and then conducted quantitative analysis using pixel accuracy, mean class accuracy, mean intersection over union, and frequency weighted IU equations. The numerical results showed that ResNet34 had the lowest accuracy and that the encoder-decoder type models had higher accuracy. Considering the number of parameters required for calculation and the learning time, we confirmed that U-Net is the most efficient model.