Seismic Facies Classification Algorithm Based on the EarthTransNet

Abstract

Seismic exploration is an interdisciplinary subject. Combined with artificial intelligence, it can automatically identify seismic dips and distinguish faults. The application of Deep Neural Network can reduce the error of manual recognition and improve the efficiency of recognition. Most seismic exploration datasets lack labels, so the supervised learning algorithm cannot be used to extract image features in order to obtain better seismic facies classification effect. Due to the proposal of the F3 dataset which contains real labels in 2019, the supervised learning algorithm can be used on 3D seismic data to take less time and get better prediction results. It is an effective means of evaluation. However, the classification effect of some deep learning models is not satisfactory, especially the neglect of underlying features and the misclassification of small categories on the F3 dataset. Therefore, we apply firstly the TransUNET to the F3 dataset, and modify the input method to 3D volume data, the Transformer layers are added at the end of the CNN layers to collect deep and potential information. The output of the decoder needs to be integrated in the X, Y and Z directions to get the final result. Finally, we propose EarthTransNet, which is applied to the seismic dataset to obtain higher accuracy and better boundary characterization ability.