2D Multi-Spectral Convolutional Encoder-Decoder Model for Geobody Segmentation

In this paper, we propose a workflow of automatic facies analysis with seismic amplitude data and its spectral decomposition information. Although the spectral magnitude and phase are shown to have higher signal-to-noise ratios, it is rarely involved in recent methods. The existing methods usually use a convolutional neural network (CNN) to extract features from raw seismic data without considering that the rock properties are often carried by spectral responses at particular frequencies. In this paper, we incorporate the spectral voice components that contain spectral magnitude and phase at given frequencies into a convolutional encoder-decoder architecture. This combination considers multi-spectral fine features as well as seismic broadband features for facies classification. Experiments are conducted on an open dataset from geoscience research community for classifying eight classes of seismic facies. Performance on the test set using the proposed workflow has shown good results with an IoU of 0.882. Then we apply the trained model on the whole 3D seismic data (including the non-labeled data) and visualize the spatial distribution of the prediction masks. The result shows the robustness and effectiveness of our architecture.