Abstract

The bright spot and seismic chimney, which are prominent gas indicators observed in seismic data, exhibit distinct geophysical features such as high amplitude, phase reversal, low continuity, and frequency attenuation. As a result, seismic attribute analysis has been widely applied to derive gas distributions within subsurface media. However, seismic attribute analysis has limitation that it is difficult to clearly distinguish gas indicators from other strata with similar geophysical properties. Therefore, this study proposes a machine-learning method to predict the distribution of bright spot and seismic chimney within seismic data. To effectively predict gas indicators in complex seismic data, the study constructed training data by simultaneously using noise-reduced seismic data and various seismic attribute analysis results. The proposed method was applied to 3D seismic survey data acquired from the F3-block in the North Sea, Netherlands to verify the effectiveness of the proposed method. The gas indicator distribution predicted by the trained model demonstrated higher accuracy and consistency compared to traditional multi-seismic attribute analysis results. Additionally, sensitivity analysis and a forward selection method were applied to optimize the selection of input data, confirming that the prediction accuracy was improved when input with low sensitivity was removed.

Keywords seismic attribute analysis, machine learning, gas distribution, gas indicator, confusion matrix