The mismatched scales of seismic attributes and geological layering limit porosity estimation directly from seismic data. This can be addressed by utilizing spectral decomposition attributes and machine learning techniques. First, we decompose the seismic data into different frequency components. A variety of seismic attributes then are extracted. We simultaneously predict porosity logs, filtered to different resolutions, using conventional and deep machine learning algorithms. Methods used include support vector regression (SVR), random forest (RF), and the multilayer perceptron (MLP). We then sum the results to create broadband porosity log predictions. We first use synthetic seismic data created from rock physics modeling to test the efficacy of the proposed method, followed by the testing of field seismic data from the North Sea. We compare our method with several different conventional methods and workflows commonly used in industry. The porosity prediction results indicate that the proposed method performs better than other conventional methods and workflows, with a highest correlation coefficient of 0.94 on synthetic seismic data and 0.81 on the field seismic data example.

This content is only available via PDF.
You can access this article if you purchase or spend a download.